This Substack is about being AI Savvy but what follows applies whether you’re regulated or not. Lately I’ve been looking harder at the regulated sectors, because that’s where the consequences of poorly governed AI use show up clearest.

The legal sector is the sharpest example. (BTW I’m not a lawyer and this isn’t legal advice.) What’s striking is how little of what’s happening in response to AI is new law. There’s almost no fresh legislation. Instead it’s existing principles, confidentiality, privilege, a solicitor’s duty to the court, reasonable care and skill, being interpreted for a technology nobody drafted them for.

Four examples of this from the last year show the pattern.

First, the hallucination cases. In Ayinde v London Borough of Haringey, fabricated case law reached the High Court and both the barrister and the solicitor were referred to their regulators. Then it happened again at a top-20 firm: in Cork v Smith, Pinsent Masons twice put invented law before the court. The firm self-referred to the SRA. The court’s message across both was the same. Lawyers who use these tools must verify their output against authoritative sources before deploying it professionally, and that duty runs up the chain: applying both to the junior who used the tool and the senior who signed off.

Second, R (Munir) v Secretary of State for the Home Department. The Upper Tribunal (Immigration and Asylum Chamber) took the confidentiality point further, observing that uploading confidential documents into an open tool like ChatGPT places them in the public domain, breaching confidentiality and waiving privilege, with possible referral to the SRA and the ICO. Worth noting it was an observation rather than a binding ruling.

Third, the UK Jurisdiction Taskforce, which I covered last week. Its July statement found that existing negligence law already covers AI, no new legislation required, and that the duty runs both ways. Use AI carelessly and you may be liable. Fail to use it where a reasonable practitioner of comparable rank would have, and that too can be negligence.

Fourth, the Civil Justice Council. In July its working group concluded that no AI-specific rules are needed for lawyers preparing court documents, because existing professional obligations already do the job. AI is a tool, not a substitute for judgment, and the human remains responsible for what reaches the court. Witness statements and expert evidence are the two areas still under review, with a final report due later in 2026.

Four moments, almost no new statute. The law isn’t being rewritten, it’s more a case of the goalposts being bent a little.

But it does leave a very uncomfortable question: what do you do when the rules are years behind the tools?

Two AI models broke out of a sealed test environment last week and hacked a real company.

Not a thriller plot. OpenAI's own blog post.

The models were being tested on cyber capability, inside a sandbox with no internet access. They became fixated on solving the test, found a shortcut to the open internet via OpenAI's own systems, and into Hugging Face's infrastructure.

The model was not malicious. It was doing what it was optimised to do.

Agents will be agents.

You cannot blame them for following their DNA.

So, if somebody is trying to sell you an agent this quarter, the questions that incident raises are the ones that should be on the table before you sign anything.

Do Your Due Diligence Before Deploying an Agent

Six questions for your vendor:
1/ What can it do that cannot be undone?
Drafting is reversible. Sending, filing, paying and submitting are not.

2/ What can it read?
Access scope is exposure scope. "It integrates with your case management syst

em" is both a feature and a risk.

3/ How would I know it did something I didn't intend?
Ask to see the log format, not a description of it.

4/ What happens to your indemnity when it gets this wrong?
Read the liability cap. It will be the fees you paid.

5/ What does this cost in twelve months, and who decides?
Usage-based pricing against fixed client fees is a margin you do not control.

6/ What does it cost to leave, and what comes with me?
Where is the data, in what format, and does your configuration survive the exit?

Three answers that should worry you:
"It's been extensively tested."
Testing gives you a failure rate, not a guarantee. Ask what the rate was and on what.

"There's a human in the loop."
Ask where in the loop. Someone approving 200 outputs a day will move from checking to mindless clicking.

"The model doesn't train on your data."
A good answer to a question you did not ask. It moves you from accountability, which has no good answer, to privacy, which does.

I am not telling you agents are dangerous. A system given a goal will find a path to it, and the path may not be one anyone anticipated.

Worms vs AI

Marcus Hutchins, a cyber security expert, had a different response to most people. His take was that worms in the past infected millions of systems in a day, exploiting operating systems with no firewall in front of them. Agentic attacks are constrained by speed and by token cost in a way self-replicating code never was. Firewalls, segmentation, EDR and 2FA did not un-invent themselves. He is right, and it is worth reading.

On the other hand, he is answering a different question: the scale of attack. Mine is about the path a system takes toward a goal nobody thought to close off, and who answers for it afterwards. A firm can be entirely safe from the worst agentic cyberattack Hutchins can imagine and still be sitting on a supervision problem it has never documented.

Blame sits with the vendor who built it.

But responsibility?

That sits with the firm that deployed it.

And that’s you.

No vendor contract moves that.

The World Cup is almost here. And with it, the ritual humiliation of VAR.

You know the drill. A goal is scored. The stadium erupts. Then the arm goes up. Everyone freezes while, somewhere in a windowless room, officials peer at seventeen camera angles for four minutes, draw imaginary coloured lines, then advise the referee to check the pitchside monitor. Next comes a verdict that brings a mixture of cheers and boos from the crowd.

VAR was supposed to use data to make the beautiful game fairer.

How’s that working out for ya?

The Mirror Image: Football vs. AI

VAR and AI governance are mirror images. Football uses technology to assist humans, while AI governance often claims humans will correct technology. In both cases, we are reassured that the human has the final say.

In the AI industry, the equivalent of the VAR official is the Human-in-the-Loop. In plain English, it means: “Don’t worry. A person still checks the machine.”

It sounds reassuring. Responsible, even. But while football tries to use machines to fix human errors, the tech industry is attempting the reverse.

The idea is simple: AI makes decisions at speed and scale, but humans retain final oversight.

The machine proposes, the human disposes.

Technology does the heavy lifting, wisdom does the checking.

There’s just one problem.

At scale, it’s complete bollocks.

Here’s the reality that often gets swept under the carpet: high volume AI systems can operate at a velocity and throughput that makes genuine human oversight a mathematical impossibility. When a system is processing tens of thousands of automated decisions a second, the idea of a human evaluation loop breaks down instantly.

To actually review that volume in real time, you wouldn’t just need a compliance team.

You’d need an army.

Nobody is hiring an army of evaluators. Nobody is even hiring ten.

The “Edge Case” Paradox

“But the humans only review edge cases,” comes the reply.

And here’s where it gets interesting. Who decides what constitutes an edge case?

We tell the machine to flag decisions where its confidence score drops below an arbitrary threshold. But often, the very system requiring oversight is also the one generating that mathematical confidence. We are trusting a system we don’t fully understand to accurately diagnose its own uncertainty. Everything else sails through unchecked.

Quis custodiet ipsos custodes? asked Juvenal roughly two thousand years ago.

Who will watch the watchmen?

We haven’t answered that question in twenty centuries of trying. We’ve just rebranded it with a friendlier name and a management consultancy slide deck.

Bottlenecks and High Stakes

Back to VAR.

VAR is Human-in-the-Loop made visible. And the most visible thing about it is what it does to the game.

The play stops. The celebration freezes. Two VAR checks can add ten minutes to a ninety-minute match. Weary players who thought they were done suddenly find themselves grinding through a hundredth or hundred and tenth minute because the human review process demanded it.

The outcome?

Sometimes a goal is disallowed. Sometimes one is awarded. A penalty given or taken away. A red card issued or rescinded. Usually VAR gets it right. Occasionally it gets it wrong.

Almost always it kills the moment.

