HD-OPM Archive

Using AI and deep-research tools for health questions

How the group's admins use chatbots such as Gemini, ChatGPT and Claude to fact-check and summarize health claims, why they insist on verifying every reference, and where AI is (and is not) being used in cardiology itself.

10 min read25 group posts cited9 studies discussedEvidence through Oct 2026AI Summary of posts by Louis Nervegna and Gokhan Can
Key takeaways
  • Louis calls AI tools useful for researching health topics but warns they can hallucinate citations or misquote references, so the group asks members to look up the actual references before relying or posting.
  • Generative AI is "designed to sound extremely convincing, even when it is wrong" in Louis's words, so the group's guidance is to be extremely careful about using it for medical decisions.
  • Louis recommends free tools such as Gemini as a quick fact-check of claims heard online, including a worked prompt: "Please critique and fact-check this text for scientific accuracy."
  • AI fact-checks are a starting point, not a verdict: in the group's examples, Gemini corrected overstatements in one post, while in others it opened by praising the text it was asked to check.
  • Research AI systems are reported to match or beat physicians in diagnostic studies (303 real cases for Google's AMIE, 304 NEJM cases for Microsoft's MAI-DxO), but the posts stress this is research, not a substitute for your own cardiologist.
In this summary
  1. What AI is useful for
  2. The pitfalls the admins flag
  3. AI inside cardiology and medicine
  4. What this means for you
  5. Frequently asked questions
  6. The evidence behind this summary

Can you use an AI chatbot to research a heart-health question? The group's admins say yes, with conditions. Louis Nervegna describes generative AI tools such as ChatGPT as very useful for researching health topics, but adds right away that they "often make mistakes", especially when citing references1. His practical rule is simple: treat AI as a fast first draft or a second pair of eyes, and check the underlying sources yourself.

In the group, AI has become a routine tool for three jobs: fact-checking claims made by online influencers, summarizing a body of evidence quickly, and checking the group's own posts for errors. It is also a news topic, with several shared items about AI systems that diagnose or analyze images in cardiology. This article covers how the admins use these tools, the pitfalls they flag, and practical habits for using them responsibly. It is based on the group's posts and does not endorse any specific product.

303challenging real-world cases on which Google's research AI, AMIE, outperformed clinicians at differential diagnosis2
80% vs 47%diagnostic accuracy of Microsoft's MAI-DxO versus physicians on typical ER cases, in initial results3
Almost 1 in 4patients without heart disease met lipid-lowering criteria in the CLAiR retinal AI trial4

What AI is useful for

Fact-checking claims you hear online

The most common use in the group is fact-checking. Louis says you need to fact-check everything you hear about health online, because "most of it is nonsense", and recommends doing it with AI8. In another post he encourages people to use tools like Google Gemini to fact-check medical information they hear online, noting it is easy and free, and that the major AI tools are "getting tough to fool now" on cholesterol and statins9. He also notes that you can fact-check a claim such as the one comparing the insulin response to beef versus oatmeal or pasta by asking an AI directly10.

A worked example from the group shows how this looks in practice. Louis posted a claim about a statin trial and TG/HDL ratio, asked Gemini whether it was true, and the response called the statement scientifically inaccurate as a categorical rule: it said the ratio is a marker of insulin resistance and small dense LDL, but there is no landmark double-blind, placebo-controlled trial showing a "zero benefit" threshold at a ratio of 111. Louis presents this as a way to test a claim rather than as proof.

Summarizing the "totality of the evidence"

Louis argues that influencers cherry-pick single studies, and that a good scientific check considers the totality of the evidence. He once asked Claude to summarize the evidence on whether Paleolithic humans were carnivores; the reply described coprolites with plant remains, grinding stones, dental adaptations and isotope data, and concluded they were omnivores12. He describes this as a good way to do a basic scientific fact check. The same post is also a reminder of the limits: the summary named no specific papers, so it can only serve as an overview to be followed up12.

Group members have also shared AI-generated condensed summaries of other material, for example a ChatGPT summary of Ancel and Margaret Keys' 1975 Mediterranean diet book that Pankaj Kapahi generated and Louis passed on13. Louis has also shared a Gemini summary of a trial result alongside a post on resistance training14.

