What is the issue?

The details don’t really matter. The big picture is this: sometimes these records are very helpful in prosecuting people who did very bad things. Is that really different than allowing your personal journal in court or your phone records? I think it is.

Why we care

The thing is that there are certain protected, privileged relationships. Your religious leader, your doctor, and your lawyer. You can tell those people things that no one else ever will hear. It is a guaranteed right. (*there are special exceptions, but they are very circumscribed*) At least over here in the US. As a result, criminals can discuss very bad details with their lawyers and it stays private. Patients can discuss thoughts of harming others, or themselves, with no penalty. People can talk to their local religious leader about wrongdoing, and get solid advice without punishment. These are really important things.

Which tools?

What about chatbots like ChatGPT or Claude or Grok? Or more broadly, our AI tools, since AI has been graduating to bigger, more capable things like Codex, Claude Code, Grokbot, Open Claw and more. These agentic systems are not “chatbots”. They do things for us, they check things for us, they build things for us, etc. All these AI tools have increasingly been entering our private lives - our hard drives, our email, our private thoughts, our private conversations. We let them in because they bring tremendous value. Nevertheless, there need to be assurances that these things are private.

Do you personally care? Should you?

Here is an experiment. Think about the most embarrassing thing you have ever written or said or shown to an AI. Would you like that to be put into the public court record? If your answer is no, you agree we now have a problem. If your answer is more like - “what do you mean, embarrassing? I would never do that!” then you might be ok for now, assuming you have thought through the meaning of “ever written or said or shown”. Most people don’t realize how much privacy is lost with “Hey Siri” and “Alexa” and Meta glasses and AI tools that can see the contents of their computer.

My opinion

Anyway, in my biased opinion, I think it is time for us to consider something as a society. More and more, people will turn to AI for private but important conversations or interactions. If they thought they could not have those, they would not turn to AI. Think about the teenager who wants to understand the changes in their body, the college kid who wants to discuss the difference between consensual and non-consensual activities, the banker who wonders if their actions might be crossing the line into fraud, the person who wants to know without religious guilt what to think about their body’s urges, the soldier who wonders if this kind of act might be crossing into the realm of treason, and so on. If there were a safe space to have these conversations, society broadly would benefit. If that space exists in AI, why would we squander that?

I suggest we make a button called “AI privilege” which can be turned on whenever we want. When it is on, no court or law enforcement can ask for it. When it is on, no spouse or parent can demand a copy. But when it is on, the user is allowed to explore thoughts and ideas without judgment.

Feel differently? Tell me about your perspective. I’m sure there are many angles to consider.

“Relentless incremental improvement”

Leave a comment

General AI updates:

• Some new models: GLM 5.3 and 5.3-flash, Qwen 3.8-Flash-Next, Gemini Omni 1.1 Flash

• Rumors of new models: Gemini 4 is coming? GPT 6 or Astra is coming? Claude Fable 5.1 and Opus 5.1 are coming?

• NVidia is probably buying HuggingFace. This is important because of their respective identities.

• Anthropic offers 10,000 free accounts to scientists, builds a standard protocol for AI to run physical experiments called MHS,

• OpenAI releases details and 2 reports about the “hacking incident” and the details continue to be surprising.

• OpenAI has co-designed a chip (Jalapeño) that seemingly beats NVidia for computational efficiency on inference, they talk about benchmarks here.

Neurology AI updates:

• UCL reports live-AI assist for brain surgery. I am impressed less about the details, more about the paradigm shift that this represents.

• A new ARPA-H deal for Philips, Hopkins and BU will combine imaging, robotics and AI to help stroke intervention.

• New data from Hyperfine suggests that their AI-enabled MRI (Swoop) which uses low field strength and is portable is quite similar to traditional MRI when tested in blinded review.

• The Westover team published data on MORGOTH, which is now redefines state-of-the-art in EEG analysis across many domains, including spike detection, seizure detection, ICU patterns, sleep staging and more. The tool works at the event and entire report level. Of course it is open-source.