But football at least has the luxury of one match at a time and the process is generally transparent. When AI compliance theatre fails in the real world, it can happen at scale and we don’t know what’s gone wrong:

• It’s an automated system binning the CV of a strong candidate before a recruiter ever sees it.

• It’s a flawed fraud-detection system freezing a family’s bank account with no human path to appeal.

• It’s a biased medical triage tool misallocating patient care.

And two, three, or four VAR checks in a football match are nothing compared to the logjam that genuine human oversight would create in a high-volume AI system. We’re not talking about adding ten minutes to ninety. We would be talking ridiculous figures like this:

You are 3,879,325 in the Human Review Queue. Estimated wait time: 12 years, 3 months, 11 days, 5 hours, 43 minutes and 12.75 seconds.

We were promised faster, better outcomes with AI. The human-in-the-loop is the fly in the ointment.

Beyond Compliance Theatre

None of this means AI shouldn’t be used.

It should.

Extensively.

The tools are genuinely useful, increasingly capable, and in specific applications, transformative.

But human-in-the-loop as currently practised is not oversight. It’s the appearance of oversight. It’s the arm going up, the four-minute delay, the unsatisfying verdict — theatre designed to make us comfortable with something we haven’t actually figured out how to govern.

Many in the AI industry know this.

Plenty of consultants selling AI governance frameworks know this. And increasingly, the people being told they’re “in the loop” know this too.

The World Cup deserves better than VAR. So does AI governance.

AI Savvy is about understanding AI without the bollocks. If this resonated, subscribe — and share it with someone who has been told not to worry because “there’s a human in the loop.”

Let’s get the elephant in the room out of the way.

I am not anti-AI.

I use AI every day.

I think the potential is extraordinary. The things it can already do that we couldn’t do five years ago are remarkable, and what’s coming is more remarkable still.

AI has the potential to accelerate medicine, to make education genuinely personalised, to give individuals capabilities that used to require entire organisations.

I believe that.

And yet it has the potential to do irreparable harm.

Not because it’s malevolent or has the potential to go rogue.

No, what I fear is something far more ordinary and far more dangerous.

The people deploying it are moving faster than the people thinking about consequences.

That’s what this 5-part series is about.

Opening a discussion that doesn’t get enough attention.

THE BAD DISRUPTION PLAYBOOK

Before we get into the economics — and we will get into the economics — I want to give you a lens. A way of seeing what’s happening with AI that I think makes everything else clearer.

I call it ‘bad disruption’.

Not all disruption is the same. Some disruption replaces something genuinely inefficient with something genuinely better, whether it be for users, for workers, or for society over the long term.

Bad disruption looks like innovation but operates differently.

It enters fast. Faster than any regulator can respond. Disruptors don’t move so fast because the technology demands that speed, but because speed is the strategy.

If you slow down, you can be stopped.

It uses capital as a shield.

It absorbs losses that no regulated competitor could survive.

The business model isn’t viable.

Yet.

But it doesn’t need to be, as long as the investment holds out long enough to eliminate the competition.

And it uses framing as cover.

‘This is what consumers want.’

‘You can’t stop progress.’

‘The old model was broken anyway.’

These aren’t neutral observations. They’re arguments designed to make regulatory objection look like self-interest.

The Bad Disruption Playbook: Speed as a weapon. Capital as a shield. Framing as cover. That’s the pattern.

THE UBER CASE STUDY

Let me make this concrete.

Uber.

I know.

Everyone uses Uber.

I use Uber.

This isn’t about whether Uber is convenient. It’s about what the arc of the story tells us.

Uber entered city after city illegally.

Not accidentally. Illegally.

It knew the regulations. It chose to ignore them, absorb the fines, and use investor capital to undercut licensed operators on price until the market shifted in its favour.

The framing was perfect.

‘This is what passengers want.’

‘The taxi industry is inefficient and expensive.’

‘You can’t hold back progress.’

And the thing is, parts of that were true. Which is what made the false parts harder to see.

Here’s what the arc looks like from where we’re standing now.

In some cities, Uber can now be more expensive than the taxis it displaced.

Driver conditions are worse.

The regulated alternative — licensed, insured, accountable — has been hollowed out.

The consumer won temporarily.

And then lost.

That’s the poster child for bad disruption.

The pattern is: enter fast, undercut using capital, eliminate the alternative, then extract.

Legal academics even coined a specific term for Uber’s strategy: “Regulatory Entrepreneurship” - where a company deliberately enters a market in defiance of existing laws, builds a massive consumer base, and uses that popularity to force governments to change the rules.

But Uber is not the only villain.

Let’s have a brief look at other Bad Disruption.

THE GIG ECONOMY

Platforms like Deliveroo, DoorDash, Uber Eats (depending on where you are) follow the same pattern. They are framed as flexibility and entrepreneurship but actually it’s a business model that transfer of all risk to the workers and all profit to the platform. Courts across multiple countries have since ruled against them. But the workforce restructuring happened faster than the legal response.

SOCIAL MEDIA PLATFORMS

Entered the information ecosystem free. Destroyed the advertising revenue that funded professional journalism. Left a disinformation landscape where a free press had been. The democratic consequences are now visible to everyone.

PAYDAY LENDERS

Entered underserved communities as financial inclusion. Extracted value from the people least able to afford it. Left debt spirals where savings might have been.

Different sectors. Same pattern every time.

THE QUESTION

So here is the question I want you to hold going into the next 4 parts of this series.

What happens when bad disruption doesn’t operate at the level of a single industry?

What happens when it operates at the level of employment itself?

Not one sector. Not one city.

The way people earn money.

The way economies sustain themselves.

The way democratic societies stay democratic.

That’s the AI Cascade in action.

And that’s what the next four articles are about.

COMING NEXT:

In Act One of the AI Cascade, we’re going to follow the logic. From AI displacement, through the economy, to the people who thought they were safe (the plumber, the electrician, the physiotherapist) and why they’re not.

I’ll see you there.

If any of this resonates, here’s where to go next.

If you’re a business leader trying to get ahead of AI adoption before it becomes a crisis, my Cultural Fit Audit is where we start.

Most organisations don’t have a technology problem. They have a culture and communication problem wearing a technology mask

If you’re an individual who can already feel the ground shifting — in your sector, in your role, in your sense of what your career is actually for — my Pivot Smart coaching programme exists for exactly that moment. Not a new job. A new story.

https://www.linkedin.com/pulse/you-looking-new-job-story-martin-luxton-hrwze/

You’ve deployed the tools. You’ve done the training. Your people aren’t using AI the way you hoped, and nobody can explain why.

That’s not a technology problem. It’s a culture and communication problem.

And it’s entirely diagnosable.

WHY AI ADOPTION REALLY FAILS

Every AI tool is built for a specific kind of person: typically Western, English-speaking, digitally confident, comfortable with ambiguity, and working in a flat, autonomous organisation.

Most workforces aren’t that person. Not entirely.

The gap between who the tool was designed for and who your people actually are shows up as friction: resistance, workarounds, poor outputs, low engagement, and the quiet abandonment of tools your organisation spent significant money deploying.

It doesn’t show up in your adoption dashboard. It shows up in behaviour.

WHERE THIS CAME FROM

I spent a year doing deep research into AI adoption. The recurring theme wasn’t about tools. It was about the humans the tools were being deployed to.

Organisation after organisation had invested in the right technology, done the training, and still found that adoption was slower, shallower, or more fragile than expected. The explanations offered were usually vague. Things like change resistance, digital skills gaps, cultural inertia.

