Source: The Guardian

Subscribe now

Earlier this year I wrote a Substack piece on “Slowing Down the AI Train.” After years of thinking there was no reason for serious concern about AI development, I decided that as a society we need to figure out some way to limit the pace of development. I was hardly the first to do so, but since then many more voices have joined this particular chorus. Two weeks ago, for example, 1100 AI leaders and professionals signed a letter calling for “an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.” Last month, 200 economists—including 16 Nobelists—and executives signed a letter arguing for safeguards against large-scale job loss from AI. In addition, organizations like the Future of Life Institute, Pause AI, Stop AI, and several others have raised substantial concerns about the uncontrolled pace of AI development. Recent automated hacks by systems from Anthropic, OpenAI, and Meta haven’t helped the image of generative AI models. The anti-AI chorus is becoming increasingly loud, although there are still plenty of anti-AI regulation boosters (“accelerationists”) out there.

Political Levers

As I stated in my piece last February, there aren’t a lot of sure-fire methods to slow down the development of AI. But there are some potential levers that could help. The most commonly discussed are legal and regulatory approaches. Since the Trump administration’s only consistent AI policy is to favor Big Tech firms and owners who support the administration financially, it seems unlikely that there will be federal regulation in the US on AI for the next several years. Although Trump signed an executive order prohibiting states from passing their own regulation, that appears to be unconstitutional and hasn’t stopped states from proposing (47 pieces in total) or enacting (8 thus far) legislation regulating AI, according to this tracking map. There may be more bipartisan consensus across states on regulating AI than on any other issue! Therefore, if you want to reach out to legislators to express your desire to slow AI down, you’re probably better off contacting state officials rather than federal ones.

Of course, what we really need is global AI regulation, since multiple countries are developing AI and any problems with it are likely to have global implications. Alas, however, meaningful global regulation is highly unlikely. A survey of AI experts released recently by the Council on Foreign Relations found that 80% of the experts believe that regulation will continue to be fragmented, and that 70% believe the most likely trigger for change would be a major AI accident. Yes, that’s a bit depressing.

Economic Levers

There are also various economic levers that could slow down AI development. The primary vendors seem hell-bent on investing and circular funding themselves into bankruptcy, so there may be no need for further economic intervention. But in case they don’t pop their own bubble, consumers can slow down the train by simply spending less on AI subscriptions and tokens. Taking advantage of the increasing number of open-weight models is one way to do that. Not using AI for highly frivolous purposes is another.

Many businesses have already realized that generative AI costs a lot and yields economic value—particularly in terms of individual productivity—that is difficult to measure. So they (82% of them according to an EY survey) are already cracking down on tokenmaxxing and using AI routers to find the cheapest generative model that will get the job done. The fact that this newfound thriftiness takes place just as AI vendors are doubling down on capital expenditures may also by itself bring down startup valuations and the resulting pace of development.

I realize that generative AI can be useful, so I’m not advocating any sort of boycott—not that it would succeed if I did. I don’t want AI to go away—I just want fewer new data centers, a new model once every few months instead of every day, and greater attention paid to testing and guardrails.

Technology Levers

Technology got us into this mess, and there’s a good chance it can help get us out if we want it to. The “hair of the dog that bit you” approach attempts to use software to help us use AI in the right—and often less aggressive—ways. Software, for example, can monitor whether we are using AI as a cognititive crutch or a Band-Aid. Skillbench, for example, is an early-stage startup cofounded by my friend Matt Beane to help software developers use AI to build their skills rather than deskill them. In education, tools like Cherry Pot help students learn complex problem-solving domains like advanced math. As they work on problems they receive real-time, line-by-line feedback on their reasoning and methodology. Even the most commonly-used models can act as a coach to inform users how they can write better, for example. I hope we will see many new possibilities along these same lines.

There are a variety (more than I can easily name) of AI governance-oriented software offerings that provide usage controls and guidelines. Some are primarily oriented to preserving data privacy and cybersecurity, while others focus on limiting AI costs. While these are generally a good idea, it’s important for organizations to strategize about what governance objectives are most important, and to remember that only some of these tools will slow down AI use and hence AI development.

A final type of software that’s still emerging focuses on one of the AI industry’s primary rationales for unbridled acceleration: if we don’t do it, China will. It measures compliance with specific rules—as yet largely undeveloped—that would govern international competition in AI model capabilities. If, for example, the US and Chinese governments agreed on rules about what models from each country could and couldn’t do, this software—from startups like Lucid Computing and Amodo Design—would assess compliance with the rules. This category of software is described in an unlikely location—an article in Time magazine.

What’s Not Allowed

To be clear, I am absolutely opposed to approaches to AI slowdown that are generally illegal and to me immoral. That includes threatening or harassing AI company leaders, which is already happening and will only reflect poorly on the AI slowdown movement. Threats to sabotage AI-related facilities like data centers are also increasing. Such activities are a great way to find your way into jail, but are the antithesis of calm, reasoned discussion about potential harms of AI and how to prevent them.

Source: Ben Hickey, The Economist

Subscribe now

I have co-authored several articles over the last year or two that together suggest organizations are not likely to get business value with genAI if their primary focus is improving individual productivity. But everybody doesn’t read Harvard Business Review, and I have never put all the reasons for this in one place. So here goes—a laundry list of explanations for why you’re probably not going to achieve measurable productivity gains from genAI.

1. You need to redesign end-to-end processes around AI capabilities if you’re going to get substantial value from AI. And that, as I wrote about a couple of months ago, is hard. Even if you do it successfully it’s going to take years.

2. We’re not training people well. I saw a good summary of this argument in Robert Scoble’s newsletter last week. To get productivity from genAI demands that individuals incorporate AI into their daily work. But if you only get generic education—unrelated to your industry or your function—that’s unlikely. AI education should be highly segmented, with each business and job class getting their own version if it.

3. If you use genAI in the “right way,” you don’t save a lot of time and effort. The right way, at least if you want to create high-quality content and avoid rotting your brain, is to employ multiple prompts until you get something close to what you want, and then to review it for possible hallucinations. Then edit it to take out the cliché phrases and to add your unique perspective. When I insist that my students do this with their AI-assisted essays, they often tell me it doesn’t save them any time over writing it all themselves.

4. Measuring the aggregate impact of individual productivity gains is really hard. You have to measure how long it took you or someone else to perform a task pre-AI. Then you have to measure post-AI. Then you have to do that for multiple tasks and determine how all that rolls up to entire jobs and organizations. Not surprisingly, few organizations do this type of measurement.

5. It’s unclear what people are doing with the saved time. Let’s assume that you do the needed measurement and figure out that the average employee in a particular job saves an hour a day. What are those people doing with that saved hour? Maybe they’re doing more work, or maybe they are streaming a World Cup game. Theoretically you could lay off 1/8 of the people who do the job, but that’s hardly ever done. Most organizations aren’t confident enough about the measurements to take that step.

6. Token costs are going up, which makes getting productivity improvements harder. Now that tokens are being measured and billed by vendors more assiduously, to claim any real productivity gains you need to compare the savings from any labor time reductions to the cost of the tokens needed to perform that task. Again, it’s a measurement hassle that hardly anyone undertakes.

7. To really get productivity requires that we explore all the capabilities of genAI to transform our jobs. But most people don’t invest a lot of time in personal infrastructure development. I read a lot about individuals who create agents to do much of their work. But I don’t know any such people and I strongly suspect there aren’t too many of them. It takes a substantial investment in time, tokens (see above), and learning to automate significant aspects of your life. Most of us would rather watch the World Cup, Desperate Housewives, or YouTube videos in our spare time.

8. Organizations don’t do controlled experiments. To really figure out whether, for example, genAI improves the productivity (and quality) of blog post creation, you should do an experiment. You might assemble a group of bloggers who use a lot of genAI help to write their posts, a group that just uses it for brainstorming, and a control group that doesn’t use it at all. You then measure their average time to produce outputs and have some independent group score the output quality. Again, a few university researchers have done this sort of experiment, but they are rare in real organizations.

9. Poor genAI outputs lower productivity. This is the “workslop” problem that many observers have begun to notice and write about. Your own productivity may not suffer when you use AI to create workslop, but other people down the line may need to improve your output, and their productivity will suffer. Matthias Hollweg and I recently described the “process slop” problem, in which interorganizational use of AI leads not only to productivity loss, but reduced trust in the integrity and value of the entire process.

10. AI agents will help with the individual productivity problem, but probably won’t solve it. Teams of agents are typically for enterprise use cases more than individual ones. Even then, somebody is going to have to monitor the quality of agents’ work, and that will detract from any productivity gains. It’s true that some aggressive individual users are already adopting agents, but their efforts will run into some of the same issues involved in enhancing productivity that I’ve described above.

What does this mean for our AI-driven economy? It probably indicates that AI providers are over-valued by the market, that GDP growth that’s largely driven by data center construction is unhealthy, and that generative AI is not enough to power the economy on its own.

These factors may also mean that we are unlikely to see substantial productivity gains from AI as an overall economy. Thus far we certainly haven’t. Over the last seven years—half of which period includes the genAI era—average annual nonfarm productivity in the US (as measured by the Bureau of Labor Statistics) has increased by 2.1%. That is exactly the amount that productivity increased between 1947 and 2026. In the first quarter of 2026 productivity increased by only 0.3%. If genAI is increasing our productivity, it’s apparently not easy to measure at the overall economy level. Perhaps it’s just too early for us to see AI-driven productivity gains, but how long do we need to wait?

Subscribe now

Stanford professors Bob Sutton and Melissa Valentine at AI + Organizations Conference

Much of the effort and attention around AI for the last several years has been around technical developments. New model announced! New benchmark surpassed! New contract for massive data centers! New world-class technologists hired! You know the drill.

I am happy to say, however, that things are beginning to change in this regard. AI companies are beginning to realize something that many corporate executives knew intuitively. What matters isn’t the technology—OK, that’s important too—but the ability of organizations to deploy it effectively and get value from it.

There are several signals of this shift. I participated in one a couple of weeks ago, giving a talk at the first conference held by the “AI and Organizations Lab” at the Stanford Institute for Human-Centered AI (HAI). The lab, headed by Stanford professor Melissa Valentine (a co-author I have enjoyed working with), is one of the first—and certainly the most prominent among—AI research centers to focus on organizational issues.

The first day of the conference was attended by academics, reflective business practitioners, and representatives of the lab’s key sponsor—none other than Google DeepMind. That one of the world’s great centers of AI technology believes that organizational factors are important to the success of AI overall is a big step forward. There is an ongoing research center at DeepMind on organizational AI topics, headed by Martin Gonzalez.

The second day involved a large group of organizational behavior academics and Ph.D. students. The goal was to begin outlining some areas of interest and methods for effectively researching and publishing about generative AI. I am often frustrated with how long it takes for academic research to be published, but there was an exception represented at the meeting. Aruna Ranganathan and a co-author published the HBR article “AI Doesn’t Reduce Work--It Intensifies It” earlier this year, and it’s a fine example of how organizational behavior research can relatively quickly shed light on AI’s impact on people and companies.

AI Companies Addressing Deployment as Well

In addition to Google, a couple of other major AI providers recently announced initiatives to address deployment issues. OpenAI, for example, announced “Frontier Alliance” partnerships with BCG, McKinsey, Accenture, and Cap Gemini to implement AI in customer firms. OpenAI had previously begun to hire “forward deployed engineers”—the new phrase for a technically-focused consultant—in 2024 to help their customers deploy AI use cases and move beyond pilots, and the consulting firms will partner with them. The widespread use of the FDE role, perhaps driven in part by the stock multiple of Palantir, which invented the term, probably still isn’t sufficiently focused on organizational and behavioral issues judging from this job post, but it’s a step in the right direction. OpenAI also announced recently that it would partner with several large private equity firms to help deploy AI in their portfolio companies. The OpenAI Deployment Company, like many OpenAI initiatives, also seems like a way to raise more capital.