Did I miss something cool from the past week? Let me know.

Leave a comment

Up until recently, the answer was not really. Many special things were done by EEG readers that AI couldn’t manage. But something changed with a recent publication (8/25/26) by Sun et al from Dr. Westover’s lab - the MORGOTH paper (PMID: 42642264).

It was not easy before

There are probably lots of reasons for why it was not too easy until now. Here are a few:

• Lack of available high-quality datasets

• Strong inter-rater variability (experts disagree a great deal)

• Many sources of “noise” in EEG

• Inadequate AI tools

• Poor validation

Several years ago, some important advances got the ball rolling. SpikeNet in 2020, ScoreAI in 2023, Sparcnet in 2023, and SpikeNet2 in 2025 all showed that parts of EEG automation seemed possible. At the same time, AI tools were advancing, labs were beginning to collaborate across centers and countries, and huge datasets such as HEEDB were beginning to appear.

It is easier

Then Morgoth happened.

According to the paper, just published, they used (1) a massive dataset with many experts, (2) many centers, and (3) internal and external datasets. It exceeded 20% of experts on 17 tasks that EEG interpreters do, and exceeded 90% of experts on 3 multi-expert tasks. These include things like annotating epileptiform activity, seizures, slowing, ICU-EEG patterns, and sleep stages. Moreover, the tool was explicitly tested at the event-level and the whole-EEG level compared to experts on multiple categories. They also checked for effects of age, sex and channel loss - all quite favorable.

(Figure 2 from MORGOTH paper PMID: 42642264 which is open-access)

Are we in trouble? Is this bad?

For those neurologists who make money by interpreting EEGs, the above paragraph looks pretty scary. Will the robots take my job? Is EEG reading over? Will the government stop reimbursing? I don’t think we are there… yet. But let’s be honest - the writing has been on the wall for several years now. This convergence of data, technology, research willpower, and clinical necessity is making the complete automation of EEG interpretation inevitable. Will it happen this year? I don’t think so. Will it happen soon? Not completely. Will it happen in coming years? Almost certainly.

The SCORE-AI study should show the way. They published this study, then published a validation showing it really does work. They built the tool into a popular EEG reading system, and even the creator of SCORE-AI has not stopped reading EEGs himself. Rather he teaches others to use it as a speed boost for productivity, since some aspects of SCORE-AI work very well, and others are far from complete or trustworthy.

Will MORGOTH work the same way? Will we build it into our tools and work faster? Yes, for now. Even the MORGOTH paper itself hints at more powerful capabilities coming soon, and since I have a bit of a front-row seat on that, I can attest to more cool things coming from Dr. Westover’s highly productive lab.

The MORGOTH paper mentions a few of the limitations. It doesn’t annotate breach or phoroparoxysmal response. It assumes standard 19 EEG channels without extra electrodes, and without EOG. It does not account for clinical context. The does not perform accurate localization and seizure typing. Though the paper does not say so, the tool can still overcall seizures and spikes as prior AI tools have done.

Is it time to stop training EEG readers? I don’t think so. Is it time to stop reimbursement for EEG interpretation? Still no. At the same time, has the time come to begin to rethink how these things work, because… the day will come. It might come soon, or in some number of years. But it seems unlikely we will always use visual analysis of voltage patterns from brains to diagnose patients in the same way we’ve been doing for >100 years now that we have advanced systems capable of much better accuracy than us.

How can the informed EEG reader prepare for this?

Start working with automated EEG tools

• Develop both literature-based and also experiential knowledge of the strengths and weakness of the automated tools - current and new ones as they become available

• Diversify sources of income (e.g. - read some EEGs, see patients in clinic, take care of inpatients, etc)

• Imagine what you will do if next week reimbursement goes to zero for EEG interpretation. Will you lose your home? Will you need to go back to training for new skills? Will you be devastated? If you answered yes, the time to begin preparation is right now.

