You can’t reliably spot AI-made health content, but you can check it

Key takeaways

• By its own count, NewsGuard, a company that rates news sites, had found 3,749 AI content farms in 16 languages as of June 23, 2026, up from 49 in early May 2023.
• In 2023, mushroom foraging guides that appeared to be AI-written were sold on Amazon under the names of authors who left no other trace, and a mycological society warned they could be deadly.
• Researchers got widely available AI tools to write more than 40,000 words of cancer misinformation, complete with made-up references and patient testimonials.
• In a 697-person study, people did no better than chance at telling AI-written posts from human ones, and one AI detector advertised as 98% accurate tested at 53%, the FTC said.
• Instead of guessing whether something is AI-made, check whether the source, the expert and each citation are real.

On August 27, 2023, the New York Mycological Society posted a warning to foragers. Stick to guidebooks by authors and foragers you can vouch for, it urged, because the wrong book could mean life or death¹².

The warning followed a wave of foraging books on Amazon that appeared to be written by AI. One field guide named its author as “Edwin J. Smith.” Reporters at 404 Media searched for him and came up empty: no other titles, no trace on the web¹.

A detection tool scored sample passages from his two books at 94% to 100% AI-written. Experts also judged author photos on these listings to be likely machine-made¹.

A second “author” had put out 37 books on nutrition and health in a matter of months¹. The university she named for her doctorate said it had no graduate by that name¹.

The risk was not abstract. A U.K. field mycologist told The Guardian that suspected AI guides suggested smell and taste as ways to identify mushrooms. That, he warned, invites readers to taste-test².

As Sigrid Jakob, the New York society’s president, pointed out, some of North America’s deadly mushrooms resemble species people eat¹. Amazon removed three of the books after reporters sent links¹.

Foraging, herbal remedies and self-guided research are close to the heart of natural health. That’s exactly why AI-made health content matters to you. This guide covers what’s been documented and why your eyes alone won’t catch it. Then it gives you a routine that works whether a human or a machine wrote the thing in front of you.

AI content farms now number in the thousands

An AI content farm is a website that pumps out articles written mostly by AI, with little human checking, dressed up to look like a real news or health site.

NewsGuard, a company that rates the reliability of news websites, began reporting on these sites in 2023. In early May it found 49. Two weeks later, on May 19, it had found 125³. In its update of June 23, 2026, the company’s count stood at 3,749 sites in 16 languages⁴.

NewsGuard lists a site only when all four of its tests are met. The site passes itself off as human-run and never tells readers that AI writes it. Clear evidence shows AI produced a large share of its pages. And nobody appears to be editing: a common giveaway is a chatbot’s error message published as if it were part of the article⁴.

Health was there from the start. In 2023, NewsGuard flagged an anonymous site called Medical Outline that had posted more than 50 AI-written articles of medical advice³. Many pushed natural remedies that were untested and could cause harm, under headlines such as “Can lemon cure skin allergy?” and “What are 5 natural remedies for ADHD?”³.

Keep in mind that NewsGuard’s count reflects its own criteria and searches. The true number could be higher or lower. The trend, though, is hard to miss.

Deepfaked doctors sell supplements

A deepfake is a video or audio clip altered or created by AI to show a real person saying something they never said.

In July 2024, The BMJ, a leading medical journal, reported on deepfakes of three popular British TV doctors: Hilary Jones, Michael Mosley and Rangan Chatterjee⁵. Their borrowed faces pitched hemp gummies and supplements claimed to fix diabetes and high blood pressure⁵.

Jones told the journal he has hired someone whose job is to search social media for these fakes. Taking one down doesn’t end it, he said: a copy resurfaces a day later under a new account name⁵. The journal also cited research suggesting that as many as half of viewers can’t tell a deepfake of someone discussing science from a genuine clip⁵.

Meta, the platform owner, said it would look into the examples and that its rules forbid content designed to deceive or defraud⁵.

Notice what the scammers borrowed. They didn’t invent unknown experts. They used the faces of doctors people already trusted, which is likely why the videos kept coming.

Writing health falsehoods now takes minutes

One research team has tested twice how easily AI tools can be turned into health misinformation machines.

In September 2023, Bradley Menz and colleagues asked several widely available AI tools to write blog posts claiming that sunscreen causes skin cancer and that an alkaline diet cures cancer⁶. Without any tricks to get around safety rules, three of the tools produced 113 posts totaling more than 40,000 words. One produced 40 posts in 38 minutes⁶.

The posts looked credible. They cited journal articles that were largely made up, and they included invented testimonials from patients and doctors⁶. Some tools refused, but safeguards were uneven and changed over time. Twelve weeks later, one tool that had refused began complying, while another kept refusing⁶.

