Synthetic doctors have been turning up on social media for well over a year, and people keep finding them. Rolling Stone reported in March 2025 on AI-generated doctor videos being used to sell unproven supplements. CBS News found deepfaked doctors on TikTok that August. In January 2026 the investigative outlet Indicator found dozens of synthetic-doctor accounts, the vast majority of them on Instagram, Facebook and Threads. The Polish fact-checking group Demagog traced AI doctor personas selling a supplement called shilajit; Science Feedback documented French-language YouTube channels impersonating medical professionals, sometimes using real doctors' names.

On TikTok specifically, the trail runs CBS News in August, Science Feedback in February, and the Guardian today — and at each of the three, the platform removed what it was shown. CBS reported that the videos it flagged were taken down; YouTube, by contrast, declined. Each time, roughly the same sequence: someone outside the platform finds the accounts, shows them to TikTok, and TikTok removes what it was shown. Science Feedback said in February that the accounts breaching its impersonation rule had been taken down; the Guardian says TikTok took down examples of the videos it was sent this month.

That sequence is the story — more than any single statistic in it.

The July research, and who published it

The Guardian reported today on research published this month by Hallam, a digital marketing agency, which analysed 1,198 videos — the top 50 across a series of search terms including "health tips" and "doctor advice", plus general "life-impacting" advice on personal finance and education. Hallam says AI-generated content appeared in 40% of top health-related TikTok videos, most of them posted by deepfake influencer accounts designed to resemble real doctors. For some specific search terms, such as "health tips", it puts the figure at 84% classed as AI-generated or AI-assisted — a broader category than AI-generated alone, and worth keeping distinct. It says the top AI-generated doctor and health videos in its search received an average of 2.5m views.

Those figures come from an interested party, and TikTok said so directly: "These findings from a marketing agency with commercial interests are not an accurate reflection of our platform."

That objection is worth taking seriously. An agency that sells digital marketing has an interest in the finding that social platforms are full of synthetic content. The numbers are not peer-reviewed, and we found no independent replication of them.

So set them aside for a moment, because the phenomenon does not rest on them.

February: the same thing, found by a non-profit

In February 2026, Science Feedback — a non-partisan non-profit that crowdsources reviews from scientists — ran its own investigation. It has no commercial stake in the answer, though it is a fact-checking organisation whose remit presupposes that misinformation matters; the relevant point is that it is not selling marketing services.

Its method was mundane, which is the point. Researchers searched TikTok for "doctor advice", followed the platform's own Suggested Accounts recommendations to find similar ones, and ran the videos through Hive Moderation software, which generally indicated a high likelihood that AI had been used to generate the speech. They identified 18 accounts in, by their own account, "just a few hours."

What they found was not generic slop. Accounts were impersonating real, named clinicians — the weight-loss doctor Garth Davis, and the orthopaedic surgeons Paul Zalzal and Brad Weening. Two of the most popular videos, from @healthvibes888, racked up more than 14 million views. Most of the accounts had several hundred thousand followers. One video showed a man in surgical scrubs claiming raw cacao and beetroot would "unclog your arteries"; another, in Weening's likeness, said turmeric reduces high blood sugar. Science Feedback's assessment is careful about that one: some studies have found curcumin, a compound in turmeric, could help reduce blood sugar — but turmeric cannot replace diabetes medication.

The accounts generally did not label their content as AI-generated, even though TikTok's rules require creators to do so.

Science Feedback shared the accounts with TikTok. TikTok said the accounts violating its rule against impersonation had been removed. Two days after publishing, Science Feedback amended its own headline following feedback from TikTok, broadening it from "on TikTok" to "on social media" — a detail worth stating, since this piece leans on its work.

The pattern

Five months later, the Guardian sent TikTok examples of the videos. TikTok took those down too.

Outsiders, none with special access to the platform, keep finding this material and reporting it. Each time TikTok acts — on the specific items it is handed. In between, and since, the material has evidently kept coming: the July research describes remedies including Coca-Cola mixed with onion for pain relief, a fake anti-ageing product called "Hyalethinap Plus Pro Max", and disproven cancer myths about microwaving food in plastic, using deodorant, and sleeping next to a phone.