Checking your own writing

Louis also uses AI to check posts before and after sharing them. In one case he posted a text on ceramide and insulin resistance, then replaced his takeaway with a revised version from Gemini that "corrects some technical inaccuracies and overstatements" in the original, with the full fact-check placed in the comments15. In another case, he used Gemini to test his own explanation of how short trials understate the long-term benefit of statins, and encourages everyone to do likewise16. On a dairy and dementia post, he again put his own summary through Gemini and urged others to do the same to help ensure factual accuracy17. He also uses a standard prompt, shown in one post: "Please critique and fact-check this text for scientific accuracy. Do not include tables in your response."6.

Several of Louis's posts are labeled as written "in collaboration with Gemini AI", a disclosure habit that lets readers know which text was AI-assisted, for example on the genetics of statins and diabetes, on PCSK9 inhibitors versus statins, on hsCRP, on a plaque-reversal trial summary and on cocoa butter1819202122.

The pitfalls the admins flag

Hallucinated and misquoted references

Louis's clearest rule, posted as a group rule, is about references. Generative AI tools "can hallucinate citations that do not exist, or hallucinate information or quotes from references that are wrong." Anyone posting AI-generated content in the group is asked to look up the actual references the AI cites and to check what the AI says about them1. In plain terms, a "hallucination" means the model produces plausible-sounding but false content, such as a paper that does not exist.

Persuasive tone is not reliability

Louis shared a short video by Terence Tao on what he calls AI's plausibility problem, with the warning that generative AI tools "are designed to sound extremely convincing, even when they are wrong," and that you need to be extremely careful using them for medical decisions5. Fluent writing, in other words, says nothing about whether a claim is true.

AI can agree too readily

An observation from the group's own examples: in two of the shared fact-checks the AI's reply opened by endorsing the text it was shown. One begins "This is a great summary of the recent Swedish study" and another says the text "is scientifically accurate and aligns closely with the current consensus"617. The first of these appeared on an observational study in which Louis had already stressed that observational studies show association, not causation17. This does not prove the AI was wrong, but it is a reason to ask explicitly for a critique, as the prompt Louis uses does, rather than "is this right?"6. The Gemini fact-check on the ceramide post shows what a useful critique looks like, since it found specific overstatements15.

AI output is not itself evidence

When Louis explains that an elevated creatine kinase on a statin points to a real drug effect rather than a nocebo effect, he supplies his own reasoning and rough calculations, then adds a Gemini answer as supporting text, and advises speaking to a doctor23. The same approach is a good model: AI text sits beside the primary reasoning and the data, and does not replace them.

Don't swap AI for your cardiologist

In a 2024 post, Louis says that if you are at high risk of a heart attack, rule number one is to find a good cardiologist and listen to them, and not to gamble on social media guesses or single cherry-picked studies. If you want to try natural or alternative approaches, it is safest to add them to what your cardiologist suggests and not to replace it7. The same logic applies to chatbots: they help you prepare questions, not make decisions on your behalf57.

The three everyday uses, and the check the admins pair with each:

UseHow the admins do itThe check
Fact-checking a claimPaste the claim into a free tool such as Gemini and ask for a critique69Look up the references it cites and confirm what they say1
Summarizing the evidenceAsk for the totality of evidence, as with Claude on Paleolithic diets12The summary named no papers, so follow up with primary sources12
Checking your own writingRun a post through Gemini, as Louis did with his ceramide text15Ask for criticism, since replies can open by praising the text617
Posting AI-assisted textLabel it "written in collaboration with Gemini AI"1819Readers know which text was AI-assisted1819

AI inside cardiology and medicine

The group also follows news about AI used by clinicians. These items are shared as news, and some are explicitly not heart-specific.

