Case study

The MAHA Report: Studies That Don’t Appear to Exist in a White House Health Report

In May 2025, a White House commission published a major report on children’s health. Some of the studies it listed as sources don’t appear to exist. Whether AI was used has never been confirmed.

What are the labels? The Evidentiality Framework asks an AI to mark each claim it writes: (g) generated, its own work; (u) given, passed to it by someone else; or (m) checked against a named source. The marks stay on a claim while people work with it. How the labels work.

What Happened

In February 2025, a presidential order set up the Make America Healthy Again (MAHA) Commission, led by the Health Secretary, and gave it 100 days to report.
The report came out on 22 May 2025. Who wrote it, and with what tools, has not been made public.
On 29 May, the news site NOTUS reported that seven of the studies it cited did not appear to exist. One was listed as a JAMA Pediatrics paper, “Changes in mental health and substance abuse among US adolescents during the COVID-19 pandemic”. The researcher named as its author said it was “not a real paper that I or my colleagues were involved with”.
The Washington Post reported that some source links contained “oaicite”, a marker that has appeared in ChatGPT output. That points to AI use but doesn’t prove it.
The White House press secretary called them “formatting issues” and said they did not “negate the substance of the report”. A corrected version replaced the citations. Asked whether AI was used, she referred questions to the health department.

Follow the Claim

This is an illustration of how the labels would have worked, not a test. It only holds if the conditions under “What Would Have Had to Be True” held.

Follow one source: from a draft to a White House report

Key: as it happened, the type gets bigger as the claim sounds more certain. With labels: red (g) generated: written by the AI; green (u) given: passed on, with who said it; blue (m) checked against a named source.

As it happenedWith labels

Before 22 May 2025 · AI tool (suspected) → report writers

As it happened“Changes in mental health and substance abuse among US adolescents during the COVID-19 pandemic”, JAMA PediatricsA source that looks real, with a real scientist’s name on it.

With labels(g)Changes in mental health and substance abuse among US adolescents during the COVID-19 pandemic, JAMA Pediatrics(/g)If an AI tool wrote it and used the labels: marked as the AI’s.

Before 22 May 2025 · Draft → the list of sources

As it happenedListed as a source in the draft reportNow it looks like research someone read.

With labels(u)Changes in mental health and substance abuse among US adolescents during the COVID-19 pandemic, JAMA Pediatrics(/u: AI tool, unconfirmed)Passed on with where it came from, and that nobody has opened it.

Before 22 May 2025 · The check

As it happenedNobody looked it upThe source was never opened.

With labels(m)No such paper by the named author(/m: NOTUS, 29 May 2025, checked 3 Oct 2026)This check could be made at the time. We name the source that confirmed it later.

22 May 2025 · Report → the public

As it happenedPublished as a source in a federal reportNow it’s evidence in national health policy.

With labelsLeft out. The check found nothing, so the claim never reaches the finished document.Finished documents carry no labels. They only carry claims that passed the check.

The study does not appear to exist. Who wrote the report and which tools they used has not been made public; the first row shows how such a source would look if an AI tool produced it.

How the Labels Could Have Helped

Each source shows where it came from. A reference a tool suggested arrives marked as the tool’s, not as research someone read.
Publishing waits for a check. A source nobody has opened doesn’t go out under a government seal.
Readers can trust the rest. When one source fails, readers can see which others were checked.

What Would Have Had to Be True

The four conditions every case shares: the AI tool used the labels; it labelled its own work correctly (the weakest link: in our tests, AI sometimes mislabels its own work); the label stayed on when the text was copied; and someone owned a rule that unchecked claims don’t go further. In this case:

The rule is owned by the commission staff who clear the report for publication. And an AI tool has to have been used at all, which hasn’t been confirmed.

What Already Existed

Review before publishing. Federal agencies have review and clearance steps for scientific reports. How this report was reviewed hasn’t been made public.
A simpler check would have caught it: looking each source up, for example by its DOI, the code that points to one exact paper. Labels add one thing: they show, line by line, which sources someone has actually opened.

What the Labels Wouldn’t Have Caught

The deadline. The commission had 100 days. Labels don’t make time.
No named authors. Labels say where a claim came from, not who signed off on the report.
Real studies, wrong claims. A label says a source was checked, not that it supports the point being made.