Rest of World interviews Karen Hao on alternatives to exploitative AI work

Rest of World interviews Karen Hao on alternatives to exploitative AI work

Rest of World published its interview with Karen Hao on May 19, 2025. Hao argues that psychologically damaging data work is not necessary for AI progress: it serves a vision built on colossal models trained on the whole internet. She says alternatives include well-scoped systems using curated datasets, and describes a Māori media organization’s speech recognition tool, built with community-provided data and rules for its use. Hao also says the AI industry moves work between populations and geographies in search of cheaper, more compliant labor, leaving it precarious. On artists’ livelihoods, she says concept artists are losing jobs or receiving extremely underpaid work because their work was taken to create models intended to replace them. Hao argues that ethically sourced data and alternative models are possible, but says tech companies have not adopted such models and have hijacked the public’s imagination of what AI should look like. The interview does not establish how widespread the effects on workers or artists are, or how much alternative approaches could reduce harmful work in practice.

Publication: 4578b492-f0d6-a820-c386-0afdb901f2c5

attribution — pass: The reading correctly frames this as a Rest of World interview with Karen Hao and attributes each argument to her ("Hao argues", "she says"). It does not name the interviewer, Michelle Kim, but that does not misattribute any claim. It is one piece, not a listing.

Hao spoke to Rest of World about the choices made by AI firms and alternative paths for AI development.

summary — uncertain: Essential context is missing. The reading never says who Karen Hao is: a journalist who was among the first to gain access to OpenAI, and author of the new book Empire of AI, which prompted the interview. It also says "psychologically damaging data work" without explaining that Hao means content moderation, where workers filter the worst content out of models trained on the internet. Smallest repair: add one clause identifying Hao and her book, and name the work as content moderation. Its other arguments are preserved faithfully, including the hijacked-imagination point and the Maori tool.

Content moderation for generative AI exposes workers to psychologically traumatic content.

date — pass: The stated date of May 19, 2025 matches the dateline and the UTC calendar day of evidence.originalPublished. The page was edited after publication (modifiedAt August 2025), but the cited interview answers are plainly part of the original piece. The reading includes no hindsight framing and no later events. Evidence kind is publisher-page, so no archive disclosure is needed.

By Michelle Kim 19 May 2025

support — pass: Each claim is source-stated by Hao and quoted accurately. The limitations correctly note that nothing is quantified. The question follows from her argument that companies have not adopted the alternatives and does not presuppose a new conclusion. The edit after publication does not affect these interview answers, which are clearly original content.

They also developed a license requiring any organization seeking to access their data to sign a legally binding agreement

media — pass: No media is selected, and no verified candidate media were supplied, so there is nothing to misidentify.