Conclusion

Again - the sky is decidedly not falling. Even MORGOTH makes embarrassing mistakes that EEG fellows after 6 months of training would never make. But MORGOTH is getting smarter. And things in AI are changing very fast. So it is time to begin the evolution of what an epileptologist thinks about and does.

Daniel Goldenholz

“Relentless incremental improvement”

I will present two hypothetical stories.

Version 1: Once upon a time, a certain Famous Hospital IT system decided that it would use new software to ensure that a very important policy (HIPAA in this case) would be followed, and that employees do not violate it by entering protected health information (PHI) into non-secure AI tools. To do this, they checked the exact text that employees entered into certain websites, and made sure there was no PHI included. If there was a match, the software would prevent the PHI from being entered. In this way, the IT team protected the privacy of the patients.

Version 2: Once upon a time, a small group of people, who were not elected or chosen on moral grounds, decided that they could abridge privacy and spy on employees of their company. This decision was not granted to them by the employees - this decision was made in private. Spying was ok because it was for a good cause. The spying began with detecting PHI in non-secure websites on any hospital computer.

It seems to me that version 1 sounds good, and version 2 is pretty close to terrible. This same hypothetical hospital IT department puts special software on mobile devices, tablets and laptops used by thousands of employees. Many of those devices have cameras, microphones, and likely all of them are internet connected. If the IT department felt that some hospital policy or some law was important, they could also make use of this vast network to help enable that policy or law as well. For example, there is a policy that says clinicians should not talk about their patients in public spaces. That is definitely a good policy. Should the IT department be allowed to use mobile phone microphones to spy on every conversation in order to ensure that rule is followed and if AI detects a violation for the phone to start a loud siren? That seems pretty bad. Another example, there is a policy that says that workers should be allowed to have a safe working environment, free from discrimination or gender harassment. This is clearly a good policy. To make sure it works, should the IT department activate cameras from cellphones, laptops, desktop webcams, and send all video for AI review to make sure there are no violators? That seems to me like a police state, similar to what was described in 1984. It does not sound like a place I would want to work.

So I would propose the following idea:

If a policy or law is good, all methods for enforcing that policy or law may not be good. In fact, some methods may violate other societal goods and be a net bad.

To conclude - in my opinion, not knowing all the variables that were considered by the IT department of our hypothetical Famous Hospital, it nevertheless seems to me that a Line Has Been Crossed when the IT department is actively using AI to spy on employee activities. By crossing it, they open the door to additional behaviors that can be justified yet would seem extremely invasive to many. In their defense, they might not call this original act spying - they might call it “protecting the patients from harm”. However, real spies also use this kind of re-definition trick to avoid describing their behaviors as socially unacceptable.

What do you think? Have I taken this too far? Am I over-reacting to a well-meaning policy that is in the best interests of the patients of Famous Hospital? Am I claiming a Slippery Slope when none exists? Or am I looking past the moment and seeing precedent being set? Am I worrying about tomorrow and the next day and the next policy? Is anyone worried about the possibility that AI can make mistakes, and will classify innocent behavior as a violation?

Lets see what subscribers think:

Sincerely,

Daniel Goldenholz “Relentless incremental improvement”

Carmichael ST. “Five Issues of Artificial Intelligence in Science: Sailing the Ship of Theseus.” Annals of Neurology. 2026;100(3):453–457. doi:10.1002/ana.78298

https://onlinelibrary.wiley.com/doi/full/10.1002/ana.78298 (it is open access - you can read it)

Anyway, his paper is interesting because, unlike some commentators, Dr. C clearly sees nuance in the debate about AI vs humans in the realm of creativity. Some might suggest that AI cannot be creative, and any attempts that appear creative merely devalue human work. Such a stance is too extreme for me.