Anthropic’s moves are very similar to OpenAI’s. The Claude vendor created an “Applied AI” capability to aid deployment and has also formed a partnership with private equity companies, the goal of which is to establish a new AI services company.

Granted, neither of these organizations are particularly focused on the human issues involved in putting AI systems into production. The deployment-oriented roles they have created are primarily focused on technology architecture, optimizing performance, and other technical aspects of AI implementation. The umbrella and somewhat generic term “change management” doesn’t appear anywhere in job descriptions or marketing copy.

Is Tech All That Matters?

Even at Google DeepMind’s organizational AI group, I suspect there will be a strong focus on using technology to address organizational issues. For example, DeepMind and Stanford HAI sponsored a “grand challenge” among university teams that proposed various problems and approaches to organizational adoption. The winner in what I assume was an objective evaluation (there were judges from Stanford and Google, but also several other universities) was a proposal from Stanford. That’s fine, but it was pretty technical, as this description suggests:

Wang and Goldberg will use the modern transformer architecture for machine learning, which excels at finding patterns in sequential data, to build a “large coordination model” that learns how successful teams coordinate their work and predicts which sequence of actions will work best for a given scenario in the future.

I have no objection to this effort, although I suspect that human individuals and teams often employ some behaviors that are difficult to predict with a generative AI model.

I am confident that Valentine and her collaborators at HAI (of which I hope to remain one) will not reduce all issues in the AI and organizations domain to technology issues. Perhaps someday even the AI tech firms will realize that the adoption and effective use of AI results from a complex web of human factors including workflows and workarounds, skills, rational and irrational behaviors, emotions, power, and economics. Maybe someday AI will be able to predict and resolve all these factors, but I don’t think we’re close to that yet.

Subscribe now

There is increasing consensus that to get value from AI, you have to redesign your business processes and embed AI within them. A McKinsey 2025 survey, for example, found that redesigning workflows is the factor most highly correlated with getting value from AI. A 2026 paper from MIT researchers argues that AI benefits will come from supporting “chains” of business activities, i.e. processes. My friend Erik Brynolfsson, head of Stanford’s Digital Economy Lab, has long argued that in a “J curve” situation, productivity with AI initially lags as companies re-engineer processes, but then increases significantly once AI is fully integrated into new, redesigned workflows. The distinguished user interface designer Jacob Nielsen wrote recently that: “In a controlled field experiment, startups that redesigned end-to-end workflows around AI generated 90% more revenue than equally equipped peers that used AI mainly to speed up individual tasks.”

None of these assertions or survey results is terribly surprising. AI is undoubtedly a powerful tool that can enable new ways of doing work. If it’s just piled on top of existing work activities, its potential will be limited. Processes are intimately connected with the information systems that support them, and a recent Gallup survey found that whether AI is integrated with the processes and technologies used to accomplish work is a key factor in whether workers will adopt it. In short, it’s a no-brainer to use AI to redesign processes, as I have previously argued with co-authors in this article and this one.

So Why Isn’t It Happening?

If in response to the question of, “If this is such a good idea, why isn’t everybody doing it,?” you answered, “Because it’s hard,” you’d be right. There are a variety of obstacles to widespread adoption of AI-driven reengineering. To start, most companies (particularly in the US) aren’t process-oriented. They don’t have clarity on what their processes are, who is responsible for them, how they are currently performing, etc. There isn’t even consensus on what to call them. Note that some of the authors and experts I mention above use the term “workflows” or “chains of activities.” Some companies don’t like the “process” term because it sounds somewhat bureaucratic, but it is both the most historically respected and accurate term in my view.

Of course, processes can be small (as in Lean Six Sigma, and AI can help with those too) or large, end-to-end and cross-functional (as in business process reengineering). I’m a fan of the latter because of their potential for large-scale improvement, and I think that broad end-to-end processes like order management and procure-to-pay provide the greatest benefit when combined with AI. But working on any size processes is better than nothing.

Process design also requires some effort. You need to know where the process begins and ends, what technology currently supports it, and how it’s structured and performing at the moment. Lars Reinkemeyer and I argued in this article that process mining can help companies to understand the performance of existing processes and bottlenecks in their performance, and in general to make companies more process-oriented.

Implementation Is Really Hard

But implementing a new process design is much more difficult and time-consuming than designing it in the first place. That requires not only new process flows but also the new systems, skills, behaviors, etc. that accompany them. And getting people to actually do their work in new ways is never easy. For a major cross-functional process the implementation effort can stretch into many months and many millions of dollars or other currencies.

Perhaps you’re beginning to see why this doesn’t happen more often. Process reengineering for end-to-end processes is not for the faint-hearted or poverty-stricken. It can also fail for a variety of reasons—from lack of stakeholder support to politics to some human or technical component of the process not functioning correctly.

In pre-AI business process reengineering in the 90s and 00s—I was present at the creation, and wrote the first article and book on the subject (though neither were the best-selling versions)—there were several companies I was familiar with that claimed to do reengineering, but were really just laying off people (as described in this 1995 article). It’s not surprising that companies are doing that now with AI. It’s much easier to “AI wash” than to redesign and reimplement new ways of doing work.

Source: Google Gemini

Subscribe now

All those people (4 of them, 3 of whom are related to me) who have followed my writing for decades remember that in 2001 (Sept. 10, unfortunately) I co-authored a book called The Attention Economy. I’ve already written about it in Substack, so I won’t cover too much of the same ground. In this diatribe I will provide more detail on one line from that previous piece:

…with the ease of creating vast amounts of content—and the frequent absence of the human touch in the output—the reader probably won’t pay as much attention to it.

I was motivated to revisit the glut of content issue by an article in last week’s New York Times about “too much code.” I hadn’t realized this was coming, but in retrospect it seems obvious. Some have argued that code generation is a great illustration of Jevon’s Law, in which improved efficiency in the creation of something motivates an increase in demand for that good. In the case of code, the argument would go that if generative AI helps us produce more code more quickly, the desire to use that code would rise at a proportional rate. That the need for code is insatiable may indeed be true, but there is a catch: someone needs to look at the code before it is applied to an important programming need. Otherwise the organization involved is likely to have some breakdowns in major systems (as The Financial Times argued with regard to Amazon Web Services’ recent outages, though AWS denies that genAI code was a factor in them).

In short, more code requires more attention from human coders to test and certify its quality. This is one of the reasons why some measures of coding productivity haven’t improved or have declined with code generation. However, expectations among the senior executives I’ve spoken with for coding productivity have definitely risen, and some have decided not to hire entry-level programmers at all. That, of course, creates an even greater attention deficit.

Many organizations seem to be treating mistakes in code generation as a temporary phenomenon. But as long as genAI continues to work by predicting the next word, sometimes it’s going to predict the wrong one. Even if generated code compiles, it may be sloppily structured and may take too long or too many computing resources to run. So the need for an experienced human coder in the loop isn’t going away soon.

It’s Not Just Code

The content glut is appearing not just in programming code, but pretty much everywhere these days—and genAI seems to be at least partially responsible. One of the things I do is write books, and I am proud to say that (thus far, anyway), AI hasn’t generated a line of text in any of my books. This is partly a contractual matter, as my publishers insist that no significant component of the text is written by AI. You may have heard about the novel Shy Girl that will not be released in the US because its publisher, Hachette, believed that much of it was written by AI, and that wasn’t allowed in the author’s contract.

However, publishers can’t stop people from writing and publishing books themselves. The number of self-published books grew by a million in 2025 over 2024, and AI text generation is suspected (by Publisher’s Weekly) to be a big part of the growth. One wonders, however, whom these authors believe will read their books. From 2003 to 2023, the percentage of Americans who read for pleasure fell 40%. Forty percent of Americans didn’t read a single book in 2025. My guess is if they are going to read something, it will be mostly created by humans.

Things are perhaps even worse in the marketing domain. The majority (76%) of marketers use blog posts to describe their products and services, and 90% percent of content marketers now use AI writing tools, according to a Siege Media survey reported by HumanizeAI. Content creation leads all AI use cases, and 85.1% of AI users deploy it for blog content generation. I suspect that an even smaller percentage of the 600 million blog sites are read than books.

How Does This Work Out? With Work!

So there you have the problem—increasing content generation with AI, but a decreasing amount of human attention to look at it. The problem will only get worse as AI is increasingly used to generate other media—videos, ads, marketing messages beyond blog posts, etc. According to one study, the percentage of Internet content created by AI passed 50% in November 2024, and it is surely increasing. And as humans become more inured to low-quality AI-created content, their attention will go elsewhere.

What is to be done about this? Of course, the imbalance of AI content over human attention only matters if you care that humans read and are affected by the content. Maybe some AI-assisted content creators don’t care, or are creating the content for other AI systems to consume.

If you do care about human impact, there are, I believe, two ways to address the issue:

1. Create the content yourself

2. Work really hard to make it different, amusing, interesting, touching, etc.

This is not a new idea; in 2015 Geoff Colvin, a Fortune magazine senior editor (for 48 years!), published a book called Humans Are Underrated: What High Achievers Know that Brilliant Machines Never Will. He argued that the more AI “slop” (a term created since the book) was created, the more we would appreciate human writing, art, relationships, etc.

Of course, creating content that’s better than AI takes a lot of work, and not everyone is capable of it. One who is capable is my son Hayes Davenport, who has co-founded a new media outlet in Los Angeles called LA Material. He’s the Creative Director there, and he oversees creation of the daily newsletter every Friday. I’m proud of his writing—it’s clever and interesting and fun. For example, he recently wrote a somewhat bizarre investigative piece entitled, “Why Is There a Mosaic of Jim Jones Outside a Silver Lake Elementary School?” When I complimented him on it he bemoaned the fact that it takes him most of Wednesday and Thursday to produce the Friday issue. But AI could not have produced that article, or any of the other things he writes. There’s no shortcut if you want your efforts to get human attention.

As Colson Whitehead noted today (print edition anyway) in a New York Times opinion piece called “Don’t Use AI to Do This,”

I believe in an old-fashioned virtue called Doing the Freakin’ Work.

Read the book, not the summary.

Write the piece, not the prompt.

Suffer like the artist you are. It ain’t easy, but if it were easy, it wouldn’t be worth doing.

Thanks for reading Tom’s Substack! Subscribe for free to receive new posts and support my work.

Subscribe now

My first paying data analysis job was more than 50 years ago. And for the first 45 years or so of my career in this field, it was a pretty consistent set of tasks. You put your data into rows and columns, figured out what to do with missing data, specified what type of analysis you wanted on what variables, and fed it all into a statistical analysis program. You hoped you didn’t make any spelling missteaks (sic), because you’d have to do it all over again. Interpretation of the results, of course, was up to the analyst, as was the explanation of the implications for action for your decision-maker boss.

How GenAI Can Change Data Analysis

But generative AI is clearly changing the process of data analysis. I’ve been experimenting with quantitative data analysis using ChatGPT for several years now. I was certainly impressed by the LLM’s ability to generate Python code to analyze structured data and create machine learning models. With only a short, two-line prompt it could do the following:

• Figure out (in a CSV file) which were the independent variables to analyze as predictors of the dependent variable; all it had to be told was the label of the dependent variable.

