Reposted by Adrià Auladell
This was a Stentor timelapse we made earlier this week from her trip to Boston (which she also had to crowdfund). She was magically able to locate one ‘on demand’ despite that I very rarely find one in Boston. She also brought me samples from all the way up the east coast. #sciart #microscopy
![However, recent work demonstrates that demographically prompted LLMs cannot accurately simulate diverse identities. Prompted LLMs are more likely to represent majority groups' stereotypes about minority groups than minority groups' own standpoints [12,13]. For instance, one study [13] prompted LLMs with the identity of Black woman. GPT-4 started generations with phrases such as 'Hey girl' and
'Oh, honey'; Llama-2 started nearly every response with 'Oh, girl', and used phrases including I'm like, YAASSSSS' and 'That's cray, hunty!' In addition, demographic prompting flattens diverse identities into narrow categories, limiting researchers' ability to observe and understand diversity within those categories [13,55]. Demographic prompting misrepresents people who are chronically under-represented in cognitive science research, a form of representational harm [56].](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:ho22o3jwfn4k7diklsrfnazt/bafkreifiqpzt5r2qjkewoxwxyutln7i4rg3x65pzbxbbgo7zp2by4wdqs4)