He brings up 5 important concerns. I will cartoonify them, and then address them. What I mean is that I am subtly changing his message, and addressing that version. Please read his excellent article and decide how decent my cartoons are.

His 5 problems:

1. Language problem: AI is improving everyone’s language, so it is harder to see the difference between good and bad grants.

2. Agency problem: scientist + AI ideas are deeply not equivalent to scientist ideas

3. Review problem: AI is shifting ideas to look more like ideas that were good from the past.

4. Doership problem: AI will propose grants/ideas that the human-run lab cannot do

5. Identity problem: after AI gradually changes us enough, when do we become a mirror of the AI that molded us?

My answers are:

1. Language - no problem. We want better language for everyone. This is a good thing. The fact that we used poorly written text as a proxy for poor science never made sense, but we did it. Those days are ending. Now we can focus more carefully on the science. Imagine this: native English speakers and non-native English speakers should be given equal weight if they both propose good ideas. Cool, right? I agree with Dr. C that we don’t want bad science to look like good science. But equalizing the quality of language won’t do that. Dressing up a cat to look like a dog won’t change the fact that it is a cat. Bad science still comes with a bad science smell - poor methodology, shoddy conclusions etc, but it can be written in very proper scientific language.

2. Agency problem - no problem. AI+ scientist ideas can be better than scientist alone. In those cases, we want those better ideas. The goal of science is not to have great human originated ideas. It is to search for truth. We can be assisted by our rulers, calculators, coffee, computers and yes, our AI in that search. Moreover, suppose that we get worse ideas when the “intellectual center has shifted”… if that happens, we hopefully will have enough judgement to discard those. I disagree with Dr. C’s implied view that the “agency” of the scientist is primary. I would argue that the science is primary, and the scientist is (currently) the vehicle that delivers it.

3. Review problem - no problem. This is gesturing at a real problem, but AI is exposing and potentially amplifying a defect that already exists in grant review of human-made grants. The right response is to redesign review. AI isn’t the issue. To elaborate: our current human system strongly rewards ideas that look like ideas that were good in the past. That entire process is deeply flawed, and it is one of the reasons we invest so much and get relatively so little in return. Don’t get me wrong - I absolutely think we should invest so much in basic science. But many of the things that get funded are… not designed to change much of anything. Grant applicants are incentivized to not be bold and revolutionary, but rather very small incremental steps. So Dr. C’s criticism is an excellent critique of the whole system - we should never be incentivizing that kind of behavior, and we (collectively) have been for decades. This is a great time to rethink that. Perhaps we should reward ideas that don’t look like ideas that worked in the past? Sure, many will fail, but so do many of the opposite kind.

4. Doership problem - no problem. If the AI churns out ideas that are infeasible for a lab to do, the AI will also need to churn out lies or exaggerations about the capabilities of the lab in order to convince funders that the thing is feasible. If the author signs off on those lies…they should be held accountable. This gets back to basics: if AI lies and you sign off…you are responsible.

5. Identity problem - no problem. As I mentioned above, the project of science is discovery of truth. If gradually my inferior ideas are replaced with correct different ideas from an outside intelligence, hey! That is what we were hoping for! Truth! I don’t want to be a scientist with wrong ideas. As a scientist, I am prepared to throw away my wrong ideas and replace them (as many as needed) with correct / true ideas.

Take all this with a grain of salt. The article prompted me to write this whole rebuttal. Clearly it was thought provoking either way.

Meanwhile, society’s views and mine continue to evolve as we learn more. We will see where this adventure leads.

To give you a flavor, a locally run version of Qwen 3.8 came out which was very strong, grokBot is getting a lot of attention because it seems that many are finding it useful - more for the style that it works than anything else, the politics around building data centers is getting quite heated, there was an exciting announcement from Moderna that I wrote about yesterday regarding an AI-powered mRNA treatment for melanoma.

I’m sure I missed something, so feel free to let me know in the comments.

