Post

Nina Jankowicz
@ninajankowicz.com

"We find a sharp rise in non-existent references following widespread LLM adoption, with a conservative estimate of 146,932 hallucinated citations in 2025 alone....hallucinated references disproportionately assign credit to already prominent and male scholars." arxiv.org/abs/2605.07723

arxiv.orgLLM hallucinations in the wild: Large-scale evidence from non-existent citationsLarge language models (LLMs) are known to generate plausible but false information across a wide range of contexts, yet the real-world magnitude and consequences of this hallucination problem remain p...

35 reposts76 likes

Monde Boeuf
@mondeboeuf.bsky.social

The term "Hallucination" for AI errors is a marketing word to mask blame. LLMs are software programs, so the right word for fails is "Bugs". When an LLM returns bad data, it's because the code is buggy. If Excel failed to calc formulas it would be laughed off the market. Why is AI given a huge pass?

Julia
@foxgrrl.northsky.social

It's not a "bug", the LLM is always doing the exact same thing every time. The LLM doesn't know anything. The LLM doesn't know what's true or false. The LLM doesn't know what's real or fake. The LLM outputs an ordered list of words based on the probability of those words occuring together in English

Monde Boeuf
@mondeboeuf.bsky.social

It's no f'ing excuse for shipping without error control. The output of AI could be checked for validity before outputting to the user. But that would cost extra and people like you are only too happy to give the AI companies a free pass to sell their misinformation firehoses. The damage is immense.

Julia
@foxgrrl.northsky.social

What I'm saying is that an LLM text generator IS ALWAYS HALLUCINATING by design! That's not a "bug", the program is working as designed. Everything that it outputs has nothing to do with truth, correctness, or reality! It is always always always hallucinating everything all the time!

Julia
@foxgrrl.northsky.social

I'm not giving it a free pass. I'm telling you that the only thing an LLM does is output statistically likely combinations of words based on words given to it as input. The actual words mean nothing to it! It is a very large Markov Chain. IT DOESN'T KNOW ANYTHING EXCEPT ENGLISH WORD FREQUENCY!

Julia
@foxgrrl.northsky.social

It's a language model. It's not a model of anything other than language! It is not a model of anything other than language, for example, physical reality. It can't do math! It is not a model of a brain, body, consciousness, society, etc. It doesn't even contain a database of bibliographic citations!

Dedekind Slut
@chasmat.bsky.social

Might check out state court filings next. The high profile cases have been in federal court and it’s definitely not that federal court practitioners are more likely to use AI (it’s that federal judges check all your cites)