  • Google AMIE. Louis shared a note on Google's research AI for diagnostic reasoning. In one paper it matched or exceeded primary care physicians in simulated text-chat consultations with patient actors. In a second, on 303 challenging real-world medical cases, it outperformed clinicians when generating a differential diagnosis, and clinicians assisted by it improved. The shared text also claims clinicians did worse when assisted than if they let the AI work alone2.
  • Microsoft MAI-DxO. A news item reports a "virtual panel" of AI agents that more than doubled physicians' accuracy (as measured with reference tools) on 304 complex New England Journal of Medicine cases and reached 80% versus 47% on typical emergency room cases, while ordering fewer, cheaper tests. The item itself cautions that getting a diagnosis right is only one step in care and that real clinical testing is still to come, planned at Beth Israel Deaconess in Boston3.
  • Retinal imaging. In the CLAiR randomized trial of an AI algorithm analyzing retinal images, almost a quarter of patients without heart disease met guideline criteria for lipid-lowering therapy, based on a 7.5% 10-year ASCVD risk. Louis notes the AI tools are still in the research stage and imagines tracking disease in the future4.
  • AI plaque analysis. Louis shared new American College of Cardiology guidance on AI-based quantitative coronary plaque analysis on CT angiography (tools such as Cleerly and HeartFlow). Unlike a calcium score, which only measures calcified plaque, the analysis characterizes all plaque types and can help follow treatment response24.
  • Drug development. Louis comments that AI is speeding up early (preclinical) drug development at a fraction of the cost, even with free or low-cost tools. No evidence is cited in the post25.
The chart Louis shared from Google's AMIE study: diagnostic accuracy across the top 1 to 10 suggested diagnoses for AMIE alone, clinicians assisted by AMIE, clinicians assisted by search, and clinicians unassisted. AMIE alone scored highest and unassisted clinicians lowest.
The chart Louis shared from Google's AMIE study: diagnostic accuracy across the top 1 to 10 suggested diagnoses for AMIE alone, clinicians assisted by AMIE, clinicians assisted by search, and clinicians unassisted. AMIE alone scored highest and unassisted clinicians lowest.2
The graphic on the ACC statement on AI plaque analysis: two patients with the same calcium score of 0 can have very different plaque, because a calcium score sees only calcified plaque while AI analysis of CT angiography sees all plaque types and can track treatment over time.
The graphic on the ACC statement on AI plaque analysis: two patients with the same calcium score of 0 can have very different plaque, because a calcium score sees only calcified plaque while AI analysis of CT angiography sees all plaque types and can track treatment over time.24

These are reports of research and guidance, not proof that a consumer chatbot can diagnose you. The MAI-DxO coverage itself says this still has to be proven at scale3.

What this means for you

The posts are opinions of group admins, not medical advice. Taken together, they suggest these habits:

  • Use AI to generate a first summary or to challenge a claim you heard, such as an influencer's statement8912.
  • Look up every reference the AI gives you and check that the quote or number is really in it, since citations can be invented or misquoted1.
  • Ask the AI for a critique, not a confirmation. Louis's prompt asks it to critique and fact-check a text for scientific accuracy6.
  • Expect that AI will sometimes be wrong and sound sure. Louis says to be extremely careful before using it for medical decisions5.
  • Put the primary sources first: trial names, journals and numbers should come from the paper itself, with AI used to help you read and organize them112.
  • If you post AI-assisted text, say so, as Louis does with "written in collaboration with Gemini AI"1819.
  • If you have heart disease or are at high risk, the group's admins recommend following a good cardiologist and treating AI answers as questions to bring to your appointment, not instructions7.

Frequently asked questions

Is it safe to rely on a chatbot for a health decision?

The group's guidance is no. Louis says generative AI tools are designed to sound extremely convincing even when wrong, and urges great care for medical decisions5. For high-risk people, he says to listen to a good cardiologist7.

What does it mean when an AI "hallucinates" a reference?

It means the tool cites a paper that does not exist, or attributes to a real paper a quote or finding that is wrong. Louis asks members to look up the actual references before relying on, or posting, AI content1.

How do I fact-check an online health claim with AI?

Louis suggests pasting the claim or text into a tool such as Gemini and asking for a critique, using a prompt like "Please critique and fact-check this text for scientific accuracy"69. Then check the sources it names yourself1.

Can AI really diagnose better than doctors?

Reports shared in the group say research systems such as AMIE and MAI-DxO matched or beat physicians in specific studies of simulated chats and published cases23. These are research results, and the Microsoft coverage notes that clinical testing and wider proof are still ahead3.

Does AI have a role in tracking heart disease?

Possibly. The ACC guidance on AI plaque analysis on CT angiography and the CLAiR retinal imaging trial point toward AI-assisted risk estimation and treatment tracking, with the retinal approach still at the research stage424.