• Determine how best to handle missing data

• Do a bit of feature engineering to see if it could create a better fit for the model

• Try several different algorithm types and run a little competition among them;

• Describe which features had the greatest importance in the best-fitting model.

Even more impressively, the LLM could provide a page or two of recommendations for a manager seeking to maximize or minimize the dependent variable (I often used an attrition dataset from the telecom industry). It might suggest, for example, what products a company should offer in what bundles in order to reduce attrition.

There was only one problem with this approach to data analysis. I noticed that minor differences in prompt language led to differences in the results. I’m not experienced enough in Python to determine why this happened, but it made me hesitate to recommend ChatGPT-based data analysis to others. Perhaps that’s why LLM-driven machine learning isn’t practiced all that much.

Enter Claude

A few months ago I tried uploading a CSV dataset to Claude for analysis, but it told me that it couldn’t accept uploaded CSV files. But over the December holidays I had some data to analyze for a survey on the economics of AI. My collaborator on the project, Laks Srinivasan, suggested that we see what Claude could do with the analysis. He said that Claude still had issues with uploading CSV files, but that it seemed to accept SPSS data files, which we had received from the survey provider. Laks fed Claude the SPSS files and gave it some prompts with context about our study and what we were expecting to address in the reports and articles emerging from it.

And in general, Claude did an amazing job. It quickly identified what it thought were the most interesting and important findings. It created an outline for the final report and would no doubt have written a draft of the report had we let it. It performed crosstabulations of several variables, and identified some subtle findings that I am not sure I would have noticed.

We all, including Claude, were heavily focused on one issue. Laks and I had previously identified the survey questions and variables involved in AI economic maturity. Claude made that the focus of its analysis. As Claude pointed out, “Every consulting firm has an AI maturity model. Most are useless….We took a different approach: economic value creation maturity.” (bold font supplied by Claude).

Claude also found it very important to point out that many companies in the survey were “stuck” in the middle of the economic maturity model, particularly on Stage 3 (in which economic value from AI is measured but not aggregated across the company).. It suggested several times that the surveyed companies had been stuck there for a median of six years. It commented:

“Nearly one-third of organizations (30.1%) are stuck at Stage 3 for a median of 6 years. They achieve respectable value (44.2%) but can’t aggregate because each use case is measured differently. The escape requires standardization, not more measurement.”

I found that conclusion interesting, since we didn’t ask companies in the survey how long they had been at each stage. When Laks asked Claude how he came to that conclusion that many companies were “stuck at Stage 3,” it replied that 6 years was the median time Stage 3 companies had been using AI. Claude did not acknowledge that companies could have been at a higher stage and fallen back, or that they had perhaps been in Stage 3 for a short time.

Claude may have been wrong about this finding, but it was never in doubt and continued to push that idea. To its credit, however, when we asked for an explanation it said that these companies “have achieved a fair amount of value but can’t standardize value across use cases.” The evidence it cited included:

• 52.8% cite “lack of AI-ready data”

• 50.5% cite “technology difficult”

• 49.2% cite “lack of framework”

• 38.3% cite “leaders lack understanding”

That logic was moderately compelling, so we left the “stuck at 6 years” in the report.

It also turns out that Claude is also a big evangelist for agentic AI. Perhaps that’s not surprising, as it represents an important direction for the LLM’s future. Even though a relatively small fraction (about a third) of companies in our survey were using agentic AI, and those using that technology got the least amount of high value from it (2%) than any other technology, Claude continued to be a strong advocate for it as a “sophistication signal.” Claude commented:

While the first three AI types are table stakes (80%+ adoption), Agentic AI is the sophistication signal. Organizations using Agentic AI consistently outperform across multiple value dimensions—not because Agentic AI itself creates all the value, but because deploying Agentic AI requires organizational maturity (measurement frameworks, governance, risk management) that also drives value in other AI initiatives.

We weren’t sure that any of these conclusions were true, and they were certainly overstated. The high value of AI that agentic adopters are getting is almost certainly coming from other AI technologies. But its comments were interesting, and we did hint at the “sophistication signal” in the report.

A Smart but Pushy Intern

Others have drawn this analogy before, but on our project Claude was akin to a very smart and pushy intern who saved us considerable time. We were able to engage in a dialog with it—in many cases an effective one. The LLM sometimes went too far beyond the data, and rarely backed away from the ideas it advocated. Once we did point out that it was over-generalizing, and it apologized for this behavior. But despite a few glitches, our work was definitely better as a result of having Claude on the team.

Will Claude replace a lot of data scientists or data analysts? I think the most likely effect of these capabilities will be the democratization of data analysis and some data science. More people will be analyzing data, not less. Some only moderately qualified practitioners may need to move into other fields. However, highly skilled analysts and data scientists will still be necessary. It’s occasionally important for somebody to tell Claude and its LLM relatives what to do, to figure out whether it’s making sense, and to push back on this pushy intern when it ventures too far away from the data.

Source: Jill Lepore, If/Then

Subscribe now

I’m reading Jill Lepore’s book If/Then about the origins of analyzing human behavior data with computers. One interesting aspect of it is the automation paranoia arising from the introduction of the IBM 704 mainframe computer in 1954 (the year I was born). The book even includes an image from an automation-focused campaign leaflet for John F. Kennedy’s 1960 presidential campaign—see it above.

Further research led to a website that discussed automation concerns, and a newspaper story from four days before my birth that could have been published today with minor changes in technology and gender wording (emphases from the blog post author, Matt Novak):

What future competition the officeworker will meet from the mechanical brain still seems to be in doubt. In the United States, live office employees are more than holding their own, the Labor Department reports. They now number about eight million, 64 per cent more than in 1940. The department predicts clerical employment will continue to expand despite all the automatic files, and cash resisters, adding machines and “thinkers” coming on the market.

However, the International Labor Organization has reported in a worldwide survey, the transfer of work from men to machines is proceeding in offices much faster than it did in industry. Office machine production in the United States especially is booming, sales being about four times what they were before the war. As, a result, says the report, many businesses, including especially banks and insurance companies, are reducing office personnel. The report predicts that in the long run world demand for such jobs will exceed the supply.

It is likely that the officeworker, while continuing in demand, will, like the horse and the bicycle, take on new functions as machines absorb more office routine. Hence those expecting to compete with the office robots will need more diversified education. They would do well to acquire a few skills that the robots cannot duplicate.

So the concern that computers and software are going to take our jobs has been around for a long time. I’ve been focused on it for more than a decade. At about this time In 2016 my old friend and co-author Julia Kirby and I were planning the launch of our new book. Called Only Humans Need Apply: Winners and Losers in the Age of Smart Machines, it was not the opening salvo on the question of whether and when AI would replace human workers. A year earlier Stanford lecturer Jerry Kaplan had published Humans Need Not Apply—suggesting that human jobs were about to become obsolete (they haven’t)—and futurist Martin Ford published Rise of the Robots, which posited that even burger flippers would soon lose their jobs (they haven’t). Even earlier in 2013, two Oxford researchers published a widely-cited paper called “The Future of Employment,” which employed technical analysis of tasks and AI capabilities to predict that 47% of US jobs were “automatable” with AI (they haven’t been automated).

Our book was one of the first to argue in favor of augmentation, but it didn’t sell terribly well. I think that this is a problem for optimistic books in general—at least none of my optimistic books have sold well, and I guess people prefer scary books over reassuring ones. And Julia is a great writer, so the book was definitely well-written.

But it turns out that in addition to being reassuring we were also correct—at least for a decade. We argued that augmentation of workers with AI and vice-versa was both the most likely and the smartest way to think about AI and jobs. Over the last ten years augmenting work with AI has been far more common than automating jobs out of existence. Even though we did not anticipate the rise of generative AI in our book, that powerful technology has helped hundreds of millions of people do their jobs more quickly and (in some cases) effectively, but by and large they are still employed.

In the book (and in an article preceding the book by a year) we also described five “steps” that motivated humans could take in order to effectively collaborate with AI:

• “Step in”—keep your regular job but learn how AI works and apply it to your tasks—I discussed it here recently

• “Step up”—oversee the application of AI to the business

• “Step aside”—pick a job that doesn’t require the use of AI

• “Step narrowly”—a job that could be automated, but is too rare to be worth the trouble

• “Step forward”—build and maintain AI systems.

We suggested a decade ago that most jobs would involve “stepping in,” which has indeed turned out to be the case. We also argued that “stepping aside” would become increasingly difficult as AI became more capable, which also turned out to be true. We were probably wrong, however, in advocating for “stepping narrowly”—it turns out that almost any job has tasks that can be automated, and it’s not that difficult or expensive to do it.

Aside from that relatively minor mistake, our guesses about what would happen in the decade after the book was published were more or less correct. But that’s what they were—guesses. All of the prognostications thus far about the impact of AI on the job market have been guesses as well, and as Miguel Paredes and I argue in a recent article, they’ve all been wrong.

We concluded that article with a call to fewer predictions and more careful observation and description about what’s actually happening in the workplace. Some of that is beginning to happen. The Stanford working paper called “Canaries in the Coal Mine,” suggesting reduced hiring of entry level workers in a couple of jobs, is a good example. We just have to be careful not to over-generalize such findings; there are some entry-level workers (e.g., digital marketers) who are doing pretty well.

Steve Miller and I did some observation of 29 AI-augmented workers in our book Working with AI. Things seemed to be going well for them and almost all didn’t think there would be large-scale automation of their jobs anytime soon. But as we noted in the heading of the last section of the last chapter, “If the Singularity Comes, All Bets Are Off.”

I will make one prediction, however. Getting back to the JFK ad above about automation, I am predicting that if the current level of hype, fear, and (much less) actual AI-driven job loss continues, that issue will be much more prominent in the 2028 presidential election than it was in the 1960 one. And it should be!

I am both a humanist (as a human how could one not be?) and a pragmatist, and thus far I have believed that limiting the pace of AI development would be both impossible to enforce and not necessarily a good idea. I still think it would be difficult to enforce, but I’ve changed my mind about whether it is the right thing to do. The drumbeat about the current and potential perils of AI has become louder and louder. Over the past few days, for example, there have been these stories in the online and mainstream press:

• When AI Bots Start Bullying Humans Even Silicon Valley Gets Rattled

• The Dark Side of AI Weighs on Tech Stocks

• Welcome to the Voyage of the Damned

• We’re All In a Throuple with AI

• More Than Ever, Videos Expose the Truth. And Cloud It, Too.

• How Fast Can AI Change the Workplace?

• Something Big Is Happening (over 85 million views!)

• America Isn’t Ready for What AI Will Do to Jobs

• AI Is Already Making Online Crimes Easier. It Could Get Much Worse

And these articles, all negative about the impact or potential of AI, were only in the past week. It no longer seems viable to stick your head in the sand about the negative impacts of AI on jobs, social life, political life, cybersecurity, non-AI software businesses, the environment, etc. Yes, there are some potential positive impacts as well—particularly in healthcare. I do think there is high potential to lower the cost and improve the quality of everyday healthcare, although there will no doubt be some negative impacts when AI freely dispenses health and treatment advice. And I would argue that current AI healthcare systems are already good enough to yield a major improvement in the quality and the cost of care if they were broadly implemented.