Removal on notification is a real thing platforms do, and it is not nothing. But it is a different capability from prevention, and it scales with the number of people willing to go looking — not with the volume of material being posted.

It is worth being precise about what the platform found on its own, because the numbers are easy to misread — and the looseness starts in TikTok's own summary. Its Q3 2025 Community Guidelines Enforcement Report headlines a "highest-ever rate of proactive removal" of 99.3%. Read the underlying data and that figure covers removed short-form videos — 204.5 million of them — not "content" generally; it excludes the comments, fake accounts and LIVE suspensions removed in the same quarter. In the same dataset — in the raw figures rather than the dashboards TikTok itself renders — videos taken down under its Misinformation sub-policy were removed proactively 98.5% of the time.

Those are shares of what was removed, not of what exists. A very high proactive rate is entirely compatible with a tail that an outsider can find by searching.

The medical concern is broader than the 40% figure

The medical response does not stand or fall on whether 40% is the right number.

Dr Emma Runswick, deputy chair of the British Medical Association council, flagged the risks posed by AI accounts that "peddle medical myths and promote so-called miracle cures", and was blunt about where responsibility sits: "Platforms like TikTok must do more to stop the rise of dangerous fake medical advice … otherwise, it is the public who could be paying the price with their health and perhaps even their lives."

Prof Frankie Swords, national medical director at NHS England, described what clinicians are seeing rather than what a study counted: "Many clinicians are reporting an increasing number of patients who feel worried and confused due to false advice that has been made solely for clicks."

Alex Ruani, a health misinformation researcher at UCL, described a "frightening amount of exposure to fake medical content", and put the business model plainly: "We're looking at industrialised exploitation of trust."

William Pett, interim director of policy and public affairs at Healthwatch England, located the risk in the audience rather than the platform: "Without the skills to distinguish reliable information from commercial hype or misinformation, [people] risk being misled in ways that can seriously affect their health." (The bracket is the Guardian's.)

Healthwatch England has a number of its own, and it is the most useful figure in the story precisely because nobody in the AI industry produced it: a Savanta poll last year of 3,575 men and 3,832 women aged 18 and over in England found that about one in five people were using social media for health information. That is the exposed population. It was measured independently of this research, by a pollster, for a patient champion — and it does not move whether Hallam's 40% is right or wrong.

The research also follows warnings from the NHS that health misinformation on social media was a "real threat to public health".

The Royal College of GPs' president, Prof Victoria Tzortziou Brown, was reported as concerned by the research and as warning of a real risk of harm.

A doctors' union, the health service, a royal college, a university researcher and a patient champion, describing the same problem from five directions. Two of them — NHS England and the RCGP — were responding directly to Hallam's findings, and the Guardian says so. None is quoted commenting on its method.

What TikTok says it does

TikTok's full response is more substantive than the line about commercial interests, and it deserves to be read in full:

"These findings from a marketing agency with commercial interests are not an accurate reflection of our platform."

"The facts are that we remove harmful health misinformation, including when it's AI-generated, we partner with the World Health Organization and the NHS on providing reliable health information in-app, and we continually invest in AI literacy resources and labelling technologies."

All of that can be true alongside two investigations that found impersonation accounts by searching for them. A very high proactive removal rate is compatible with a long tail that persists, and a partnership with the WHO is compatible with synthetic doctors outperforming real ones in the feed. The figures measure different things, and neither side's numbers settle the other's.

The part that isn't really about TikTok

The accounts worked because they looked like doctors. That is the same mechanism behind voice-clone scams and the finding we covered in June that 85% of people can no longer reliably tell a scam from reality. Synthetic authority is cheap now, and medical authority is among the most valuable kinds to fake, because the people most likely to act on it are the people least able to check it.

Labelling rules exist. On the evidence of every investigation above, they were not being applied to the accounts that mattered most — and each time it took someone outside the platform to notice.