LLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes | Hacker News Reader

LLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes

arxiv.org

20 pointssbulaev14 comments

The fact that even the abstract is very clearly AI-generated does not fill me with confidence in the rigour of the research.

I understand the answer to the question I'm about to ask, but how does any human read that abstract and not think, "This is entirely too many colons."

EDIT: And Pangram agrees that the abstract is 100% AI generated.

While I agree I have to say that AI-detectors are complete bullshit

Max from Pangram has done a bunch of interviews recently. I'd give them a listen; he might change your mind.

I had to use an AI to undertand the abstract made by AI…

Interesting. My initial thought was “this is so horribly written it must be human”

"Verify Presence, Not Absence" is very LLM-coded all on its own.

The entire paper is almost certainly Claude-generated, given the clear Claude-speak throughout. It seems like they didn’t even try to have Claude write a polished paper: it reads like Claude writing up a lengthy report, down to the needless sectioning with idiosyncratic title language. There appears to be no disclosure, and the author contribution section appears to falsely claim that a particular author wrote the text.

I know arXiv has taken some measures to combat spam like this, but it seems like they’ll need to do more. There’s just very little barrier now to creating giant slop papers like this and then dumping them anywhere that won’t reject them. It is an insult to everyone’s time, and I can’t imagine they expect people to actually read this. If the expectation is that everyone will use an LLM to interpret it, then maybe they should have at least had a few more rounds of tightening and polishing the paper, even via LLM, to save the redundant token use.

Not sure about the paper but the results make sense, we see it in PDF extraction too. Fields that aren't in the document are being made up 11 to 40% of the time depending on the API.

The whole thing is nasty partly because it isn't just AI problem. When checking the human labels that we used in our evals 40 out of 142 answer keys claiming absence were wrong. Tricky one.

> Fields that aren't in the document are being made up 11 to 40% of the time depending on the API.

Do you have an example?

I'm in healthcare and ambient documentation is obviously a huge thing now but I don't have any experience with it. We have anywhere from 5-10 companies reach out a week trying to sell us on their product and the demos are mostly okay (though you can tell they're rely on the happy path through a lot of it), but we haven't actually pulled the trigger on anything. Thanks in advance.

Sure. Here's two good examples from the published outputs.

Receipts where there's no subtotal line, only a total. One model returns a subtotal anyway, "63.000", "88.000", numbers it copied or computed from elsewhere in the document. Ours invented a 2% discount on a receipt that has no discount line.

TV ad contracts where the station's address isn't on the contract. Several models fill it with a real address that is on the page, just the agency's or the advertiser's. Given your industry this pattern should worry you most, since these aren't made up from nothing. Just the model tripping and taking it from the wrong place, which might look plausible on the surface.

Here's what I'd suggest for your vendor demos. Run them on a handful of your own documents where you know a field is genuinely absent, and count how many come back filled. Their own samples won't tell you that.

All the raw outputs are here if you want to look: https://velrim.com/research/fabrication-on-absent-fields

Sure. Here's two good examples from the published outputs. Receipts where there's no subtotal line, only a total. One model returns a subtotal anyway, "63.000", "88.000", numbers it copied or computed from elsewhere in the document. Ours invented a 2% discount on a receipt that has no discount line.

All the raw outputs are public, so you can check the examples above yourself. My reply to you got filtered so I'm omitting the link. If you want to read the writeup, search for "velrim" and "fabrication on absent fields" for the article plus the repo.

May be AI written. Used my AI to get the gist of it. One thing that I believe is LLMs need to be considered a pure play tool and can be guided to point out misses as well. The reason is simple. To find whats missing, one needs to ask the right questions on how to judge and when that is there.

"the floored note before its clean twin"

People using AI like this should be run out of their jobs.

Reply on news.ycombinator.com