I’ve also been influenced by the increasing number of AI scientists who are trying to warn the world that the possibility for global damage from AI is too high. They include Nobelist Geoffrey Hinton, Turing Award winner Yoshua Bengio, Berkeley AI pioneer Stuart Russell, and most recently Google DeepMind CEO (and another Nobelist) Demis Hassabis. He warned in an interview a few days ago that agentic AI poses particular risks of taking actions its creators didn’t intend and could get out of control. He argued for international cooperation to “set minimum standards before existing institutions are overwhelmed.” Several other prominent researchers at major AI vendors have recently resigned for safety and human alignment concerns.

OK, I’m convinced. In short, I’m ready to put on the brakes on AI. The fact that AI systems are now improving themselves and generating their own code has been a big factor in my increased concern. I’d like to see researchers and vendors stop making Ai systems smarter. I’d like to see agentic AI restricted from coming up with its own goals, or even its own means of achieving those goals. I’d like to see regulations that prevent it from doing harm. Let’s focus on implementing the AI we have, which is already plenty intelligent to power improvements in many business and organizational processes.

How to Slow It Down?

The obvious answer to slowing AI down and making it safer is government regulation. I don’t see the current US federal government issuing any AI safety regulation; Congress is dysfunctional, and the Trump administration is hell-bent on eliminating all regulation and beating China on AI. In the US, that means that states and even cities will have to regulate AI. It’s also obvious that having a patchwork of state- and city-based regulation is nutty, but until there are major changes in the federal government that may be the only regulatory remedy. California has already passed some AI legislation, one component of which focuses on the safety of foundation LLMs. Since OpenAI, Google, and Anthropic are all headquartered in California, maybe that legislation will have some impact. It hasn’t thus far, however. New York City has also enacted legislation about the use of AI in hiring (Local Law 144), but thus far it appears not to have been enforced.

It’s not surprising that if state regulation of AI is the only alternative in the US now, AI vendors and investors are lining up to fight it. Meta, for example, just announced that it was giving $65 million to two political action committees—one for Democrats, one for Republicans—to support anti-regulation candidates. Last summer, vendors and AI investors put even more millions into Leading the Future, another anti-AI regulation PAC. To their credit, Anthropic seems to be the only major AI vendor that is still advocating for regulation, although I’m sure they would prefer that it were federal rather than state-based.

Max Tegmark, the MIT physics professor who leads the Future of Life Institute, said at a Davos panel (video here) that he believes that openness to AI regulation is increasing. I hope he’s correct, but I am not quite as optimistic.

On the same panel, historian Yuval Hariri argued for an international agreement not to recognize AI as legal persons. That would mean that AI systems can’t have bank accounts, can’t own property, and can’t form corporations on their own. He feels that if even one country allows personhood, the global structure preventing AI-wreaked havoc would break down. Avoiding AI being corporations seems to me a good idea, but I am not sure we still have a structure to create and enforce international agreements. It certainly wouldn’t hurt, however, for the UN to take up the issue.

There is also the possibility of a consumer revolt—particularly in the US, where most citizens don’t trust AI already. The industry can only thrive if people use its products, and once potential consumers realize that AI can cost them their jobs, hurt their children, increase crime, nurture political conflict and even war—and perhaps kill us all, though I am not yet ready to envision that possibility—they may stop using it. Lack of use would perhaps slow down the whole industry and the somewhat crazy valuations and investment levels in AI vendors and data centers.

On the corporate side of AI consumption, I’ve been doing some research recently on the economic value that companies receive from AI. I’ll publish more about it soon. It does suggest that companies are starting to receive substantial value from AI, but only about half of the respondents in a global survey say they are getting “a great deal of value.” One interesting finding is that they say they are getting more value from analytical AI (which is less scary) than generative AI by a substantial margin. But if companies don’t get sufficient value and start spending less on AI—and the survey suggests that only a few plan to spend a lot more—that could slow the pace of AI research and development considerably.

As I mentioned at the beginning, I’m a pragmatist, and I’m not sure how an AI slowdown can be accomplished. But I am increasingly sure that it is desirable or even necessary.

Source: Gemini NanoBanana

Three years ago or so I wrote this passage as part of a foreword for John Thompson’s excellent book Data for All:

In 1998 I saw the movie Enemy of the State with Will Smith in the starring role and an excellent cast overall. I highly recommend seeing the movie, even if you are not interested in data privacy. But if you are, it raises a particularly interesting set of issues. Will Smith plays a lawyer named Bobby Dean. He accidentally ends up with some highly damaging information about a government intelligence agency. A corrupt member of that agency is able to use all of the data about Dean to track him down instantaneously, tap into all his calls, plant all sorts of false information about him, and generally make his life miserable.

In 1998 that wasn’t really possible; data couldn’t be captured, integrated, and acted upon instantaneously. Today, however, the Will Smith character would have left a much richer data trail, and the communications and computing power available to the intelligence agency would make it relatively easy to track and find him (and perhaps to shoot him dead with a drone!).

Things have only gotten worse since then. Over the last couple of days I read the story of Nicole Cleland, a Minnesotan who is justifiably disturbed by the efforts of ICE over the past several weeks in Minneapolis. Cleland is among a group of volunteers that employs non-violent observation of ICE agents to ensure that their activities are lawful (which, you may have heard, they often are not). On January 10 she was following a truck she suspected was being driven by ICE agents. The truck stopped, an ICE agent in camo uniform got out, and told Cleland that he had identified her through facial recognition. He accused her of impeding his work and told her she’d be arrested if she continued her (perfectly legal) observation.

You may have heard that three days later, Cleland received an email from the Transportation Security Agency telling her that her Global Entry status was being revoked. It gave no reason but listed several possible explanations, including violation of customs and immigration laws.

This is not an isolated incident. A Maine woman was capturing ICE activities on video in Portland, Maine in January and was told by an agent that he had photographed her license plate because she was a “domestic terrorist.” This is becoming a common strategy for the US federal government, as described in a December Cato Institute blog post by David J. Bier. The Cato Institute has strong libertarian leanings, and you can bet that if they complain about something it’s a violation of basic democratic rights.

The Palantir State Got Us Here

As I suggested in my comments about Enemy of the State, the government didn’t always have the ability to identify and punish its citizens for legal dissent. But over the past decade or so—including during the Biden administration—technology from companies like Palantir has been increasingly used to integrate US citizen and resident data so that the government can punish individuals it doesn’t like. It’s not Palantir’s fault that the government is misusing the data, but it should now know that its data integration and analytics capabilities are being put to nefarious purposes.

It’s only in the current Trump administration, however, that such identification and punishment has become common. Trump and his appointees have taken a desire to control illegal immigration well beyond that somewhat reasonable objective. It’s now possible to punish perfectly legal protest activities with denial of easier travel, stopping law-abiding citizens at borders, auditing tax returns, etc. It seems likely that locking up such citizens—and perhaps killing them in the process, as Renee Good’s and Alex Pretti’s fates in Minneapolis illustrate.

I believe it’s fair to assume that the majority of Americans don’t want this kind of behavior from their government. Even Republicans and the MAGA faithful must realize that they too could fall prey to this drastic overreach. The Biden administration may have paid for contracts for data integration to Palantir in the name of efficiency, but once the inter-agency citizen data is available, it can used to punish any kind of enemy.

US law does not incorporate any citizen data privacy regulation that prohibits sharing of data across federal agencies, or that defines misuse of such data. The last major data privacy legislation was passed in 1974, and as Tressie McMillan Cottom describes in a recent New York Times editorial, we haven’t made any significant progress since then despite some attempts. As she rightly points out, the ICE phone cameras are almost as dangerous as their guns. The private sector’s data collection is scary, but it is of minor importance compared to what the US government is doing.

Subscribe now

Source: NanoBanana

There are plenty of discussions about whether the US is in the lead in the creation of generative AI models. The Wall Street Journal recently published an article, for example, suggesting that Chinese AI developers fear they can’t keep up with the US because of a shortage of high-powered GPU chips. This may be true, although I suspect that any such shortages will lead to some creative development of smaller models, a la DeepSeek. But that’s not my primary concern these days.

I think the US will do fine in AI development, although my research suggests that organizations seeking AI capabilities should look more broadly than the US and China. I am more worried about the relatively slow pace at which US companies seem to be adopting AI to support their existing operations and enable new ways of doing work.

Survey Data on AI Value

This concern was sparked by a recent survey (sponsored by Scaled Agile’s AI Native training and done in collaboration with Laks Srinivasan, CEO of the Return on AI Institute) of 1006 business and government leaders in eleven countries around the world. The primary focus of the survey was to better understand the economics of AI and develop an AI economic maturity model that could guide organizations in getting more value from the technology.

One striking finding is that in terms of getting a “great deal of value” from AI, the US ranked near the bottom. Only 38% of executives surveyed said their organizations were getting a high degree of value. Canada and Italy were tied at this low percentage.

Who’s winning the race to get value from AI? Mostly European companies, it turns out. Germany in particular had the highest percentage claiming a great deal of value from AI, with 56%. The UK was at a respectable 50%, and Spain and France were in the 45% range.

Asian countries were also pretty good in terms of value, with Australia, Japan, and the UAE all coming in at 50% or a bit above it.

Not the First Time

It isn’t terribly surprising that the US would lag in adoption of a new technology. There have been several other recent technology advancements that were adopted much faster in other countries than in ours. Take, for example, mobile electronic payments. China’s use of that technology is roughly ten times that in the US. We’re also making slow progress in the use of digital health technologies—I am ashamed to say that we still use fax machines in many healthcare settings. China and Southeast Asian countries, on the other hand, have hundreds of millions of patients using intelligent diagnosis and treatment applications. How about robotics in manufacturing, electric vehicles, electronic voting, smart home technology, universal broadband, etc.? We’re far down the list of early adoption on all of them. Don’t even get me started on high-speed rail.

Why are we so dilatory on employing many new technologies? Maybe business owners are too conservative, or maybe our consumers aren’t open enough to trying new things. But mostly I blame our government, and not just the current administration. We often either don’t support new infrastructure technologies, politicize their usage (as with electric vehicles and renewable energy), tie them up in regulatory knots (see Dan Wang’s new book Breakneck contrasting China and the US in this regard), or worry too much about government’s ability to correctly pick winners with industrial policy. In a few cases where the US government has taken an active role in mandating or subsidizing adoption of a new technology, we’ve done pretty well. Notice, for example, that basically every US hospital and medical practice these days has shifted to electronic medical records. That’s because the US government heavily subsidized their adoption with the HITECH Act of 2009.

What Are the Implications?

The point is that if companies within a country don’t adopt a particular new technology, it’s less likely that the country’s vendors can maintain a long-term lead in that technology. We don’t have strong robotics companies in the US in part because US companies are slow adopters of robotics. Electric vehicle manufacturing is doing better in China than the US because many more Chinese consumers are buying them. There are exceptions to this rule, of course. TSMC in Taiwan does quite well with semiconductors even though Taiwan is a very small market for its output. The same with ASML and the Netherlands. But as a rule of thumb, a vibrant and large domestic market for your products is a good way to stay in business, particularly when trade wars are raging as they are today.

I’m not the only person saying that AI adoption (versus creation) is a problem. Satya Nadella, speaking in Davos at the World Economic Forum, said that there will be an AI bubble if non-tech firms don’t make more use of AI. He didn’t speak about the US situation specifically, but clearly we have a lot of non-tech firms.

And back in October of last year, Salesforce CEO Marc Benioff bemoaned the slow pace of AI adoption by customers as well. “The speed of innovation is far exceeding the speed of customer adoption,” he said in an interview. “These customers have to go back and modify massive architectures they have and systems they’re running.” Benioff may be using slow adoption as justification for the mediocre success of its AgentForce agentic AI product, or for the decline in the company’s stock price. But he’s probably correct about the ratio of innovation to adoption speed, at least in the US.