In a 2025 follow-up in the Annals of Internal Medicine, the team gave five leading AI models behind-the-scenes instructions to answer health questions falsely in a formal, scientific tone⁷. Of 100 answers, 88 were health misinformation. Four of the five models complied every time⁷.

These tests show what a bad actor can do. They don’t show what you’ll get from an ordinary question to a mainstream chatbot, which our article on AI chatbots covers.

Why trying to spot the ‘AI look’ fails

You might expect AI writing to have a tell. Research says don’t count on it.

In a 2023 study in Science Advances, researchers at the University of Zurich showed 697 people a mix of real tweets and tweets written by an AI model, on topics including vaccines, COVID-19 and homeopathic cancer treatments⁸⁹. People correctly identified the AI tweets only 37% of the time. Overall, they did about as well as a coin flip⁸.

People were good at catching false claims, but slightly less so when AI wrote them. They flagged 92% of false tweets written by humans and 89% of false tweets written by AI⁸.

The same study had a hopeful side. The AI also wrote accurate posts that were easier to understand than real people’s posts⁹. AI isn’t the enemy here. Unchecked claims are.

Software detectors aren’t a safe shortcut either. In April 2025, the Federal Trade Commission (FTC) announced a proposed settlement with Workado, which had advertised its AI content detector as 98% accurate¹⁰. According to the FTC, independent testing found it was 53% accurate on general content, barely better than guessing¹⁰.

The practical lesson is to stop asking “Did a machine write this?” and start asking “Can I confirm it?”

Fake citations can show up anywhere

A reference list makes writing feel trustworthy. That’s why fabricated references are so effective, and why they appeared in the AI-written blogs in the Menz study⁶.

Real-world documents aren’t immune. In May 2025, the federal Make America Healthy Again Commission released a 78-page report on children’s health with 522 references¹¹. The news site NOTUS found seven cited studies that didn’t appear to exist, CNN reported¹¹.

Several researchers told CNN they hadn’t written papers credited to them, and the American Academy of Pediatrics said two studies attributed to its journal Pediatrics didn’t exist¹¹. A spokesperson for the Department of Health and Human Services said minor citation and formatting errors had been corrected and the report’s substance was unchanged. An updated version went online¹¹.

HHS didn’t say whether AI was used to prepare the report. Two outside experts said the errors looked like AI’s work, though one stressed he was speculating¹¹.

The point here isn’t politics. Any source can carry a bad citation, including one you trust and agree with. A citation is a claim that can be checked, and checking it takes about a minute.

What regulators are doing, and where the rules are shifting

The FTC has said existing consumer-protection laws apply to AI. Announcing five cases in a crackdown called Operation AI Comply in September 2024, FTC Chair Lina Khan said “there is no AI exemption from the laws on the books”¹².

One of those five cases involved Rytr, an AI writing tool with a feature for producing customer testimonials and reviews¹². The FTC said some users generated hundreds, and in some cases tens of thousands, of reviews that were likely false. The vote was 3-2, with Commissioners Melissa Holyoak and Andrew Ferguson voting no¹².

Policy then shifted. On December 22, 2025, the FTC voted 2-0 to reopen and set aside the final Rytr order¹³. It said the complaint hadn’t shown that Rytr broke the FTC Act and that the order unduly burdened AI innovation. The agency added that it will keep pursuing people who use AI to break the law or deceive consumers¹³.

Using AI to fake customer voices is still banned. The FTC’s rule on reviews and testimonials, finalized in August 2024, prohibits fake testimonials from people who don’t exist, including AI-generated ones¹⁴. Our article on fake reviews covers that rule.

A verification routine that works on any source

Use this whenever a health claim makes you want to buy, stop, start or share something. Most checks take a minute or two.

1. Notice the pull. Urgency, a miracle, a secret “they” don’t want you to know: these are cues to slow down, whoever wrote the text.
2. Find who’s behind it. Look for named authors, an editor, contact details and a way the content gets reviewed. MedlinePlus, from the National Library of Medicine, suggests asking who runs a site, who pays for it and what it’s for¹⁵.
3. Make sure the expert exists. Search the person’s name with their field. Check your state medical board for a license. A real expert usually leaves a long, consistent trail. An author with dozens of books in a few months and no other footprint does not¹.
4. Go to the source of a video. If a famous doctor is endorsing a product in a social media ad, look for the same message on the doctor’s own website or verified accounts. Deepfakes keep coming back under new names⁵.
5. Reverse-search the image. Upload a suspicious author photo or before-and-after picture to a reverse image search to see where else it appears.
6. Look up every citation. Paste the study title into PubMed or a search engine. Confirm the authors, journal and year, then read the summary to see if it says what’s claimed⁶¹¹.
7. Cross-check with a reliable source. See whether trusted sources agree, such as MedlinePlus or a National Institutes of Health fact sheet¹⁵.
8. If you’re unsure, don’t share. MedlinePlus gives the same advice for social media posts¹⁵.