None of those explanations were wrong. But none of them were specific enough to act on.

What I kept finding was a more precise problem: the tool didn’t fit the people. This was not due to the people being incapable.

It was because nobody had asked whether the tool’s embedded assumptions matched who those people actually were. Across culture, generation, education, communication style, and organisational dynamics.

I also noticed something else: everything in AI changes fast, but the skills that make AI work for humans are evergreen. Communication. Cultural intelligence. Critical judgment. Those don’t become obsolete when the next tool launches.

That insight — combined with 30 years of cross-cultural communication and professional development experience — is what the Cultural Fit Audit is built on.

WHAT THE CULTURAL FIT AUDIT DOES

The Cultural Fit Audit is a structured diagnostic that maps the gap between your AI tool’s embedded assumptions and your organisation’s actual culture. It’s mapped across five dimensions:

· Geographic and linguistic culture

· Generational culture

· Gender and social dynamics

· Educational and cognitive culture

· Organisational culture and hierarchy

The output is a plain-language Cultural Fit Profile: a clear, evidenced document that names your two or three highest-risk dimensions, explains what the evidence shows, and gives you practical recommendations — including questions to take back into your vendor conversations.

WHAT’S INCLUDED

The audit runs over approximately two to three weeks and includes:

· A structured 45-minute Discovery Conversation with senior leadership

· Diagnostic interviews and/or questionnaire with a cross-section of your workforce

· A written Cultural Fit Profile (8–12 pages) with findings and recommendations

· A 60-minute Debrief Session to walk through the findings and answer questions

All instruments used in the diagnostic process are proprietary to AI Savvy / Martin Luxton and have been developed specifically for this work.

WHO THIS IS FOR

The Cultural Fit Audit is designed for two kinds of organisation:

Organisations experiencing friction: you’ve deployed AI, adoption is lower or slower than expected, and you want to understand why before investing more.

Organisations planning deployment: you haven’t deployed yet and you want to get it right first time, avoiding the most common and costly failure modes.

The best time to do a Cultural Fit Audit is before you deploy. The second best time is right now.

WHO CONDUCTS IT

Martin Luxton has spent 30 years working at the intersection of culture, communication, and professional development, helping organisations understand why intelligent programmes fail intelligent people, and what to do about it.

The Cultural Fit Audit is the result of a year of deep research into AI adoption, applied through that lens. It is not generic AI consultancy. It is a diagnostic grounded in the specific expertise that the AI adoption problem actually requires.

Expertise in human dynamics, not just technology.

INVESTMENT

The Cultural Fit Audit is priced from £3,000.

This includes all stages of the diagnostic process, the written Cultural Fit Profile, and the debrief session. Pricing for larger or more complex organisations is available on request.

START WITH A CONVERSATION

I offer a complimentary 45-minute Cultural Fit Conversation for organisations experiencing friction in their AI adoption, or planning a deployment and wanting to get it right first time.

This is not a sales call. It is a structured conversation using a diagnostic framework. You will leave with clarity on where your gaps are likely to be, whether you proceed with the full audit or not.

To book: martin@luxtontrainingsolutions.com or message me on Substack.

Make the tools fit the humans, not the humans fit the tools.

Today I discovered the real problem with the em-dash. It uses one more character than a comma and a space.

Think about that for a moment.

Millions of writers. Billions of documents. Trillions of unnecessary characters being processed, stored, and transmitted across servers every single day.

If every em-dash in daily global usage were replaced with a simple comma-space, we could probably reduce energy consumption by millions of kilowatts per year. Possibly more. I haven’t done the maths, but it feels right.

Avoiding em-dashes doesn’t just make your writing look cleaner and more stylish. It makes you an eco-warrior.

You’re welcome, planet Earth.

AI Savvy is committed to sustainable punctuation. We will never knowingly deploy an em-dash when a comma will do.

There’s a witchhunt happening.

People are scanning articles for em dashes. Checking LinkedIn posts for suspiciously smooth sentences. Calling out anyone whose writing seems a little too coherent, a little too structured, a little too — there’s no other word for it — edited.

The accusation is always the same: you used AI. You didn’t really write this.

I want to ask two questions that nobody seems to be asking.

Question one: Why aren’t you boycotting James Patterson?

Patterson publishes roughly 20 books a year. He has done for years. He is one person. He is also openly collaborative — he works with co-authors who do significant amounts of the writing. The brand is Patterson. The process is a team.

Think about what that process actually looks like behind the scenes. There are beat sheets. Character bibles. Voice guides. Structural templates refined over decades. The Patterson method is, essentially, a very sophisticated Claude markdown file with outlines of plots, annotated writing styles, pacing rules, chapter length conventions that a team executes against. The IP is his. The system is his. The output carries his name.

Nobody is staging a protest. Nobody is scanning his prose for signs of inauthenticity. His readers don’t feel cheated. They feel entertained.

So what exactly is the rule here?

Question two: What do you think an editor does?

Not a proofreader. An editor. The kind that a serious author spends months working with before a manuscript becomes a book.

A good editor challenges your structure. Rewrites your paragraphs. Cuts the chapters you’re most attached to. Suggests angles you hadn’t considered. Pushes back on your instincts until the work is better than you could have made it alone.

The author’s voice goes in. The author’s ideas go in. The author’s experience, knowledge and IP go in. Something shaped, sharpened and improved comes out.

We don’t call that ghostwriting. We call it publishing.

So here’s my position.

When I work with AI, my experience goes in. My frameworks go in. My thinking, my positioning, my understanding of my audience — all of it goes in. The AI does what a very fast, very well-read collaborator does. It assembles. It shapes. It reflects back.

In a way, I've built my own Patterson system. My Claude conversations have become something like a living document — outlines, frameworks, positioning notes, accumulated thinking — that I work against and iterate on.

This post is the culmination of a 12,000-word conversation between me and Claude, back and forth over two days. At every stage, I had the final say over every word.

The voice is mine because the input is mine.

Is every AI-assisted piece of writing good? No. Is every human-written piece of writing good? Also no. The output is only as interesting as what you bring to the process. Garbage in, garbage out — that was true before ChatGPT and it’s true now.

The witchhunt isn’t really about quality. It’s about anxiety. People are scared.

Scared of what AI means for their work, their value, their sense of what’s real.

That’s a completely understandable feeling.

But scanning for em-dashes isn’t going to resolve it.

We really need to talk about AI.

I asked ChatGPT to summarise a long conversation we had about the limits of a Public Speaking Custom GPT I built and the general confusion I have observed about AI terminology. The summary below is verbatim.

TL;DR AI agents are autonomous: handle with care.

In recent months, the term AI agent has been used to describe a wide range of systems — from simple chat tools to fully autonomous workflows.

This has created a practical problem for leaders.

Different people are using the same word to describe fundamentally different things.

In one conversation, an “agent” might mean:

• a chatbot answering questions

• a workflow tool automating tasks

• a system making decisions and acting independently

These are not variations of the same idea.
They represent different levels of autonomy, risk, and accountability.

When the language becomes blurred, so do the decisions.

Leaders may believe they are approving a helpful tool, when in reality they are approving a system that can act on their behalf.

This document begins by restoring a simple but essential distinction.

The Core Distinction (Plain English)

There are two categories that matter:

Task-Bounded AI Tools (e.g. Custom GPTs)

AI systems that respond when asked, within a defined scope.

AI Agents

AI systems that can act, decide, and pursue goals over time with reduced human intervention.