Of course, some industries are adopting AI faster than others. Our analysis of the same survey data found that the IT/tech sector is the highest in terms of percentage achieving a great deal of value from AI, with 71%. Other industries with high levels of value include financial services (60%) and healthcare (58%). All of these have lots of data for AI to analyze, and a strong need to either sell more AI or improve internal business processes with it. Agriculture has the lowest percentage reporting a great deal of value, with only 15%. Although a majority (72%) of respondents in the agriculture industry said their company has adopted generative AI, it hasn’t led to much value yet. Other industries with relatively low levels of reported high value include hospitality (26%), construction (23%), and retail (21%). Other industries fell between these high and low extremes.

How do we get American companies to use more AI? Vendors and consultants can help by documenting how value is really achieved with the technology. A modicum of support from governments in the form of tax credits or incentives might also help. I would even argue that mild regulatory interventions by government (federal, not 50 states with different regulations) would help—consumers would trust AI more, and buy more products and services with AI built in. I try to do my part by focusing not so much on how many billions of parameters are in the latest LLM, but rather on how to use and manage the technology. It would be great to see other analysts and authors doing more of that too.

Subscribe now

For as long as Microsoft’s Copilot AI has been available, I’ve been less than enthusiastic about it. I thought it was just a personal feeling. But since I decided to research this post I discovered that my “meh” assessment is broadly held. A quick Google search with the topic “people who hate Copilot” yields a litany of discontent. Google’s AI Overview (admittedly not an unbiased source) reports,

People dislike Microsoft Copilot due to forced integration, data privacy concerns, poor quality/inaccurate responses (hallucinations), disruption to workflows, lack of control (can’t easily disable), generic output, security risks for businesses, and frustration with its inability to handle simple tasks or feedback gracefully, with some feeling it hinders learning or adds bloat rather than value.

It goes to cite a broad list of “Common Criticisms and Concerns,” and includes a section on “Who Hates It?”

My Own Private Meh

I have several reasons for my own Copilot antipathy. First of all, it feels like generic food in the grocery store. I thought that one of our largest and most profitable tech companies should be develop its own language model rather than white-labeling ChatGPT. Three years after that technology entered the world, Copilot still relies primarily on GPT-4 and 5.

Then there was the primary purpose of CoPilot, which appears to be to aid people in the use of Microsoft’s Office suite. Is the best that we can do with this breakthrough to create more Word, Excel, and PowerPoint outputs? It makes CoPilot seem like an advanced version of Clippy. It’s better than Clippy, but perhaps not enough to get excited about.

I have used Copilot on occasion, but it often doesn’t seem to work for me. For example, I just tried clicking on Word Copilot (there are many versions), and I got this helpful message: “AADSTS7000024: Inconsistent broker application IDs asserted by incoming credentials.” Maybe that’s code for “I’m not going to help you write an article that attacks my very existence.”

The Value Issue

But my greatest objection to Copilot is that it contributes to the lack of value for corporate customers that may bring down the entire AI-driven economy. The problem is that Copilot’s aids to personal productivity enable a benefit that is virtually impossible to measure. The savings in creating Office documents is typically a few minutes here and there. Even in the unlikely scenario in which those saved minutes are aggregated—one company I studied did say they had saved more than ten million minutes—but what of it? What are individuals and companies doing with those saved minutes? No organization I’ve come across seems to know.

I spoke with a data and AI leader in a large bank today about this. His organization has a well-defined approach to measuring the value of AI use cases. But he said he struggles with measuring the productivity benefits of Copilot usage. “Maybe they save a few minutes in writing an email or creating a spreadsheet,” he said. “But how do I know that they’re not using that time to check their social media feeds?”

One reason why Copilot is common within large organizations is that it often comes along with the Office 365 deals that Microsoft has with some of them. It’s very easy for an IT organization to say, “We’ve got generative AI covered—we have Copilot licenses for everybody.” But that, of course, doesn’t mean that the organization will get economic value from those licenses. And if the organization has to pay a per-user addon for Copilot—often $30 depending on the Office license—it can add up to a substantial expenditure.

There Are Better Alternatives

When I talk to AI-savvy individuals, I often ask them what language models they use. They never say Copilot. If you’re serious about using AI as an individual, I’d go with ChatGPT, Gemini, Claude, etc. (not necessarily in that order). I don’t recommend Copilot unless you are a very dedicated Microsoft customer who spends all of your time in the Office suite.

The same applies for businesses. I am a fan of enterprise-oriented genAI use cases over individually-focused ones, as this article describes. I never hear an AI leader say that they have built their enterprise AI application on Copilot.

Like most of us, I have spent a lot of time using Microsoft products over the years. Even though I sometimes have problems with their products, I think they are a pretty well-managed company overall since Satya Nadella ascended to the CEO job. I especially admire what they have done with the Power Platform to encourage citizen development (or AI-driven vibe coding, although I think that there are better AI alternatives for that too). But Copilot feels like the Microsoft of old—conservative, slow, and second-rate. It’s the Bing of AI.

Source: ChatGPT, used in moderation

This is the season of holiday overeating and over-drinking—despite the fact that moderate consumption of food and alcohol is widely believed to lead to a better life. Although I sometimes agree with Oscar Wilde in advocating “moderation in all things—including moderation,” I am beginning to think that AI—particularly the generative variety—is no different than food, alcohol, or other good things that become problematic when used excessively.

Don’t get me wrong; using genAI in moderation is fine—even a virtue. It can make us more productive, stimulate our thinking, and reduce our drudgery. I think virtually everybody should learn how to use it.

But three years into our post-ChatGPT lives, it’s becoming apparent that using it too much isn’t good for us humans. There are several different aspects and implications of excessive AI use. I and others have written, for example, about deskilling that can result from excessive use. We don’t yet know exactly how much deskilling results from how much AI use, and it will probably vary by the context. But we should be thinking now about how much AI is harmful to professionals in various fields, and what we can do to prevent that harm.

It is also increasingly obvious that some people can develop emotional dependence on AI, typically in the form of chatbots. And while it is perhaps better to talk to a chatbot than to talk to no one at all, there is evidence that conversing with AI on personal topics tends to increase loneliness rather than reducing it. I suspect that more studies will further reinforce this finding. If my kids (or more likely in my case, grandkids) were going to use an AI chatbot, I would certainly emphasize moderation.

At work, genAI is a seductive aid to creating content faster and with less mental engagement. The “faster” benefit might be okay, except for the fact that AI-generated work is often not of high quality and creates a need for others to improve it. I wish that I had coined the term “workslop” to describe this phenomenon. And the “less mental engagement” benefit means that the AI user may not remember much about what has been created. If you believe that work is in part about learning from experience, you may be out of luck if genAI does much of the work.

In fact, I would not be surprised if there is someday a scientific study suggesting that heavy genAI use is associated with increased risk of dementia. I have no evidence for this hypothesis, but there is some logic to it. A Mayo Clinic study, for example, found that using your brain for mentally challenging activities helped prevent or postpone the onset of mild dementia. Creating original content is mentally challenging; asking ChatGPT or Claude to do it is not (much).

My favorite relevant research study (actually a set of studies summarized in this article) involves a group of elderly nuns who were studied for 35 years, and then allowed their brains to be studied for evidence of dementia. More education was correlated with lower risk of dementia. Even more to my point, how the nuns wrote seemed to help them avoid dementia. The nuns had written autobiographies before taking their vows, and both the idea density and grammatical complexity of their writings were found to be associated with less mental impairment in old age. Imagine what might have happened to their brains if they had let genAI do their writing for them!

Of course, this is all supposition on my part, but there is perhaps more reason to believe genAI use only in moderation can help prevent dementia more than the widely-advertised over-the-counter drugs that some people take for the same purpose. Just sayin’.

Subscribe now

Source: Google Gemini, which can’t seem to explain what “Sothins” are

Many observers of AI are questioning the level of value that it provides to organizations, particularly for generative AI. The notorious MIT NANDA study—often described as the “95% zero value” study—has an odd methodology, but its primary finding is that tactical, individual-level, “broad and shallow” implementations of generative AI don’t provide substantial value to the companies that employ them. Enterprise-level, “deep and narrow” generative AI projects that are consistent with a company’s business strategy, on the other hand, do often yield measurable value.

The most extreme version of strategic AI is found in a small but growing number of companies might be described as “all in on AI” or “AI first.” These companies are aggressively pursuing strategic returns on their AI investments. They are using the technology to enable new strategies, new business models, and dramatically new ways of performing their business processes. While they represent a low percentage of companies, they are providing trailblazing examples for the majority of companies that are more conservative.

Consider, for example, such companies as Ping An in China, DBS Bank in Singapore, and Sanofi in France. All are currently involved in, or have already been successful at, the adoption of AI at strategic levels in many parts of their businesses. Ping An has employed AI to enable and empower five rapidly-growing ecosystems in insurance, banking, healthcare, automobile services, and smart cities. DBS uses AI to create “joyful banking”—a pleasant, digital way to conduct banking transactions using a chatbot built with extensive refinements—as well as to reduce fraud and transform internal processes. At the pharmaceutical giant Sanofi, both analytical and generative AI are being harnessed to transform the process of drug development.

Contrast their approach with the goal of tactical returns on AI investments. In this context, returns from AI projects mean incremental gains on existing tasks or processes. Perhaps some customer calls are diverted to a chatbot or intelligent agent, or the company is able to use AI to extract information from documents that previously required human eyes and typing. In order to achieve substantial value with this approach, there would need to be widespread adoption of AI across the company. Even that broad approach will achieve incremental benefits for the overall company, rather than any specific function or unit.

Strategic use of AI takes advantage of all the technologies available. In the late 2010s and early 2020s, the focus was on traditional machine learning on structured data—sometimes called “analytical AI.” For example, Morgan Stanley Wealth Management began working in the early 2010s on a “next best action” AI system to recommend personalized investments to clients, which it deployed beginning in 2017. In late 2022 with the announcement of ChatGPT, many companies began to explore generative AI. However, Morgan Stanley began working with OpenAI in early 2021 to build the “AI@Morgan Stanley Assistant.” It made the company’s own knowledge widely accessible to advisors and their teams, and the application went into production deployment in 2023. Since then Morgan Stanley has created several other production use cases of generative AI. In 2025, many companies are beginning to address strategic use cases of agentic AI, which not only provides content to users but can take action on digital transactions. Morgan Stanley plans to roll out a AI agent for customer service in late 2025. The most strategically-focused companies with AI are now employing all of these AI capabilities.

The same distinctions between strategic and tactical use, however, apply to all types of AI technology. With analytical AI, RoAI Institute research presented in a 2021 MIT Sloan Management Review webinar noted that less than 15% of companies investing in AI are getting meaningful returns. Several other surveys found low levels of economic value from AI. A 2022 Deloitte survey in large, global organizations found that the most common objectives pursued with analytical AI were process improvement and better decision-making—worthy goals, but normally tactical.

More recently, the same pattern is taking place with generative AI. Most organizations are taking an experimental approach to the technology, with tactical and incremental productivity as the objective. At the beginning of 2024 only 6% had an organizational use case in production; a year later at the beginning of 2025, the percentage increased to 20-25%. Even when companies are experimenting, the experiments are typically casual and unstructured. They seldom change business processes, measure value, or address strategic priorities. A McKinsey survey in early 2025 found that only 1% of companies in developed economies described their generative AI implementations as “mature.” Another McKinsey survey later in 2025 found that only 21% of large organizations had redesigned some workflows to take advantage of AI, but that workflow change was the attribute of AI most closely correlated with economic value.