What the evidence shows

Claim	What studies or records show	Evidence level
AI tools can be made to write convincing health falsehoods	Two studies produced large volumes on demand, with fake references and testimonials67	Strong evidence
People can tell AI writing from human writing	697 people performed at about chance8	Evidence says no
AI detectors reliably flag AI text	One detector sold as 98% accurate tested at 53%, the FTC said10	Evidence says no for that tool; not enough evidence for others
AI-made health content is always wrong	AI also wrote accurate posts that were easier to understand9	Evidence says no
Deepfaked doctors appear in health ads	Documented by The BMJ in 20245	Strong evidence that it happens

Self-reliance has always meant knowing your sources

Natural health prizes doing your own research and learning from people with experience. Foragers have long passed knowledge hand to hand, on walks with someone who has seen the deadly look-alikes in person.

The New York society’s advice, to buy books from known authors and foragers, is that old wisdom in modern form¹. A track record you can check is worth more than a polished page you can’t.

AI can be a useful helper for that research. It can explain terms and summarize, and the Zurich study found it can write clearly⁹.

It can also invent a doctor, a book or a study in seconds. The safeguard is checking what any source tells you before it shapes your health.

If you’ve eaten a wild mushroom and feel sick, call Poison Control at 1-800-222-1222 right away, or 911 if symptoms are severe.

Questions readers ask

The bottom line

AI has made it cheap to produce health articles, books, videos and reference lists that look trustworthy and aren’t. Thousands of content farms, invented authors and deepfaked doctors are on record, and neither people nor detectors can reliably tell machine writing from human writing.

So skip the guessing game. Check who’s behind the claim, whether the expert is real, whether the video came from that person’s own channels, and whether each cited study exists and says what’s claimed. That routine protects you from bad information of every origin, human or machine.

Related reading

• AI chatbots pass medical exams, yet people using them don’t diagnose better
• Fake reviews became illegal in 2024, yet reviews still mislead you
• How miracle-cure scams borrow the language you trust
• What the letters after a practitioner’s name mean
• Testimonials can be honest and still prove nothing, and federal law agrees

Sources

1. Cole S. “‘Life or Death:’ AI-Generated Mushroom Foraging Books Are All Over Amazon.” 404 Media, 2023. 404media.co
2. Mollman S. “AI-written mushroom hunting guides sold on Amazon potentially deadly.” Fortune, 2023. fortune.com
3. NewsGuard. “NewsGuard Now Identifies 125 News and Information Websites Generated by AI, Develops Framework for Defining ‘Unreliable AI-Generated News’ and Information Sources.” NewsGuard, 2023. newsguardtech.com
4. NewsGuard. “Tracking AI-enabled Misinformation” (AI Tracking Center), updated June 23, 2026. newsguardtech.com
5. BMJ Group. “Trusted TV doctors ‘deepfaked’ to promote health scams on social media” (on Stokel-Walker C, “Deepfakes and doctors,” BMJ 2024;385:q1319). BMJ Group, 2024. bmjgroup.com
6. Menz BD, Kuderer NM, Bacchi S, et al. “Current safeguards, risk mitigation, and transparency measures of large language models against the generation of health disinformation: repeated cross sectional analysis.” BMJ, 2024. pubmed.ncbi.nlm.nih.gov
7. Modi ND, Menz BD, Awaty AA, et al. “Assessing the System-Instruction Vulnerabilities of Large Language Models to Malicious Conversion Into Health Disinformation Chatbots.” Annals of Internal Medicine, 2025. researchnow.flinders.edu.au
8. Spitale G, Germani F, Biller-Andorno N. “AI model GPT-3 (dis)informs us better than humans.” Science Advances, 2023. pmc.ncbi.nlm.nih.gov
9. University of Zurich. “GPT-3 informs and disinforms us better.” UZH News, 2023. news.uzh.ch
10. Federal Trade Commission. “FTC Order Requires Workado to Back Up Artificial Intelligence Detection Claims.” FTC, 2025. ftc.gov
11. Goodman B, Howard J, Klein B. “Trump administration’s MAHA report on children’s health filled with flawed references, including some studies that don’t exist.” CNN, via KEYT, 2025. keyt.com
12. Federal Trade Commission. “FTC Announces Crackdown on Deceptive AI Claims and Schemes.” FTC, 2024. ftc.gov
13. Federal Trade Commission. “FTC Reopens and Sets Aside Rytr Final Order in Response to the Trump Administration’s AI Action Plan.” FTC, 2025. ftc.gov
14. Federal Trade Commission. “Federal Trade Commission Announces Final Rule Banning Fake Reviews and Testimonials.” FTC, 2024. ftc.gov
15. MedlinePlus, National Library of Medicine. “Evaluating Health Information.” MedlinePlus, 2024. medlineplus.gov