The evidence behind this summary9 studies · 3 trials · 1 MR · 1 observational

The 25 group posts cited above discuss 9 studies and sources. Their mix of study designs shows how much rests on randomized trials and meta-analyses, which can show cause and effect, and how much on observational, lab or animal work, which mostly cannot.

Randomized trials 3Mendelian randomization 1Observational studies 1Reviews & guidelines 1Other sources 3

Strongest evidence discussed

All 9 studies and sources
Sources from the group (25)
  1. Fact-check AI-generated content before posting it in the group Louis Nervegna · Jul 28, 2025 · Facebook
  2. Google's AMIE diagnostic AI matched or beat physicians in two studies Louis Nervegna · Apr 11, 2025 · Facebook
  3. Microsoft's MAI-DxO AI beat physicians on complex NEJM diagnostic cases Louis Nervegna · Oct 23, 2026 · Facebook
  4. Retinal AI imaging may predict CVD risk; in CLAiR, ~1 in 4 met lipid-lowering criteria Louis Nervegna · Apr 10, 2026 · Facebook
  5. Generative AI sounds convincing even when wrong; be careful with medical decisions Louis Nervegna · Apr 05, 2025 · Facebook
  6. No long-lived population ate a carbohydrate-free diet; long-term effects unknown Louis Nervegna · Jun 09, 2026 · Facebook
  7. At high risk of a heart attack? Find a good cardiologist and add, don't replace, treatment Louis Nervegna · Feb 26, 2024 · Facebook
  8. Fact-check health claims with AI: a viral video on oatmeal vs eggs misleads Louis Nervegna · Mar 16, 2026 · Facebook
  9. AI chatbots now fact-check cholesterol denial claims accurately, says Louis Louis Nervegna · Jul 19, 2026 · Facebook
  10. Beef raises insulin more than oatmeal or pasta per calorie, with an AI fact-check Louis Nervegna · Sep 22, 2026 · Facebook
  11. Fact check: TG/HDL under 1 does not nullify statin or LDL/ApoB benefit Louis Nervegna · Apr 10, 2026 · Facebook
  12. Claude AI summary: Paleolithic humans were omnivores, not carnivores Louis Nervegna · May 14, 2024 · Facebook
  13. Ancel Keys' 1975 Mediterranean diet book: plants, olive oil, fish, little red meat Louis Nervegna · Jan 19, 2025 · Facebook
  14. Twice-weekly resistance training reversed prediabetes in a third of adults in 12 weeks Louis Nervegna · Jun 11, 2026 · Facebook
  15. Ceramide as a driver of insulin resistance, with a Gemini AI fact-check Louis Nervegna · Sep 01, 2026 · Facebook
  16. Number needed to treat explained: short prevention trials understate long-term benefit Louis Nervegna · Oct 12, 2025 · Facebook
  17. Full-fat cheese and cream linked to lower dementia risk over 25 years (observational) Louis Nervegna · Dec 18, 2025 · Facebook
  18. Genetics suggest statins' small diabetes risk is tied to LDL lowering (Gemini-assisted) Louis Nervegna · Sep 07, 2026 · Facebook
  19. No trial has shown PCSK9 inhibitors lower heart attack risk without statins Louis Nervegna · Aug 13, 2026 · Facebook
  20. hsCRP is not artery-specific: aging skin and visceral fat drive it in older people Louis Nervegna · Jun 28, 2026 · Facebook
  21. FLOWPROMOTE: LDL near 40 mg/dL on statin plus ezetimibe cut total plaque 21% in 18 months Louis Nervegna · Oct 31, 2026 · Facebook
  22. Cocoa butter raises LDL far less than milk fat, so avoid chocolate with added milk fat Louis Nervegna · Aug 08, 2026 · Facebook
  23. Elevated creatine kinase on a statin means muscle symptoms are real, not nocebo Louis Nervegna · Apr 29, 2026 · Facebook
  24. ACC guidance backs AI plaque analysis on CT angiography to guide and track treatment Louis Nervegna · Jun 01, 2026 · Facebook
  25. AI is speeding up preclinical drug development at a fraction of the cost Louis Nervegna · Sep 14, 2026 · Facebook

Keep exploring

Related research summaries

Topics and hashtags in this summary