Whatever the specific form of AI, if your return on AI is non-existent or all tactical now, our research points to success factors and approaches organizations can take to maximizing returns using a more strategic and systematic approach.

Strategic vs. Tactical Return

Both strategic and tactical approaches to AI returns can be useful, and some firms are aiming for both. However, the processes for proposing and realizing them are very different. Achieving strategic returns, by definition, involves strategy and its implementation. The idea to use AI in a particular setting is part of the organization’s overall strategy for how it goes to market and succeeds with customers. Strategy is usually articulated by senior executives, and if AI is going to be integrated with strategizing, senior executives need to understand all types of AI and what can be done with them. If there is a strategy organization, its members also need to be conversant with AI and its application.

Strategic implementations of AI often cost more, take longer, and involve more people than tactical ones. They usually involve major change to business processes and business capabilities. Perhaps they are the basis of a new product line or a new relationship with customers. That requires careful preparation across multiple parts of a company. Front-line employees who will use the system need to be educated in the use of the system, and perhaps will need to develop other new skills. The rollout of the technology and the associated business changes needs to be carefully planned.

Tactical returns from tactical AI projects, not surprisingly, generally involve easier implementation. Because they involve incremental change, they don’t need a lot of senior management deliberation. Changes to business processes are usually relatively minor. In many cases, AI is being used to perform a task that is already done by humans. If that leads to substantial reallocation of labor, human employees would need to be trained on their new roles. And in most cases there is still some need for integration of AI capabilities with existing information systems and architectures.

However, because tactical projects involve less commitment, they are more likely than strategic AI initiatives to start as proofs of concept or pilots and never get deployed into production. And as is perhaps obvious, no deployment means no value. Tactical projects also often employ the “so-so technologies” described by economists Daron Acemoglu and Pascual Restrepo. These are technologies that displace workers from jobs, or aspects of their jobs, without offering significant productivity benefits. One example they cite is inexpensive, off-the-shelf chatbots, which they argue displaces workers to some degree, but provides neither high-quality customer service nor significant gains in productivity.

Which to Pursue?

Strategic applications of AI are far more valuable to an organization than tactical ones if they work as planned. They can vault an organization into a new business model or strategy, and can yield high levels of benefit. For example, in a 2021 MIT Sloan Management Review/Boston Consulting Group survey, respondents who use AI primarily to explore new ways of creating value were 2.7 times more likely to improve their ability to compete with AI and capture opportunities from adjacent industries than those who use AI primarily to improve existing processes. In the 2022 Deloitte survey mentioned above, the most successful companies with AI were more than three times as likely as those with lower levels of value achieved to use AI to address new customer markets and segments, enable new products/programs/services, and support new business models.

Of course, achieving those returns is more difficult, as well as more difficult to measure. For an AI-enabled strategy to go right, many things have to fall into place, including the factor of competitor response. If they emulate the AI and business changes, the strategy could fall flat. And when a strategic use of AI does succeed, there will be many who want to claim the success. Attributing a specific amount of return to the AI itself is difficult with multiple aspects of business change being wrapped into the solution.

Tactical AI projects provide less value, but are easier to measure. The primary benefit they offer is productivity gains, particularly if they replace human employees. At a time when many companies struggle to find enough workers, tactical projects can also fill an important labor gap. When human workers are not eliminated but deployed onto other types of tasks (with reported savings in minutes or hours), measurement of value is more difficult. Elimination of outsourcing is another common source of value from tactical AI—particularly from automation projects.

Choosing strategic vs. tactical projects and returns is often dictated by the management and financial situation within the company that will implement AI. Are senior leaders aware of what AI can do and motivated to employ it aggressively? Is the strategic intent for AI clearly defined and communicated across the company? Does the company have the human and financial resources to undertake a large-scale AI project? Is there a potential strategic benefit from AI done well? Is there a deep experimentation mindset across the company? If these factors are in place, strategic projects will be a natural step to take. If they are missing, it is unlikely that any strategic project will be successful, and a company should “cut its teeth” on more tactical AI. At a minimum, however, companies should undertake tactical AI projects with an eye to how they might fit together to make possible a broader strategic objective.

Companies desiring strategic returns can start with significantly increasing the AI fluency of their top leaders including the board, CEO and management teams. Top leaders with a deep understanding of AI will position the company to ask the right questions about how AI can bring about a transformation in its strategy, business model, or operations.

Second, identify the levers required to achieve those desired outcomes from AI. These levers many include changes in business strategy, leadership, culture, AI talent, organization, and technology. It is helpful to baseline assets and capabilities and to pinpoint key gaps. Then, systematically identify and evaluate alternative strategic initiatives, and tactical projects that could eventually produce strategic value, around a coherent vision for generating return.

Third, companies should conduct a series of stakeholder sessions to align the top-down strategy with the bottom-up execution to create organizational buy-in. This organizational alignment will help to develop the right AI roadmap with components of action and corresponding investments. Finally, companies will need to monitor whether the desired strategic change is taking place, and whether capital resources are being effectively allocated. Remember that strategic returns are measured in business change, not incremental savings. These deep changes to an organization won’t happen overnight, but the investment and effort over time to achieve them can position an organization for a long and successful future.

Source: ChatGPT

There has been a rise in concern about de-skilling and loss of critical, thorough thinking in various fields because of AI. I have written about aspects of this issue, and Princeton professor Kwame Anthony Appiah wrote about it more generally in this Atlantic article.

I am most concerned about this issue in healthcare—perhaps obviously, because de-skilled clinicians could endanger human health. The other general reason why healthcare professionals might let AI de-skill themselves is that they are under considerable pressure to be more productive—at least in the U.S. The average patient visit with a physician is only about 15 minutes, and there are temptations galore to take AI-enabled shortcuts. If AI can easily summarize patient visits in clinical notes, diagnose diseases and communicate with patients about them, and help navigate highly bureaucratic insurance processes, why shouldn’t doctors and nurses take advantage of this capability?

My Family’s Experience with AI Shortcuts

My wife and I have experienced several issues with shortcuts (or the absence of them in one case, I think) in the medical field over the last several weeks.

1. The first event was with my primary care physician (PCP) in California. I had experienced an upper respiratory infection for several weeks that just wasn’t getting better. I was beginning to think that I had a sinus infection, but wasn’t sure if it was viral or bacterial. So I emailed my PCP about it, asking whether I should take something or see a doctor (I was far away from him at the time). He almost immediately emailed me back, and the message began:

Tom [in a different font than the rest of the message],

Thanks for reaching out about the patient with a lingering cough after a recent viral illness.

The emailed response went on to tell me that I should just wait out the infection and cough. But it certainly appeared that my doctor had used generative AI to prepare his response to my message, and that he hadn’t bothered to edit out the incriminating introduction. I asked him, “So what AI model are you using?,” and he admitted to using OpenEvidence, the fast-growing “medical information platform.” I don’t object to my doctor using that tool, but I would like to think that he read it enough to edit out the part revealing that the message didn’t really come from him.

2. Shortly thereafter, I was at a conference on AI in Boston. One of the speakers was Dr. John Halamka, a former prominent medical CIO in Boston and now head of the Mayo Clinic Platform. It is gathering patient information from a variety of sources and incubating healthcare startups, particularly in AI. One of its incubated companies is OpenEvidence. In his talk Halamka described a glowing future for healthcare AI, and mentioned that physicians were becoming increasingly dependent upon it. So I asked a somewhat cheeky question when audience Q&A time came around: “How does this AI-enabled future relate to the events of November 13, 2002?” I guessed that Halamka would know what happened on that date, and he did—even remembering the exact time of day it happened. He was CIO of Beth Israel Deaconess hospital system at the time, and the entire computer network at the hospital went down—for four days. During that time the hospital and its employees couldn’t use any electronic medical records, order labs or procedures electronically, or use any sort of decision support. Were it not for some old paper forms—some accounts suggest they were taken out of a dumpster—and some old modems, all care processes might have broken down. Halamka said in answer to my question that those events point out the importance of “business continuity” approaches, but what happens when people forget how to diagnose diseases and write up clinical notes?

3. My wife and I had patient checkups a few weeks later at a different health provider in Massachusetts. Like many clinical practices these days, the physician makes use of “ambient AI” to capture all conversation in the exam room. Fine, I thought, and we agreed to the use of it. However, the clinical notes for our visits were both inaccurate. My wife’s said that she had chronic fatigue, sleep problems, and a child enrolled at Northeastern University—none of which she said anything about. My notes said that I was visiting the doctor for a chronic cough (yes, the one I mentioned above—although it had passed several weeks earlier, and I made the mistake of telling this doctor about the AI-generated diagnosis message) and shoulder pain, which the doctor mentioned from my California patient records but also was resolved more than a year ago. As a result of this text in the clinical notes—which appeared not to have been reviewed by the doctor—our visits were coded as medical problem visits rather than routine checkups, which resulted in $1200 of medical bills. We’re still trying to get that problem reversed.

4. My last story was not a medical shortcut (to my knowledge) but an admission that they are a problem. I had a routine colonoscopy a couple of weeks ago (I recommend the pills for the awful prep process rather than the exceedingly nasty liquid). It went fine, but as I was chatting with my gastroenterologist before the procedure I asked him if he planned to use AI to identify potential polyps. He was quite familiar with the de-skilling issue revealed in this article, which found that gastroenterologists’ ability to identify polyps declined after only 3 months of AI use. He said he’s worried about it as AI use among doctors performing colonoscopies grows quickly. My doctor’s operating room did have a screen using AI, but he said he tried not to rely on it to identify polyps. Of course, I was sleeping during the procedure, so I have to take his word for it.

So How to Avoid De-Skilling?

These were all relatively minor problems, and our health was never in danger. But it’s easy to see how big problems could result. What can we do to avoid AI-related deskilling? Some degree of regular recurrent training in non-AI-based clinical decision-making—perhaps making use of simulations—may be necessary. Doctors do already have continuing education requirements (particularly to retain board certification in specialty areas), but they don’t generally address this de-skilling issue.

There is another field that has a similar problem—aviation. De-skilling because of automation in aircraft has already been blamed for at least one horrible airline crash. Commercial pilots do have regular recurrent training, but their simulators often use the same automated systems that real planes do. The de-skilling issue is probably greater for general aviation pilots, who have no recurrent training and whose planes are being increasingly automated over time.

In both cases, people can die if AI or automation leads to the loss of critical skills by those in charge of safety. Given the rate at which AI is advancing, professional associations in these safety-critical domains had better move fast to address the problem.

Subscribe now

Replica of Thoreau’s cabin on Walden Pond

“I went to the woods because I wished to live deliberately, to front only the essential facts of life, and see if I could not learn what it had to teach, and not, when I came to die, discover that I had not lived. I did not wish to live what was not life, living is so dear; nor did I wish to practice resignation, unless it was quite necessary. I wanted to live deep and suck out all the marrow of life, to live so sturdily and Spartan-like as to put to rout all that was not life, to cut a broad swath and shave close, to drive life into a corner, and reduce it to its lowest terms...” Henry David Thoreau, Walden, or Life in the Woods

I am fortunate enough to live in New England for half the year, and fall is the best time to be here. Last weekend my wife and I went out to Concord, Mass. and revisited (and circumnavigated) Walden Pond. That’s where Henry David Thoreau lived in a small cottage in the woods for a couple of years in the 1840s. There’s a replica of the cottage there today, and it’s about as spartan as you can imagine.