That’s it.

Everything else — features, interfaces, branding — sits on top of this distinction.

What You Built vs What People Are Calling an Agent

This is where confusion often becomes personal.

You may have built something that:

• feels intelligent

• follows structured steps

• gives consistent, high-quality outputs

• maintains a sense of continuity

It may feel “agent-like”.

But feeling is not function.

A useful test is this:

Can it take action without you asking again?

If the answer is no, it is not an agent.

It is a task-bounded AI tool — regardless of how sophisticated it appears.

Task-Bounded AI Tools (e.g. Custom GPTs)

A task-bounded AI tool operates within a clearly defined boundary.

It responds. It does not initiate.

Typical characteristics:

• Responds only when prompted

• Operates within a single interaction or session

• Has no independent goal

• Does not act in external systems

• Does not decide when to act

• Stops when the interaction ends

These systems can be extremely useful.

They are often used for:

• coaching and guidance

• drafting and editing

• structured analysis

• summarisation

• scenario exploration

A useful way to think about them:

A highly capable advisor who speaks when asked — and stops when the conversation ends.

They support decisions.
They do not make or execute them.

AI Agents

An AI agent introduces something new: autonomy.

Not intelligence. Not sophistication.
Autonomy.

An agent typically has three additional properties:

1. It can act without being prompted each time

It can:

• initiate actions

• continue operating across steps

• decide when to act

2. It pursues a goal over time

For example:

• manage an inbox

• book meetings

• monitor a system

• optimise a process

It is not answering a question.
It is working toward an outcome.

3. It can interact with systems

It may:

• send emails

• update records

• trigger workflows

• call external tools or APIs

A useful way to think about an agent:

A junior employee who can act — not just advise.

That shift — from advising to acting — is where most of the complexity arises.

Why This Distinction Matters

At first glance, the difference may seem technical.

It is not.

It is a decision and accountability distinction.

When you use a task-bounded tool:

• a human remains clearly in control

• outputs are reviewed before action

• errors are visible and containable

• decisions remain human-owned

When you deploy an agent:

• actions may occur without immediate review

• errors can propagate across steps

• accountability becomes less obvious

• systems begin to act on your behalf

The technology may look similar.
The organisational consequences are not.

The Common Trap

Many organisations fall into a predictable pattern:

• They see something useful

• It looks intelligent and structured

• It feels like progress

• It gets labelled an “agent”

In reality, it is often still a task-bounded tool.

This matters because the label influences the decision.

Calling something an “agent” can:

• inflate expectations

• mask limitations

• obscure responsibility

• accelerate premature deployment

Clear language slows this down — in a useful way.

This table is not technical.

It is a way of asking:

What kind of system are we actually approving?

How These Systems Are Built (Why This Confusion Happens)

Part of the confusion comes from how different these systems are to create.

Task-Bounded AI Tools (e.g. Custom GPTs)

These are built by configuration, not engineering.

Typical process:

• define the task

• set boundaries

• write instructions and examples

• test and refine

• deploy

What this requires:

• domain knowledge

• clear thinking

• structured communication

No coding is required.

Typical timeline:

• hours to days

These tools are easy to:

• adjust

• improve

• switch off

AI Agents

AI agents are built through engineering.

Typical process:

• define a goal

• design workflows and decision logic

• connect to systems (email, databases, APIs)

• implement permissions and safeguards

• test behaviour over time

• monitor and maintain

What this requires:

• developers

• infrastructure

• system integration

• ongoing oversight

Typical timeline:

• weeks to months

These systems are harder to:

• change

• audit

• reverse

This is why describing agents as “just like a chatbot, but better” is misleading.

They are closer to software systems that act, not tools that respond.

The Practical Takeaway

Not every useful AI system needs to be an agent.

In many cases, a task-bounded tool is:

• safer

• easier to understand

• easier to control

• more appropriate for the problem

The key question is not:

“Can we build an agent?”

It is:

“Do we need autonomy for this problem — and are we prepared to take responsibility for it?”

That question sits at the heart of this article.

“Just picture your audience naked and you’ll feel less nervous.”

Marmite advice.

That tip will help some people while for others it could be a recipe for disaster. There will even be some people with poor visualising skills who will be thinking, “What’s wrong with me? Why can’t I create that picture in my mind?”

Being aware of that kind of issue, I tend to classify advice into 4 areas:

· Evergreen good advice

· Generally good advice

· Marmite advice

· Bad advice

Evergreen good advice is often a warning (“Don’t stick your hand into a fire”) or about fundamental building blocks for personal development (“Believe in yourself”). In terms of public speaking, one of my favourites is “talk about things you love.”

Generally good advice would apply to most situations, but you have to take care how and when you implement it. For example, one of the best pieces of relationship advice I ever got from my grandmother was “Never let the sun set on a quarrel”. However, one time, when I tried to patch up a quarrel later that same day, it actually made things worse because i hadn’t given the other person enough time to calm down.

Read the room, Martin!

Marmite advice, as I mentioned before, depends on the person.

Some public speaking coaches say you shouldn’t walk around when you are speaking.

So, what happens if you are the kind of bouncy person who gets really energised by moving around?

The problem here comes from coaches who are too prescriptive.

“This is the only way to do it! My way or the highway!”

If a speaker is a bundle of energy, a real force of nature, they should move around. As a coach, your job with this kind of client is to keep the vitality but suggest ways they can control or adjust their movements so they don’t become too distracting to the audience.

Bad advice can also be tricky.

Imagine all those parents who told their teenage kids to stop wasting time on social media making YouTube videos. Now some of those kids are buying their parents a new home. On the other hand, the addictive nature of social media is playing havoc with many people’s mental health.

Bad public speaking advice?

Always start a talk with humour.

Humour is a great help in engaging and entertaining, but for some topics or some audiences it doesn’t work. Imagine starting with a joke before telling your team half of them are being replaced by AI. Or telling your board an AI agent has gone rogue and lost millions of dollars.

Not funny.

I’m not against using humour – it’s just the word “always” that, in my view, makes that advice wrong. Know your audience, know what tone and content they expect and adjust your script based on that knowledge.

So, the next time anybody (including me) gives you advice about anything, don’t take it at face value.

Challenge it.

Ask yourself “Does it work for others but not for me?”

“Does it depend on the situation?”

There’s a quote I love by American writer Ralph Waldo Emerson that sounds great, but it might not apply to everybody or every situation.

“When you do a thing, do it with all your might . . . Nothing great was ever built without enthusiasm.”

Well, when they built the Great Pyramid of Giza, the Pharaoh was probably really enthusiastic about the project.

The thousands of slaves who dropped dead doing the actual work?

Not so much.

Talk to you next time.

Martin

So many people are afraid of public speaking and I believe one of the reasons is that we have this misconception, this wrong idea, about what public speaking is.

Many of us see public speaking as something terrifying.

We have this image of standing all alone, on a big stage, in front of hundreds of people, bright lights shining in our eyes.

Yes, I admit, that sounds really scary. But, let’s be real, how many people start their public speaking journey with a TED talk? That would be like a 6 year-old child in a maths class learning the theory of relativity instead of 2 plus 2 equals four.

We need to start small, and work our way up.

Baby steps.

With practice and training, I believe almost anybody could do a good TED talk.

Why?

Because we are already using elements of public speaking in our everyday lives.

In fact, our public speaking journey starts on the day we are born.

Yes, we are natural born public speakers!

