AI's answer keys are being written for $85 an hour by people who cannot get other work

Micro1's listing asks producers to author evaluation tasks, not to produce anything. It wants more than five years of credited project experience, and it wants that experience turned into scenarios covering budget reconciliation, scheduling adjustments and vendor or crew coordination, for up to $85 an hour ^[4]. The deliverable is an answer key: a specification of what a correct schedule looks like, so a model's attempt can be marked against something. Compute does not generate answer keys. Somebody who has lost a shoot day to a permit that never arrived does.

Ruth Fowler's assignment shows how dense the key has to be. She trained a model to schedule a hypothetical two-day shoot, including filming permits, location hazards, daylight conditions for photography, personnel per shot and cast needs such as child protection, all inferred from emails, a shooting script and other production materials ^[10]. She also taught a system to assemble a pitch deck, which she describes as a production executive's job ^[11]. No generalist writes that checklist.

Then the price. The reported range for this work runs from $12 to $200 an hour ^[1], a spread of roughly 17 times within one occupation ^[4], and the Micro1 ceiling sits at about 43% of the top of it ^[5]. The skill did not move that much. The alternative buyer went away. Between July 2022 and May 2026, US motion picture and sound recording employment gave up 124,000 positions ^[1], an average of about 2,700 a month over 46 months ^[2]. Fowler told the Guardian she began training AIs because she was always broke ^[9], and said the work amounted to teaching it how to take their jobs ^[9].

That last part does not travel. The same agencies hold contracts with AI companies including Anthropic and OpenAI, and are collecting hard-won skills in finance, health, law and social work as well ^[2]. Those workforces have not shed a quarter of their jobs. Boston Consulting Group's April analysis puts 10% to 15% of US jobs at risk of elimination and more than half of roles reshaped ^[8]. A reshaped professional still drawing a salary has little reason to write answer keys at Micro1's rate.

One more thing the buyer never sees. The anonymous documentary director was asked to build visual descriptions of each speaker in a Little League recording and to characterise their accents as English, Scottish, African American or Asian, which caused concern among the freelancers doing it ^[12]. Whoever writes those categories fixes what the model treats as correct, and that authorship has been subcontracted twice before it reaches anyone accountable for the output. Netflix says it used AI in 300 of its 1,000 titles in 2026 ^[5], which is 30% ^[3]. The demand side has already committed to output whose standard of correctness was set by people it will never meet, at rates set by how badly those people needed the work.