What would someone like Thoreau think about AI? It’s a pretty good bet that he would hate it. He’d probably think that if he’d lived much of his life in front of a screen—however intelligent its messages—he would discover that he hadn’t lived. Thoreau was primarily a writer, and were he to explore having generative AI do his writing for him, he’d feel that he actually hadn’t written, or learned anything from the process of writing.

Is there anything he would like about AI? Assuming he could get a wifi signal in his cottage, he’d probably be impressed by the amount of knowledge he could access. He loved to read; he wrote in Walden that “Books are the treasured wealth of the world and the fit inheritance of generations and nations.” If Thoreau were alive today, he might occasionally resort to genAI to summarize a book or one of his fellow Transcendentalists’ speeches.

However, I’m guessing that he’d generally say that a real book is far preferable to an AI summary of one. It takes longer to read, but the reader is far more likely to internalize, reflect on, and act on the content. Thoreau’s move to the woods—OK, there was a train and his parents’ house nearby, and he could pretty easily walk to them—was a signal that he believed some struggle—both intellectual and physical—was worthwhile. GenAI reduces or eliminates struggle, but there is a price to pay for that. The great article Your Brain on ChatGPT provides a good description of what that price is—little or no brain activity.

There is also little doubt that he would scorn the “996” (9AM to 9PM, 6 days a week) lifestyle of AI builders in Silicon Valley. He wrote about the workers in his time that “The mass of men lead lives of quiet desperation,” and even in the 1840s most people didn’t work that many hours.

Thoreau was also a surveyor for the town of Concord, and he painstakingly surveyed the circumference and depth of Walden Pond. Apparently that was a very difficult project; he did it in the winter so he could stand and walk on the ice and use various ancient measuring devices. But even if AI or some other advanced technology (satellite imagery+ AI?) could do those tasks much more easily, I doubt that Henry would have made use of them. No pain, no gain—and no reason to spend time in nature if you’re surveying in front of a screen.

Thoreau wouldn’t have liked AI much, but he would have really hated social media and the AI algorithms that drive it. If he spent so much time in the woods he obviously liked being alone; he wrote: “I never found the companion that was so companionable as solitude.” He would probably despise the social media-based companionship that tends to make us dissatisfied with our own lives.

My trip to Walden Pond did not make me want to live in a small cottage in the woods, nor to recommend that for anyone else. But it did make me reconsider being connected all the time and thinking too much about the latest advances in AI. Walking around the pond or similarly getting out into nature made me realize that living a physical life should at least supplement a digital, AI-driven one. Every so often I need to get out into the world, leave my phone and laptop at home, enjoy the natural beauty, and think about what it means to dwell on this lovely planet.

Walden Pond

Subscribe now

Source: ChatGPT

Yes, there is an AI bubble. AI-related stock prices are inflated, valuations of startups are inflated, there is too much marketing hype, data center construction is excessive, etc. All of these related bubblets will deflate, and probably soon. I’ll deal with the second question below, but the short answer is that we’ll be OK in the long run.

The AI Demand/Supply Mismatch

Why do I say there’s a bubble—and that it is likely to burst soon? The most important factor in the bubble is that demand is not keeping up with supply. All of the tech companies, both Magnificent 7 and AI startups, are on the supply side. The demand side is made up of non-tech industry corporations and individuals. Both are showing signs of weak demand. Some straws in that wind include:

• It's pretty clear that bottom-up experimentation with generative AI in large organizations isn't yielding much measurable value, which will eventually lower demand; the recent MIT NANDA study reports that 95% of companies are getting no measurable value from generative AI. While both the data and the methodology are problematic, I do think it’s directionally correct.

• The more enterprise-level use cases take substantially longer to implement than bottom-up experimentation, which will probably lower demand; This HBR article is a good summary of the reasons.

• For some of the more popular use cases such as code generation, it appears that the productivity gains are not quite was anticipated (primarily because of the need for human review); that should eventually lower demand;

• Some data suggest that the general public doesn't use gen AI tools very often. Benedict Evans’ annual presentation about AI eating the world depicts several surveys in which daily use tops out at about 10%, and weekly use at 25-30%. And most of that is with one tool—ChatGPT.

There is also too much supply being developed on the data center front. While there are still some improvements in generative AI that will come from scaling training data and processors, it would appear that those benefits are leveling off, and credible alternatives to highly scaled models (e.g., the new DeepSeek V3) are still much more efficient than the primary US-based models in terms of compute.

If these trends continue, we’ll have many unused or under-utilized data centers, just as we had enormous volumes of dark fiber after the Internet bubble burst. While data center construction is so massive today that it is contributing more to the US economy than consumer spending, the pace of capacity creation isn’t sustainable and this bubble will surely burst too. The heavy use of carbon-based fuels for these data centers—nobody seems interested in powering them with wind or solar anymore, and nuclear will be slow to emerge—is also hardly sustainable for our atmosphere.

There is also a marketing bubble with generative AI, in that vendor company leaders are talking a lot about artificial general intelligence (AGI) while still being a long way from achieving it. For example, GPT-5 has some impressive capabilities, but it is hardly more intelligent than humans on every topic. Among many other shortcomings, the new LLM’s inability to draw an accurately labeled map of North America screams “not ready to beat humans.” This mismatch between marketing and actual performance doomed IBM Watson and will doom OpenAI if they keep it up.

Should We Care?

I suggested above that we shouldn’t care whether AI is in a soon-to-burst bubble state, but let me qualify that. From a financial perspective, many of us have benefitted from the U.S. stock market rise that has largely been fueled by AI. So from that narrow angle, we will care if the bubble bursts. I am not a stock market prognosticator, but I think the markets will eventually recover. There is likely to be pain in the short term, however.

From a general economic perspective, I would argue that even if the bubble bursts everything will eventually work out fine. The current AI bubble is a bubble of substance, as was the Internet and e-commerce bubble 25 years ago. Even if stock prices and startup valuations decline, AI will continue to advance and be used productively by companies and organizations. It’s probably better for long-term economic growth and for the maturation of AI if the general public isn’t discussing AI startup IPOs at cocktail parties and speculating over lunch about whether data center developers can get enough GPUs.

In my hometown of Boston we’re seeing the impact of a slowly-bursting biotech bubble. Instead of dark fiber or dark data centers, we’ve got plenty of dark lab space. There are very few biotech IPOs, venture funding is down, and startups are going out of business. There has been a mild drop in employment. Fortunately, the region’s economy is still fairly diverse and successful—including AI businesses—and we will eventually see the biotech industry boom again. But it’s perhaps an early indicator of what might happen with AI.

Subscribe now

Source: GPT-5

We’ve been promised that GPT-5 was coming for a while now. Sam Altman, OpenAI’s CEO, had predicted that AGI (artificial general intelligence) would come in 2025, and we’re now 2/3 through that year, so we might have expected that GPT-5 would accomplish the elusive AGI (which has no clear definition, of course). But even according to Altman, it falls short of AGI, though he thinks it’s a step in that direction. He described it as like having “a team of Ph.D.-level experts in your pocket.”

I wanted to try it out and I struggled to think of what I might do with a team of Ph.D.s in my pocket other than needing bigger pants. I eventually came up with an idea: what if I got it to rewrite my aborted novel? This potential Great American Novel didn’t turn out to be that. I conceived of and wrote several drafts in 2019 and 2020. What else to do during a pandemic shutdown?

The topic—I hoped soon to be a major motion picture—was robot American football as practiced at MIT and its nerdy intercollegiate rivals. Driven by an epidemic of CTE, the world has decided that football is no longer safe for humans. The MIT team members advance the robots toward greater autonomy until a robotic apocalypse occurs in the final game of the season. I thought it was a great idea, but reviews—even from my friends—were not complimentary. “Good idea, but it needs a lot of work” was the most common comment. If generative AI at its best could turn it into a great novel, that would put us well on the way to at least artificial literary intelligence.

So I had GPT-5 review and try to improve it. It made some mistakes—for example, saying it was going to interweave its ideas for new topics throughout the text, but then putting them all at the end, and repeating itself multiple times for plot ideas—but I can’t deny that it was helpful. I hadn’t tried this earlier with the various permutations of GPT-4, but I’m guessing GPT-5 was a little better than they would have been. At one point I had commissioned a human editor to take a pass through the manuscript, and I think GPT-5 may have been more helpful. That doesn’t mean it would be better than all editors, of course.

The other thing I must say about GPT-5 is that it is unfailingly willing to assist. It kept volunteering to do more and more, and I kept saying yes. I think I wanted it to say, “Why don’t I just rewrite the damn thing,” but it didn’t go that far. But it did provide extensive human relationship ideas (ironically), some ways to flesh out the villainous opposition to robot football, and helpful suggestions on avoiding too much sex—apparently the readers of tech fantasy fiction don’t like that, it confidently told me (could that be a hallucination?). It also told me that my technical speculations on what robots would be able to do in five years were a bit aggressive but within the realm of plausible tech fiction.

However, it did not feel like a team of Ph.D.s in my pocket, or at least not a team of Ph.D.s with expertise in modern fiction. I’d compare it to a smart B.A. graduate who majored in contemporary literature and was also a robotics hobbyist (there are so many of them!). Most of its suggestions arose out of the prompts I gave it. I had to tell it that I needed an image, and what it should look like (see above—pretty good but not sure why only one player has a number). I would be confident that a Ph.D. with the right speciality could do a good job of rewriting the book (I think that’s called “ghostwriting”), but it was clear to me that I still needed to be the literary integrator or uber-editor of all of GPT-5’s suggestions.

In other words, this feels more like a good upgrade to GPT-4 than a major breakthrough in generative AI capabilities. That seems to be the consensus of the other reviews that I’ve read. I’m somewhat grateful for this outcome, since I was concerned after Altman’s predictions that human writers, thinkers, and artists would pretty much be put out of business by GPT-5. But it seems that most of us are still fairly safe.

Mostly I hope that it will quiet the chorus about which genAI model is best, when we’re going to reach AGI, and when humans will become irrelevant. I’m not as negative about generative AI overall and GPT-5 in particular as Gary Marcus, but I do agree that we are going to need more than transformer-based models if machines are to beat humans on every dimension. I read a review of Alexa Plus last week from Amazon, and it suggests that while its conversational skills are better than the previous form of Alexa, the ability of generative AI to seamlessly carry out “skills” is worse than when deterministic AI was doing it. I’ve seen similar kinds of results from the early efforts at agentic AI. So I’m a fan of the idea that we will need probabilistic, deterministic, robotic, and many other forms of AI programming before we really need to start worrying about it taking all our jobs.

In the meantime, if you have any feelings about whether a novel on robot American football at MIT would be an interesting read, I’d love to hear them!

Subscribe now

For many years, my conversations with CEOs about AI-related job loss were similar to those involving Donald Trump. Privately they predicted doom, but publicly they said little or nothing. Kevin Roose of the New York Times wrote about this two-faced approach back in 2019 about the World Economic Forum discussions he heard. I worked with Deloitte on an anonymous survey of US executives back in 2018, and we found that 63% agreed that “To cut costs, my company wants to automate as many jobs as possible with AI/cognitive.” But at that time almost no CEO went on the record that AI would lead to substantial job loss in their organizations.

Things are changing rapidly in this regard, however. Several prominent CEOs have recently proclaimed that AI will eliminate jobs in their companies:

• Ford’s CEO Jim Farley predicted that AI would replace “literally half” of all (not just Ford’s) white collar workers in the US, adding that “AI will leave a lot of white-collar people behind.”