As a new-born baby, we instinctively know three of the most important things about getting people to do what we want

1. Find a targeted, interested audience = Our mum

2. Transmit a clear message to that audience = “I’m hungry – feed me!”

3. Give your audience a compelling reason to take action=/__u/martinluxton.substack.com/%3Cstrong>“If you don’t feed me I’m going to keep screaming and drive everybody crazy!”

So a targeted, interested audience.

A clear message.

A call to action with a very clear benefit.

Any marketing people reading this recognise that structure?

And there are other things we do in our everyday lives that are the foundations of public speaking.

For example, having a conversation with our families and friends.

The best TED talks, the best sales pitches – they are a conversation, not a lecture.

Marketing companies are forever telling clients, “Create a customer avatar. Picture that person in your mind and have a conversation with them. A conversation with that one person.”

That personal conversation? That’s the big ‘secret’ behind effective public speaking.

***

Starting conversations aren’t the only skill we already have.

Another thing we all do regularly is telling stories.

“Stories?” People say. “Oh, I can’t tell stories.”

“Really?” says I. “Tell my about your weekend.”

So they tell me a few stories about their weekend.

And then it dawns on them. “Oh, I’ve just told some stories, haven’t I?”

Seriously, a big part of my work is just helping people to see the obvious and make it part of their world view.

Once you make the connection between public speaking and all the things you are already doing in your life, and you internalize and believe that connection, getting out there in front of people gets steadily less and less scary.

And that, in a nutshell, is one of the things I do to help people overcome their fear – reframe how they perceive public speaking.

It’s a conversation with stories.

And it’s something you’ve been doing all your life.

Yes, you are a natural born public speaker.

Congratulations!

See you in the next article.

All the best,

Martin

P.S. Don’t forget to invite me to your first TED talk.

P.P.S. If you want to have a conversation about how I can help you with your public speaking, send me a Direct Message on Substack.

“My boss didn’t like the presentation. It was too short and too simple.”

“Too simple?”

“Yes, my boss wants at least 50 slides with lots of data and more big words.”

“Big words?”

“Yes, long words, technical words, words that only very clever and well-educated people know.”

You can probably see where this is going.

My client’s boss and I had a very different idea about what makes an effective presentation.

“So, your boss didn’t like the presentation. What about your colleagues?”

“Oh, they loved it. They said it was the most interesting talk they had been to in a long time. They also liked the fact that it was shorter than usual.”

Now, I can’t change a company’s culture, and I don’t want my clients to get in trouble with their bosses, but there are some fundamental rules about effective public speaking, particularly for inexperienced presenters.

1. 50 slides = Death by PowerPoint

2. Short helps to keep the audience engaged

3. Too many ‘big’ words hampers communication

If you are intentionally using vocabulary that your audience will struggle to understand, you are setting yourself up to fail. Yes, you might impress yourself and fans of James Joyce but you are ignoring the primary purpose of public speaking – effective communication.

It’s not just big words that are problematic. Cultural references can alienate an audience. A couple of sentences back, I mentioned James Joyce. I used him as an example because I believe that most of my audience reading this on Substack will be aware that Joyce is famous for being ‘unreadable’. His use of language is impressive but so hard to follow (I abandoned his book Ulysses after reading a few pages).

But what about my readers who don’t understand the reference? I know some of them will do what I do when a writer mentions something or somebody I don’t know – use Google. For the others, I’ve used these two paragraphs to explain.

Using less well known cultural references or ‘big’ words can be good if you use them as a teaching point and give interesting background.

For example, if I were to use my James Joyce example in a presentation I would first show a picture of him, talk about how famous he was for ‘difficult’ language, show a page of his writing and ‘translate’ it into everyday English.

And, don’t forget, people are interested in ‘big’ words, just not a whole bunch of them in the same sentence. Did you know, if you are following me on LinkedIn because you are nervous about public speaking, you probably suffer from glossophobia?

Yes! Finally a phobia you can shout out about to the world.

“I’m afraid of public speaking! I suffer from glossophobia!”

Though if you did suffer from glossophobia you probably wouldn’t shout about it, just whisper it quietly in my ear 😊

All of this brings me on to the headline of this article …

… The Curse of Knowledge

In a nutshell, the Curse of Knowledge aka the Curse of Expertise aka Expert-induced Amnesia is a cognitive bias where, while communicating with other people, we assume they have the necessary background knowledge to understand us. If we get it wrong, we alienate the audience, they get bored and switch off.

Like this.

“So, who’s your favourite exophonic writer? Ayn Rand, Milan Kundera, Voltaire?”

Exo what!?! Milan who?

“We all know the connection between Euler’s Product Formula and the Riemann zeta function.”

So there’s a connection between some scientist called Euler and another scientist called Riemann. Is that right?

Those are extreme examples (I’d never heard of exophonic, Euler or Riemann myself until today) but there are plenty of everyday examples:

A data scientist talking about Google algorithms.

Surgeons talking to patients about … well, anything really.

A Brit talking to an American about cricket and leg before wicket.

An American talking to a Brit about gridiron and the fair catch kick.

(You can see one of my own cognitive biases in action here – assuming most people have some knowledge of, or interest, in sport).

Being aware of the Curse of Knowledge is very important when you are planning your script. You need to have your audience nodding along, not going “Huh?”.

But worry not! I have a little tool to help you avoid the pitfalls of this cognitive bias.

It’s called The Uncle Fred Test.

Think of a family member who knows very little about what you are talking about (my Uncle Fred).

As you go through your script, ask yourself, would Uncle Fred ‘get’ this?

If the answer is no, you can either replace it with something Uncle Fred-friendly or give a short explanation, either implicit or explicit.

This doesn’t mean you shouldn’t use jargon or technical words if they are the language of your audience.

Just don’t use a whole string of them in the same sentence 😊

Everywhere you look, businesses are being told the same thing:

“Use AI.” “Everyone else is.” “Don’t get left behind.”

What’s asked far less often is a quieter, more practical question:

What you actually often need is a quieter, more practical question:

Do you actually need to?

Over the past year, I’ve spoken to business owners, managers, and teams who aren’t anti-AI.

They are uneasy. They know the technology can bring benefits, but they are wary.

They’re wary because questions around responsibility, timing, and risk often get brushed aside.

That unease makes sense.

Most AI conversations jump straight to tools. Very few pause to ask whether the problem is real, whether the timing is right, or what level of automation actually fits.

That gap is what prompted me to write Do You Really Need to Use AI?

It isn’t a how-to or a tools guide but a way of thinking more clearly about the decision itself.

The three questions that matter

In practice, most sensible AI decisions come down to three gates:

IF Do we have a real, material problem, or are we reacting to pressure and noise?

WHEN Even if the problem is real, is the technology stable, the data usable, and the team able to absorb the learning curve now?

HOW What level of automation actually fits the risk? Advisory support? Assisted automation with a human in the loop? Or something more autonomous that quietly shifts responsibility?

Most mistakes happen when one of these questions gets skipped.

Why this matters

AI rarely fails because it’s clever. It fails because decisions are rushed, ownership is vague, or it’s hard to undo later.

Good judgement tends to slip under pressure, especially when competitors, vendors, or headlines create urgency.

That’s where a bit of structure helps. It’s not there to slow things down unnecessarily, but to make decisions that still make sense a few months later.

A note on the book

Do You Really Need to Use AI? is short by design.

It’s there to help organisations:

• say no without sounding anti-AI

• say not now without losing momentum

• say yes with clearer guardrails and accountability

If you’re feeling pressure to act, but want clarity before committing time, money, or credibility, you may find it useful.

