there is no such thing as "responsible use of generative ai" in research
for every skeet i post here showing ai systems catastrophically failing to do what they are supposed to do, i have dudes replying with “but humans are also bad at x" just don't. it's ignorant and unhelpful
llms don't and can't love, hate, feel, care, think, imagine, dream, hope, fear, believe, reflect, aspire, wonder, empathize,... fix your your anthropomorphising language
i dunno but using a machine that always regresses to the mean (llms) to simulate “diverse populations that are not readily accessible” seems the oppose of good science to me… no matter how popular it has become in behavioural sciences
in a phone conversation with family from back home who tells me something suspicious me: where did that information come from? family: "i googled it on chatgpt" i give up
i cannot emphasise it enough: rid yourself of the idea that the purpose of your research is to develop “solutions” to the clusterfuck the ai industry has thrown at us. i certainly am not interested in that or in making existing ai systems “better” 1/
the ai industry hijacked the term “safety” and ran it into the ground, stripping it of all meaning
safety evaluation research the past year or two: "we use llms to ideate & define “harm”, then use llms to generate prompts about it, which we then feed to llms to generate responses, we then use an llm-as-a-judge to classify the responses, we then use an llm to evaluate the accuracy…"
reading AI eval papers using llm-as-a-judge is alarming for many reasons; many “scientific claims” rely on evals executed on models that are now deprecated. how are we supposed to treat “findings” that are inherently not reproducible? i despair. llms are accelerating the decay of scientific rigor
Charlton McIlwain makes such a poignant point at #FAccT2026: what has the decades of work that has measured, documented, & evaluated the failure and bias of AI systems done? we have known the negative impact for decades yet very little has changed