• Andy Jassy of Amazon told employees in June that he expected AI would lead to a smaller corporate workforce;

• Marc Benioff of Salesforce eliminated 1000 positions at Salesforce this year, and later commented that AI was doing 30-50% of the work at the company;

• JP Morgan Chase consumer bank CEO Marrianne Lake told investors in May that AI could drive 10% job loss among operations and account servicing employees at the bank.

• Vista Equity Partners CEO Robert F. Smith told a conference of investors that 60% of them would be looking for jobs next year.

And that’s not all—I could go on. While I suppose there is something to be said for honesty, I think these pronouncements are generally harmful to the companies and to the economy in general. They make me wistful for when the AI job doomers were restricted to expressing themselves quietly at ritzy executive gatherings.

For one thing, their predictions are likely to be wrong. I have reviewed perhaps twenty different predictions of AI’s effect on jobs, mostly involving losses but sometimes predicting gains. The predictions all have one thing in common: they were hugely wrong. We don’t have—and probably won’t have—good numbers on the actual jobs lost because of AI. But it’s pretty clear that we haven’t lost even the low predicted number (Gartner, with 1.8 million) and certainly not the highest predicted number (2 billion jobs, from an organization called Vista that I haven’t heard of since). It’s also obvious that we didn’t gain the 133 million new jobs from AI that the World Economic Forum predicted would arrive by 2022. In short, we don’t know the answer to this question, and making up numbers that are bound to be inaccurate just makes people (including CEOs) look bad.

But accurate or not, hearing from your company’s CEO that AI will put many of you and your colleagues out of work is also rather dispiriting to say the least. Maybe these predictions are designed to encourage mass resignation, but AI isn’t ready to take over everybody’s job who might resign. The most likely future is that many of us will be working alongside AI—my co-author Steve Miller and I discovered many people who were already doing this several years ago in our 2022 book Working with AI. We found that collaborating effectively with a smart machine requires some creativity and initiative on the part of human workers, and being told you’re going to lose your job to AI hardly encourages that type of thinking. I’m guessing that it leads to a lot of employees just going through the motions of working. Maybe the CEO AI job doomers are contributing to the lowest employee engagement figures in a decade—only 30% are engaged in their jobs.

These dire predictions may also create a social movement and a self-fulfilling prophecy. Other CEOs will feel like they have to be similarly provocative in order to keep their stock prices high (although there is no evidence that investors are impressed by these pronouncements). Eventually perhaps analysts and investors will expect large-scale job cuts because of AI.

At a recent conference I asked Daron Acemoglu, the MIT economist who won the 2024 Nobel Prize in Economics, if he thought that AI-based job cuts are the route to better economic performance. He said that is unlikely. “It leads to a race to the bottom,” he said. Your company cuts its costs, but so do your competitors. Everybody becomes less profitable. Far better, he said, to focus your AI investments on the ability to innovate and create better products and services. Incidentally, Acemoglu predicts relatively low job losses from AI.

I’m not sure what these doomsayer CEOs are thinking when they say that AI will lead to massive job losses in their companies and the broader economy. If lots of people lose their jobs, they won’t be able to spend much on these companies’ products. It’s the reverse of the Henry Ford strategy—he paid workers more in part so that they could buy his cars. Instead many companies are planning to implement AI in a way that may well bring down the economy.

Subscribe now

Source: Carbonite

Today is AI Appreciation Day! You are probably already celebrating or asking ChatGPT how best to celebrate. If you haven’t done that, I will save you the trouble. It suggests, among other things, putting on a “themed event” such as:

• Lunch & Learn: Organize a session featuring a speaker on AI ethics, innovation, or trends.

• AI Demo Fair: Encourage team members to present short demos or posters of AI projects they're working on or admire.

• AI Trivia or Game Night: Include questions about famous AI breakthroughs, movies, and researchers.

Or perhaps you could just read this Substack post in celebration! That would save you the embarrassment of telling your friends that you have organized an AI game night. Or that you even acknowledge such a Hallmark Holiday.

Like many of you, I suspect, I have mixed feelings about AI. I am concerned about its impact on human employment (see many of my other posts in this channel), our ability to think deeply and critically, our emotional health, our politics, etc. In these areas I don’t think AI has been terribly harmful thus far, but neither has it given us reason to celebrate.

I’m also not sure that we have reason to rejoice yet about the AI-enabled transformation of health care, but I do expect that will be among its most beneficial impacts—eventually. We desperately need AI to improve our health care system—to reduce errors in diagnosis, to make care more accessible and affordable, and to make available better treatments and cures. I read this week that healthcare is the largest employer in the US, but we are not getting the health and lifespans we need from all those hardworking people. In terms of AI, thus far we haven’t seen any of these potential benefits, at least in the U.S. healthcare system.

There are a few bright spots for AI, however, including:

• Radiology was one of the earliest adopters of AI in healthcare, and now, according to one survey, 68% of US radiologists make use of it. There are still issues involving trust in AI-generated diagnoses, figuring out how to get economic value out of the technology, etc., but progress is definitely being made. AI-based diagnostic pathology has been a slow follower of radiology, but it is also progressing.

• In the US we haven’t yet made much progress in making AI useful for the average patient who can’t afford or can’t get to a doctor, or who’s not sure whether to visit a doctor. In China, however, there are several of these “AI doctors” or “AI hospitals” that provide a substantial amount of care—from initial triage to diagnosis to treatment recommendations to getting a drug to the customer—with the substantial help of AI. I’ve mostly studied the Good Doctor app from Ping An, but there are several competitors. These systems are being used by hundreds of millions of citizens in China. I can’t wait until they come to the US.

• Generative AI has the potential to more easily accomplish what the Chinese AI doctor vendors have done. It is only a matter of time before there are gen AI models from the usual providers that focus specifically on healthcare, and make diagnosis and treatment much more accessible to millions or even billions of humans. Google is probably the leader in this regard with its MedPaLM, now Med-Gemini models. They are already better than the average physician at diagnosis and treatment recommendations, but even after four years of validation they aren’t quite ready to introduce the technology to the world. I hope it comes soon.

• There are many different aspects of administrative AI that are being effectively used in healthcare. It’s not exciting, but it may be beginning to ease the heavy administrative burden in US healthcare—which currently takes up about 30 cents of every healthcare dollar. AI is being used to categorize ICD codes from clinical notes, prepare claims for payers, decide whether to pay claims by payers, schedule operating rooms, etc., etc.

• It's kind of a small thing, but our conversations with physicians and nurses are increasingly likely to be captured, turned into text, and summarized by AI. Now within five minutes of leaving my primary care physician I have a visit report in my email inbox. AI also means that my doctor can actually look at me while we talk, rather than typing things into an electronic health record system.

Are these healthcare uses of AI enough to justify a big celebration on AI Appreciation Day? I’m not sure we should break out the champagne yet, but healthcare AI does already offer some value and I’m confident it will provide much more in the future.

Subscribe now

Source: ChatGPT

Everybody in the AI world is excited about agents, which of course can perform digital tasks of various types rather than just inform curious humans. However, astute observers are a bit less excited about agents now than they were a few weeks ago. Cracks in the agent armor have begun to appear:

• Gartner predicts (always with rounded percentages!) that 40% of agentic projects will be canceled by 2027. The research analysts at the company blame escalating costs, unclear business value or inadequate risk controls. Surprising that a technology in its infancy is already entering Gartner’s “trough of disillusionment.”

• Anthropic used its LLM Claude to run a small automated shop in its headquarters building. As an agent it could decide on prices, what to stock, when to restock, and how to communicate with customers. The experiment did not go well; to quote an Anthropic blog post, “If Anthropic were deciding today to expand into the in-office vending market,² we would not hire Claudius [the name it gave its LLM Claude for this job]…it made too many mistakes to run the shop successfully.”

• Carnegie Mellon researchers tried out ten different agents to run a software business. The agents made lots of mistakes; the most effective agent (from Anthropic, ironically) only completed 24% of its assigned tasks successfully.

The vendors all say that these agents will get better over time, and they no doubt will. And of course human workers also make mistakes. But these predictions and results suggest a specific set of roles for humans in a world of lots of AI agents. Are they roles that we want to play? I can’t speak for all humans, but I’m not sure I would want these jobs. Here are some I envision:

Human agents—The inability of agents to perform all roles suggests that there will be “human agent” roles to do certain steps in a process that an AI agent can’t (yet) perform well. Maybe they will involve interfacing with other humans, or performing physical tasks that robots aren’t yet good enough to undertake. This description of a new Amazon warehouse, for example, suggests that humans perform tasks like “unloading trucks and packing orders of unusual sizes or shapes.” To emphasize that these human workers are a form of agent, the article points out that, “Many of the roles at the Shreveport facility involve managing and collaborating with robotic systems.”

Agent auditors—If AI agents make a lot of mistakes, they will need humans to review and audit their performance. They’ll also probably have to reverse or fix the mistakes that agents have made (I think agent task reversal is a critical capability!). This auditing doesn’t strike me as a particularly fun job; it would seem not to involve a high degree of creativity or “agency.” You’re responding to a machine all day long—and if it’s only doing a quarter of its assigned tasks well, you’re going to be very busy.

Incremental agent improvers—A well-designed work environment involving AI agents would probably require that humans not only monitor and correct their mistakes, but also make suggestions to the agents (or their human creators) for how to improve their performance. In order to do this well, the human worker would need to be technically astute and know something about how the agents work. This could be a more fulfilling role than the agent auditor role, but my experience in talking with programmers is that they’d rather write new code than fix someone else’s bugs. Perhaps this job and the auditor one will eventually be replaced by “quality control agents” that will monitor performance and recommend fixes for errors.

Agent architects—One of the better roles in an agent-centered world is designing what they do in an overall system or process. I don’t think it will be easy, given that a broad process may involve lots of different agents, each doing a different task. And given the current agent accuracy level, it will involve both human and AI actors. At some point we’ll probably have trusted third parties to rate and authenticate agent performance, which should make it easier for architects. This service would be something like what Apple and Google do for smartphone apps. There’s nothing like that at the moment, however.

Agentic decision-makers—The most important role in an agent-centered world is deciding whether to use agents or not for a particular purpose. Again, this could be a tough job. Agents are likely to put some humans out of work, and being a hatchet person is never fun. They are also likely to substitute for existing software applications; Satya Nadella, CEO of Microsoft, is one of the software company leaders who have argued that AI agents will replace conventional SaaS software. Deciding when and how to do that within a particular organization isn’t going to be easy.

Agentic Psychologists—A somewhat tongue-in-cheek role I might suggest is that of the agentic psychologist, who would counsel unhappy humans about their daily work with AI agents. AI can usually work faster than we humans can, and doesn’t need to take breaks. The digital tasks that humans still perform are also likely to be measured. The speed and monitoring of the AI/human collaboration will probably put stress on the humans assigned to partner with them. One social media-based survey of Amazon warehouse workers suggests that they “are suffering physical injuries and mental stress on the job as a result of the company’s extreme focus on speed and pervasive surveillance.” Physical injuries aren’t very likely from collaborating with AI agents, but stress does seem a probable side effect. Perhaps the humans who do work with AI agents will have easy access to a counseling chatbot to relieve their stress.

I would stress that it’s early days for agentic AI, and we don’t know how things will eventually turn out in terms of the capabilities of agents and the ways humans will need to adapt. But perhaps we should think a bit about the implications for humans before we charge off too fast in an agent-focused direction.

Subscribe now