Daniel Goldenholz

“Relentless incremental improvement”

https://www.modernatx.com/media-center/all-media/blogs/advancing-fight-against-cancer

Longer term? My prediction is this will become the new normal - I hope soon. We will sample the patient’s complexity, show it to AI, find a specialized treatment appropriate for them and them alone, and that treatment will be better than standard therapy.

How do I keep up with all this stuff.

Short answer: I do not. Just like Epileptology, and just like Neurology, it is impossible to be fully up to date. Unlike those two, the field moved MUCH faster. And like those two, one can certainly TRY to stay a bit current.

Long answer: My method, which I don’t claim is best, includes a nearly continuous diet of YouTube, email lists, journal emails, reading journal articles, trying out many different AI tools, talking with colleagues, collaborators and complete strangers, and yes, also learning from kids.

The following list is incomplete, but really helpful if someone wanted to try to drink from firehose:

Youtube (watch at 2x): Matt Berman, MattVidPro, Matt Wolf, Machine Learning Street Talk, Wes Roth, AI explained, Dwarkesh Patel, Ai Search, Theo-T3.gg, Futurepedia, AI Daily Brief, CNET, Nate B Jones, ThePrimeTime, Lex Friedman, Prompt Engineering, OpenAI, Anthropic, HowIAI, Alex Kanrowitz, Karpathy...

Email lists: Future Forward, Matt From FutureTools, The Batch @Deeplearning.AI, Natural 20, Superintelligence, WSJ AI Business, DataPoints @Deeplearning.AI, Sam Charrington (TWIML AI)

Journals to keep a watch on: JAMA AI, NEJM AI

Pubmed searches: pubmed can email you any new articles on any specific subtopic search set you want at any interval you want. I have a bunch of these going related to my own research interests, and to more general AI and neurology search terms

I also subscribe to several AI services to try our new features and to get Real Work from my lab done: Google Gemini, Anthropic Claude, OpenAI ChatGPT. There are some free services one can use in limited ways like Qwen, Kimi, Grok, Deepseek, etc which I might test once in a while

Local AI hosting: using a strong desktop computer, one can test out locally running (Llama, Qwen, Gemma, Ministral, Deepseek, etc) AI tools with various tools. I tend to favor Ollama, but there are others too.

Anyone who drinks from the firehose and has other useful high impact sources for news and up knowledge, I’m always learning and I have been asked more than once how I do this, so please share with me (don’t reply all).

• Several very strong Chinese model updates: GLM 5.3, Deepseek v4 0813 Pro

• Several US updates: Grok 4.6, Grokbot, Meta Muse Glimmer, NVidia Nemotron Lightning and Switchyard, Google Gemini Flash 3.7

• Claude recently figured out how to loosen the restrictions on Fable 5. It now will work with you on neurology things. It is worth checking out.

• OpenAI: announced but did not widely distribute “Ultrafast mode” which is 14x faster for their strongest models

• OpenAI also announced Computer History - a way for Chatgpt to help you figure our what you did with your computer - EVERYTHING you did…

• A lot of detailed developments in cybersecurity too complex for me to get into, but mostly good

• Thanks to EU AI Act, it will now be perhaps easy to detect AI generated text and pictures from compliant AI labs that will “watermark” their AI generated images AND text. I say perhaps because there will certainly be ways to defeat these hidden markers, and only the laziest users will be unaware of those ways. High school/college kids will still cheat. Political opponents will still smear. This does not sounds like it will fix much.

One of the most seemingly innovative stories is the Grokbot, which is a variation on the OpenClaw / Hermes always on always helping AI tool. This thing was designed seemingly for “normal people” because even though it uses coding and web searching and math etc, the interface looks more like a texting app - you send text messages to different agents who help you with things. My two cents - sounds nifty, but realistically requires $100/month SuperGrok service. Not a strong value proposition for “normal” people. Not yet.

