Post

@UlrikeHahn curious:

at the same time, hallucinated references disproportionately assign credit to already prominent and male scholars, suggesting that LLM-generated errors may reinforce existing inequities in scientific recognition.

i did notice how none of the outlets the authors used are peer-reviewed, and which therefore represent an uncontrolled adversarially-driven input. i also observed that a similar analysis could be performed for peer-reviewed journals, who can be held institutionally liable for failing to publicly retract fabrications they accepted for publication, to say nothing of the authors themselves.

i would be curious about whether fabricated citations at peer-reviewed journals also disproportionately assign credit to already prominent and male scholars, since a simple application of bayes' theorem would be sufficient basis to investigate illegally discriminatory review processes.

i did notice the characterization of "human review" as "costly" in the discussion section. i would put forward for future work an evaluation of the corresponding cost for the alternative, as the absence of human review would necessarily impose a legal liability for neglect.

you begin by paraphrasing this caption:

Teams cited in hallucinated citations also deviate from conventional authorship patterns, including smaller team sizes and weaker first-last author hierarchy. These biases largely extend to valid references co-cited with hallucinated references.

i will say i've simply never seen a p-value provided without the corresponding raw data before, so to see three in a row like that was a rare treat indeed! a sign of good luck.

now granted, i did think they'd perhaps come into disuse as a result of the p-value's catastrophic laundering of widespread scientific fraud. it's just too bad for the cancer patients who could only provide partial data. i wonder how often LLMs have been used to erase their death from the record.

perhaps Legal Liability might become another Model to reach for next time?

@UlrikeHahn So not only are LLMs theft, they poison the pool of human knowledge so that no one will ever be sure what's true again? Awesome.

@UlrikeHahn just generate fake articles for fake citations to make them real? thank me later

@UlrikeHahn "This is Fine"

@UlrikeHahn Wouldn't that be easy for the publishers, just to check that the references actually refer to something existing ? Do you think that could be automated ? Shouldn't that imply automatic desk rejection ?