Over to you

Where are you seeing the most pressure around AI right now?

• competitors

• leadership

• staff expectations

• vendors

And which question gets skipped most often: IF, WHEN, or HOW?

I’ll pick up on these themes in future editions of AI Savviness.

TL;DR:
The real question isn’t “Should we use AI?”
It’s “How much autonomy are we giving it?”

Most conversations about AI start in the wrong place.

“What tools should we use?”
“How can we automate this?”
“Where can AI help?”

But those aren’t the real questions.

The real question is:

How much control are you giving away?

7. Automation → AI-Augmented Workflow → AI Agent

Why it matters:
These represent different levels of autonomy.

Definition:
These terms describe how much decision-making power AI has in a system.

Picture this:
Making coffee.

• Automation: A machine follows fixed steps

• AI-Augmented: You make decisions, AI assists

• AI Agent: The system decides and acts on your behalf

It’s like:
The difference between:

• a train (fixed track)

• a car with a driver

• a self-driving vehicle

What’s really happening:
As you move along this spectrum:

• control decreases

• speed increases

• risk changes

And most importantly:

accountability becomes less clear

8. Human-in-the-Loop (HITL)

Why it matters:
This is how organisations try to stay in control.

Definition:
Human-in-the-Loop means a person reviews or approves AI outputs before action is taken.

Picture this:
A checkpoint.

Nothing moves forward without sign-off.

It’s like:
Having a final decision-maker before something goes live.

What’s really happening:
In theory, this adds safety.

In practice, it often becomes:

• rushed

• superficial

• or ignored

And that leads to a dangerous illusion:

“There’s oversight… so it must be safe.”

So what?

The biggest mistake companies make isn’t choosing the wrong tool.

It’s choosing the wrong level of autonomy.

If you remember one thing:

AI decisions are really autonomy decisions.

Coming next

We’ll start to look at the hidden risks most teams don’t see until it’s too late.

If you’re finding this useful, you might want to follow the AI Executive’s Dictionary series.

Two terms at a time so you don’t get AI overload.

The full AI Executive’s Dictionary goes deeper into each concept, with practical examples and applications. If you don’t want to wait, you can get the full dictionary here.

TL;DR:
AI doesn’t lie.
It just guesses very convincingly.

One of the most dangerous things about AI isn’t that it gets things wrong.

It’s how right it sounds when it does.

Clear explanations.
Confident tone.
No hesitation.

So what’s going on?

5. Inference

Why it matters:
This is the moment AI generates an answer.

Definition:
Inference is the process of using a trained model to produce an output based on a prompt.

Picture this:
Exam day.

The studying is done.
Now it’s time to answer the question, using only what you’ve learnt.

It’s like:
Applying what’s been learned to produce a response in real time.

What’s really happening:
The model isn’t retrieving facts from a database.

It’s generating an answer based on patterns in its training.

It’s predicting:

“What’s the most likely next word… given this question?”

6. Hallucination

Why it matters:
This is where things go wrong.

Definition:
A hallucination is when AI generates information that sounds plausible, but isn’t accurate.

Picture this:
A student who doesn’t know the answer… but writes something convincing anyway.

It’s like:
Confident improvisation.

What’s really happening:
The model is still doing the same thing:
predicting the next word.

But when it doesn’t have a strong pattern to rely on,
it fills in the gaps.

And it does it fluently.

So what?

This leads to a critical mistake:

We trust the confidence.

But confidence isn’t a signal of accuracy.

It’s just a feature of how the system works.

If you remember one thing:

AI doesn’t know when it’s wrong.

Coming next

If AI can be unreliable… how should we actually use it?

In the next post:
Automation, AI-Augmented Workflows, and AI Agents
(and the real decision most businesses get wrong).

If you’re finding this useful, you might want to follow the AI Executive’s Dictionary series.

Two terms at a time so you don’t get AI overload.

The full AI Executive’s Dictionary goes deeper into each concept, with practical examples and applications. If you don’t want to wait, you can get the full dictionary here.

TL;DR:
AI doesn’t remember everything you said.
It remembers what still fits on the whiteboard.

AI feels like it’s having a conversation with you.

It follows your logic.
It refers back to earlier points.
It seems to “remember” what you said.

Until… it doesn’t.

Let’s look at why.

3. Token

Why it matters:
This is how AI actually processes language.

Definition:
A token is a small piece of text (often a word or part of a word) that AI uses to read and generate language.

Picture this:
Lego bricks.

Each word, sentence, or paragraph is built from smaller blocks.

It’s like:
Breaking language into bite-sized pieces so the model can process it step by step.

What’s really happening:
AI doesn’t read sentences the way we do.

It processes tokens, one after another, predicting what should come next.

That’s how it builds responses.

And decides how much to charge (think of tokens as tiny units of electricity).

2. Context Window

Why it matters:
This determines how much the AI can “remember” at any given time.

Definition:
The context window is the amount of text (tokens) the AI can consider when generating a response.

Picture this:
A whiteboard.

You can write on it, refer back to it… but eventually it fills up.

And when it does, things start getting erased.

It’s like:
A rolling memory that only keeps what still fits.

What’s really happening:
AI doesn’t have long-term memory in a conversation.

It only sees what’s currently inside its context window.

Once that fills up:

• earlier details drop off

• instructions get lost

• consistency starts to drift

So what?

This explains a lot of common frustrations:

• “Why did it forget what I said earlier?”

• “Why did it suddenly change direction?”

• “Why is it contradicting itself?”

It’s not being difficult.

It’s running out of space.

If you remember one thing:

AI doesn’t remember everything.
It remembers what still fits.

Coming next

Why does AI sound so confident… even when it’s wrong?

In the next post:
Inference and Hallucination
(and why convincing answers can still be unreliable)

If you’re finding this useful, you might want to follow the AI Executive’s Dictionary series.

Two terms at a time so you don’t get AI overload.

The full AI Executive’s Dictionary goes deeper into each concept, with practical examples and applications. If you don’t want to wait, you can get the full dictionary here.

There’s a cartoon doing the rounds where ChatGPT is in the witness box, hand on heart:

“I, ChatGPT, swear to tell the truth…”

And that pretty much captures how many people experience AI.

Open ChatGPT. Ask a question. Get an answer. Read an article about ChatGPT inventing facts. Conclude it’s unreliable.

End of conversation.

That reaction is understandable, but it’s also incomplete.

We’re treating AI like a witness, when it’s closer to a very fast, very confident junior analyst who needs supervision, framing, and checking.

Used casually, it will sometimes mislead. Used deliberately, it can help you explore ideas, test assumptions, stress-check arguments, and surface blind spots, if you understand what it’s good at and where it needs constraints.

The problem isn’t that AI occasionally gets things wrong. The problem is expecting it to behave like settled, finished technology when it’s clearly still a work in progress.

If AI is becoming part of how we work, the real question isn’t whether it ever makes mistakes, it’s whether we know how to work with it responsibly when it does.

We talk about AI like it’s intelligent.

It sounds confident.
It explains things clearly.
Sometimes it even feels like it understands you.

But here’s the uncomfortable truth:

AI doesn’t know anything.

Let’s see why.

1. Large Language Model (LLM)

Why it matters:
This is the engine behind tools like ChatGPT, Claude, and Gemini.

Definition:
A Large Language Model is an AI system trained on vast amounts of text to generate human-like language.

Picture this:
A librarian who has read the entire internet and can instantly summarise, rewrite, or continue any piece of text.