In the neurology space directly, my lab published a fun article about PAT, the Paper Analysis Tool, that scientists can use to check their manuscript before submission to make sure they follow best practices. It can be used as a Skill for your favorite AI in the cloud, or as a locally run AI that never shares your paper with the big AI companies. We tested it out on multiple papers and gave it an honest review. The paper is here: https://www.sciencedirect.com/science/article/pii/S1386505626003904?dgcid=coauthor. The github repo for PAT is here: https://github.com/bdsp-core/PAT-PaperAssessmentTool.git

In my own exploration, I want to share two anecdotes.

1. An incredible way to read journal articles. As part of MOC, I was reading some neurology papers. The problem is, a) that is boring, b) they often use jargon I don’t know, and c) sometimes the paper raises questions and no one is there for me to ask. The solution: open voice mode on ChatGPT and then read the paper. As soon as a question comes up, just say it out loud. “What does that acronym mean?”, “what years were the studies that they mentioned published?”, '“Are there other studies that have come out about that in the years since the publication of this paper?”, “can you explain what they are trying to do here in Figure 4?” and so on. It is 1000x more engaging, 10x more helpful, and much more enjoyable overall. I highly recommend it.

2. I am working on a very very large project and AI is helping me. We came upon a problem - one of our simulated results does not look like a physiologically reasonable result. So I worked with AI to track down why, and then plan and execute a fix. Of course, doing that meant that all downstream results that depended on the simulator needed to be updated. Plus all the results from our pending paper need to be updated too. So AI helped with all those tasks. That is cool, but not new. What is cool here- is that same process happened to me 3 more times, and so I had to effectively rebuild a giant project 4 times, and re-run all our analysis over again. That would have taken weeks. But with AI helping? It took a few hours. We can now demand much more from our projects, and if we need to “go back and recalculate everything” - that is no longer the dealbreaker it once was.

That’s all for now. Keep learning.

Daniel Goldenholz

“Relentless incremental improvement”

Subscribe now

1. Why this, why now

I have been writing to the American Academy of Neurology for quite some time, as well as to my neurology community locally. I continuously receive feedback from people that they appreciate the updates I provide. Maybe others can benefit?

2. What kind of community are you looking to build here

I am an engineer and a neurologist, fighting against neurological disease in general, and epilepsy in particular. I use advanced data science and AI tools to help. Anything I learn along the way that can help others I am happy to share. No one wants a neurological disease. Who would most benefit? Neuroscientists, neurologists, engineers working in the neurology field - all of whom are interested in AI being applied to their work. Anyone else that is interested in these topics of course.

3. Be specific

I don’t know how often I’ll post - mostly weekly? Sometimes a lot, often nothing. I will probably post cool updates in AI that I think will be of interest to my community, that will help someone do better clinical or research work in neurology. I’ve been doing this anyway elsewhere, so more here.

4. Who am I?

I have a PhD in Biomedical Engineering and an undergrad in Electrical and Computer Engineering. I have an MD, board certification in Neurology, Clinical Neurophysiology and in Epilepsy. I work at Harvard Beth Israel Deaconess Medical Center, currently as an Assistant Professor. I also have formal training in biostatistics, neuroradiology, and neurostatistics. I informally have a lot of machine learning and AI training from online courses, personal experience and collaborations.

Some of my titles:

• PI of the Epilepsy + Data Science Lab.

• Fellow of the American Epilepsy Society

• Director of Neurology Artificial Intelligence Innovation

• Director of the International Seizure Diary Consortium

• Founder of the American Academy of Neurology AI in Neurology Synapse Group

• Leader of the NeuroAI.Chats channel on VuMedi

You can get the idea. I like to learn, I like AI, and I like thinking about stopping epilepsy.

5. Feedback

I appreciate feedback. Let me know what is working and what is useful, and what is pointless. I am not hard to find: daniel.goldenholz@bidmc.harvard.ed

u