Think the title here is a little misleading. The citations do indeed appear to be hallucinations but it is not known if the report was written partially or entirely with an LLM. I would have no trouble believing that it was but at this point there isn't even a leak at MAHA claiming that it is.
Anthropic had a slip-up like this recently with a legal filing (totally unrelated to the report, just a similar example of bad citations). After being challenged by a judge, they said:
> A Latham & Watkins associate located that article as potential additional support for Ms. Chen’s testimony using a Google search. The article exists [...]
> [...] I asked Claude.ai to provide a properly formatted legal
citation for that source using the link to the correct article. Unfortunately, although providing the correct publication title, publication year, and link to the provided source, the returned citation included an inaccurate title and incorrect authors.
Anthropic could be lying, but apparently the link is indeed correct, so the account seems plausible.
However, the current situation is less understandable. The article says that "some correctly cited papers were inaccurately summarized", which suggests that AI either was used for the report itself, or at least was told to add citations without the author's input, which would be far more irresponsible than what Anthropic did. The apparently completely hallucinated "paper on direct-to-consumer advertising of prescription drugs" also doesn't look good.
The article also mentions that "[a]n early copy of the report shared with reporters did not include citations", which does support the theory that citations were added after the fact (whether or not AI was also used for the report itself).
I'm not saying they didn't. It's just that the title here says "WH releases health report written by LLM" but the article it links to does not claim that. The headline from the article is about the fake citations.
Also, I could imagine that the report, whether drafted by human or not, could have been pasted into an LLM with a prompt like "make this sound more authoritative" and the LLM dutifully added some "citations" because, what's more authoritative looking than citations?