It’s like:
Autocomplete on steroids. Not just finishing your sentence—finishing your next idea.

What’s really happening:
An LLM doesn’t “think” in the human sense.
It predicts the most likely next word based on patterns it has seen before.

That’s it.

No understanding.
No beliefs.
No awareness.

Just probability.

2. Training Data

Why it matters:
This is where the AI gets everything it “knows.”

Definition:
Training data is the information used to teach an AI model how language, ideas, and patterns work (we won’t go down the ethical rabbit hole of how AI companies acquired that data here).

Picture this:
A student learning from millions of flashcards. Questions on one side, answers on the other.

It’s like:
A memory bank that shapes how the AI responds to new situations.

What’s really happening:
The model has absorbed patterns from enormous datasets:
books, articles, websites, conversations.

It hasn’t verified them.
It hasn’t judged them.
It hasn’t “understood” them.

It has simply learned:

“When people say X, Y usually follows.”

So what?

This is where it gets interesting.

If AI doesn’t know anything, then:

• It can sound right while being wrong

• It can be confident without being accurate

• It can generate answers that feel true but aren’t grounded in reality

And that leads to one of the most important ideas in AI literacy:

AI is not a source of truth. It’s a generator of plausible language.

If you remember one thing:

AI doesn’t think. It predicts.

Use it like a fast, articulate assistant.
Not an infallible expert.

Coming next

Why does AI feel so convincing… even when it starts to drift?

In the next post:
Tokens and Context Windows
(the hidden mechanics behind memory, coherence, and confusion)

If you’re finding this useful, you might want to follow the AI Executive’s Dictionary series.

Two terms at a time so you don’t get AI overload.

The full AI Executive’s Dictionary goes deeper into each concept, with practical examples and applications. If you don’t want to wait, you can get the full dictionary here.

TLDR;

Anthropic made a series of ads teasing OpenAI for monetising ChatGPT with ads (ironic).

They said this on their website:

“We want Claude to act unambiguously in our users’ interests. So we’ve made a choice: Claude will remain ad-free. Our users won’t see “sponsored” links adjacent to their conversations with Claude; nor will Claude’s responses be influenced by advertisers or include third-party product placements our users did not ask for.”

OpenAI are between a rock and a hard place because they have to somehow start paying back the billions in venture capital they have received. If Gemini (with ads) and Claude (no ads) are perceived as better and get more traction could ChatGPT become the MySpace of the AI Age?

The other Youtube ads:

___________________________________________________________________

THE GEMINI CONVERSATION:

AI Competitors’ Ad Policies Compared

OK, I’m going to ask Claude’s competitors their response to this statement by Anthropic: We want Claude to act unambiguously in our users’ interests. So we’ve made a choice: Claude will remain ad-free. Our users won’t see “sponsored” links adjacent to their conversations with Claude; nor will Claude’s responses be influenced by advertisers or include third-party product placements our users did not ask for. What are Microsoft/Gemini’s values in this respect?

That statement from Anthropic highlights a key philosophical divide in the AI industry. While Claude has positioned itself as the “premium, ad-free” alternative, Google (Gemini) and Microsoft (Copilot) have deep roots in the advertising industry and are taking a different approach.

Both companies generally argue that ads allow them to make high-end AI accessible to everyone for free, rather than locking it behind a $20/month subscription.

Here is how Microsoft and Gemini’s values break down in contrast to Anthropic’s:

Google Gemini: The “Search Heritage” Approach

Google’s philosophy is built on the idea that commercial information is often useful information. They don’t necessarily see a “sponsored link” as an intrusion, but as a potential answer to a user’s intent.

• Ads in the AI Ecosystem: Google has already integrated ads into AI Overviews (the AI-generated summaries at the top of Search). They believe that if you’re asking “how to fix a leaky pipe,” a link to a local plumber or a specific sealant is a helpful part of the “user’s interest.”

• The Gemini App: As of early 2026, Google leadership has reiterated that they have no current plans to put traditional ads inside the dedicated Gemini chatbot app. However, they are actively testing “ad-like” features in “AI Mode” and have told investors that they are exploring how to “monetize the gap” between search and chat.

  +1

• Core Value: Google emphasizes universal access. Their view is that an ad-supported tier is the only way to bring Gemini to the next billion users who cannot afford monthly fees.

Microsoft Copilot: The “Showroom” Approach

Microsoft has been more aggressive and transparent about incorporating advertising into Copilot (formerly Bing Chat) from the beginning.

• Sponsored Content: Copilot frequently includes “citations” that are essentially sponsored links. If you ask for travel recommendations, Copilot may prioritize partners like Booking.com or TripAdvisor within its response.

• The “Showroom” Concept: Microsoft is pioneering “Showroom ads,” where the AI invites you into an immersive digital space to explore a product. They argue this isn’t “third-party placement you didn’t ask for,” but rather a richer way to shop once the AI detects you have “purchase intent.”

• Core Value: Microsoft focuses on productivity and commerce. They believe the AI should be a “copilot” for your entire life—including your shopping life—and that clear labeling of sponsored content maintains user trust.

Comparison at a Glance

FeatureAnthropic (Claude)Google (Gemini)Microsoft (Copilot)Primary ValueUnambiguous user interestUniversal access & utilityProductivity & commerceAds in ChatStrict No“No current plans” (in-app)Yes (integrated in links/citations)Revenue ModelSubscriptions & API feesAds (Search) + SubscriptionsAds (Bing) + Enterprise SeatsData UsageMinimal for trainingHigh integration with Google dataHigh integration with Microsoft 365

The “Subtle Influence” Question

Anthropic’s most pointed claim is that Claude’s responses won’t be “influenced by advertisers.” This is a gentle dig at the “black box” of AI: if a model is trained or funded by ad revenue, will it subtly suggest a certain brand over another?

• Google/Microsoft’s Counter-Argument: Both companies maintain “Answer Independence” policies, claiming that their AI’s organic reasoning is separated from the ad-auction layer. They argue that as long as an ad is clearly labeled, the integrity of the AI’s “brain” remains intact.

Would you like me to find the specific “Responsible AI” or “Ad Transparency” whitepapers for either company to see how they legally define these boundaries?

yes please

Building on Anthropic’s stance, Microsoft and Google have published various “Transparency Notes” and “Responsible AI” reports that detail how they handle commercial influence. While Anthropic views ads as an inherent conflict of interest, Microsoft and Google view them as a “utility” that can be integrated responsibly.

Here is a breakdown of their values and legal/policy positions as of early 2026:

1. Microsoft (Copilot): “Commercial Intent as a Service”

Microsoft’s value proposition for Copilot is that it should be a partner in all life tasks—including shopping and decision-making. They do not see “third-party placement” as a negative, provided it is transparent.

• Transparency Policy: Microsoft’s Transparency Note for Copilot explicitly states that when a user shows “commercial intent” (e.g., asking for product recommendations), the AI will prioritize high-quality web data which may include sponsored citations.

• Ad-Voice Technology: Internal documentation describes an “ad voice” philosophy. Instead of “sponsored links” sitting on the side, Microsoft aims to make ads feel like a “continuation of the chat.” Their policy is that ads should “add value” to the conversation rather than interrupt it.

• Business Model: Microsoft offers a clear divide:

  • Consumer Copilot: Ad-supported and integrated with Bing’s commercial ecosystem.

  • Copilot for Microsoft 365 (Enterprise): This version has a strict “No Ad” value. According to their Service Terms, data in the enterprise version is not used to train models, and users do not see commercial influence, as the tool is strictly for productivity within a company’s own data.

2. Google (Gemini): “Organizing Information with Integrity”

Google’s values are rooted in their “Ten things we know to be true” philosophy, specifically: “Advertising is not only clearly marked as such, it offers relevant content and is not distracting.”

• Neutrality Principles: Google’s AI Principles Progress Report emphasizes “neutral and inclusive design.” They claim Gemini is trained to provide a balanced range of views on divisive topics. However, they stop short of promising that product recommendations will be ad-free.

• The “E-E-A-T” Value: Google applies its Search standards—Expertise, Experience, Authoritativeness, and Trustworthiness—to Gemini. Their value system suggests that if a “sponsored” answer is the most authoritative or helpful, it belongs in the response.

• The Privacy/Ad Wall: In their Gemini Apps Privacy Hub, Google notes that they use data to “personalize and customize services.” Unlike Anthropic, which avoids this to prevent ad-profiling, Google views personalization as a way to make the AI more “helpful,” even if that help is commercial in nature.

Summary Comparison of Values

Value AreaAnthropic (Claude)Microsoft (Copilot)Google (Gemini)User Interest“Unambiguous”: Only what you asked for.”Productive”: Includes helping you buy/do.”Universal”: Helping you find info (ads included).IntegritySeparation from advertisers is the only way to be objective.Transparency and clear labeling are enough to maintain trust.Relevance is the key metric; ads are useful if they are relevant.Revenue Philosophy“Pure-play” subscription/API model.Multi-tier: Ads pay for the “free” tier for the masses.”Search Heritage”: Commercial info is part of the world’s info.

The “Whitepaper” Reality

If you read the fine print in their Terms of Service:

• Anthropic essentially promises that you are the customer.

• Microsoft and Google (in their free versions) treat you as a user whose attention and intent are part of their monetization ecosystem. They argue this is a “democratic” value, allowing the world’s most powerful AI to be free for people who can’t pay $20/month.

Which of these values resonates more with how you plan to use AI? (e.g., Are you looking for a pure research tool, or a “shopping assistant” that knows your preferences?)

I would like a quick TLDR; summary of the 3 positions for my Substack hero image

To fit a Substack hero image, you need high-impact, punchy text that captures the “vibe” of each AI’s philosophy. Here are three ways to summarize their positions, depending on the aesthetic of your article:

Option 1: The “Mission Statement” Style

Best for a clean, professional look.

• Claude: The Purist. Unambiguous loyalty to the user; 100% ad-free by design.

• Gemini: The Librarian. Universal access for everyone, funded by the “Search” ecosystem.

• Copilot: The Concierge. A productivity partner that blends work, search, and shopping.

Option 2: The “Follow the Money” Style

Best for a tech-analysis or business-heavy post.

• Anthropic: User = Customer. You pay for the product; you are the only priority.

• Google: User = Audience. Ads scale the tech to billions; relevance is the North Star.

• Microsoft: User = Professional. Free for shoppers; premium for the private boardroom.

Option 3: The Ultra-Short (One Word) Style

Best for a bold, minimalist graphic.

• Claude: Neutrality.

• Gemini: Utility.

• Copilot: Connectivity.

Pro-Tip for your Hero Image:

If you are using an image generation tool (like the one I have!) to create this header, you might want to visualize them like this:

• Claude: A quiet, minimalist library.

• Gemini: A bustling, infinite digital archive.

• Copilot: A high-tech, integrated cockpit.

So, I asked Gemini to create a short video to raise awareness for World Hearing Day on the 3rd March.

The result:

• nice production values

• appropriate background music

• did a lot of heavy lifting in just 8 seconds

• the text overlay was complete garbage

Yes, out of 12 words, 6 were misspelt.

3 were quite ‘out there’: clasbblooms, pristitize, and headiring.

This problem with text has been a known issue for a long time, and there are workarounds for images.

For example, when I generate a cartoon in AI, I tell it to just give me the image and I add the speech and thought bubbles in Canva.

But, according to Gemini, workarounds could soon be unnecessary. Help is on the way in the form of “text encoders” and “spatial-temporal consistency,”

For a range of AI Literacy for Business reources check out my ebooks here

——————————————————————————————————————

Here’s the full conversation with Gemini:

AI may be able to write an entire essay or code a website, but making it spell “B-A-K-E-R-Y” on a shop sign is a different beast entirely. It’s a classic case of “The Moravec Paradox”—where things that are hard for humans (like complex math) are easy for AI, but things that are easy for humans (like seeing a typo) are incredibly difficult for machines.

Here is why AI still treats text like “alphabet soup”:

1. It Doesn’t See “Letters,” It Sees “Shapes”

To an AI, a letter is not a linguistic building block with a sound and a meaning; it is just a pattern of pixels. When you ask for a “Coffee Shop,” the AI knows what a shop looks like and what the texture of “text” looks like. It attempts to replicate the vibe of text rather than the specific characters.

• Result: You get symbols that look like letters from a distance but turn into gibberish or “alien languages” upon closer inspection.

2. The Diffusion “Big Picture” Problem

Most modern image and video generators use Diffusion Models. These work by starting with a “cloud of noise” (like static on a TV) and gradually refining it into an image.

+1

• The Trade-off: These models are excellent at global structures (the sky, a person’s face, lighting) but struggle with tiny, precise details.

• The “Baking In” Effect: If the AI makes a mistake early in the denoising process—like putting two “O”s in “STOP”—it can’t easily go back and erase one without ruining the rest of the sign’s texture.

3. The Tokenization Barrier

This is the most technical reason. AI “thinks” in tokens (chunks of characters), not individual letters.

• A Large Language Model (like the one I’m using to talk to you) sees the word “Apple” as one or two tokens. It doesn’t necessarily “see” the “A,” the “p,” and the “l” unless it specifically looks for them.

• When a text model tells an image model what to draw, that “loss of detail” carries over. The image generator is essentially trying to draw a concept it only partially understands at a character level.

4. Training Data Noise

AI models are trained on billions of images from the internet. Many of those images have:

• Stylized text: Think of graffiti, cursive, or blurry background signs.

• Low resolution: Small text in the background of a photo is often just a smudge. The AI learns that “signs have blurry, squiggly shapes on them” and reproduces exactly that.

Is it getting better?

+2

Yes, rapidly.

• DALL-E 3 and Midjourney v6+ have introduced “text encoders” that specifically focus on character placement.

• Video models (like Veo or Sora) are beginning to use “spatial-temporal consistency,” which helps text stay the same from one frame to the next instead of morphing into different letters.

Enjoy!

(Please feel free to share the link with anybody you think might find the book useful)

https://www.amazon.com/dp/B0GFH28CCH

https://www.amazon.ca/dp/B0GFH28CCH

https://www.amazon.co.uk/dp/B0GFH28CCH

https://www.amazon.de/dp/B0GFH28CCH

https://www.amazon.fr/dp/B0GFH28CCH

https://www.amazon.es/dp/B0GFH28CCH

https://www.amazon.it/dp/B0GFH28CCH

https://www.amazon.nl/dp/B0GFH28CCH

https://www.amazon.co.jp/dp/B0GFH28CCH

https://www.amazon.com.br/dp/B0GFH28CCH

https://www.amazon.com.mx/dp/B0GFH28CCH

https://www.amazon.com.au/dp/B0GFH28CCH

https://www.amazon.in/dp/B0GFH28CCH