Leaked docs show Outlier and Scale AI use freelancers to write prompts about suicide, abuse, and terrorism to stress-test AI, urging creativity but banning CSAM
- BI obtained training docs showing how freelancers stress-test AI with “harmful” prompts. — Outlier and Scale AI use freelancers … Bluesky: @theopriestley.com and @hypervisible Bluesky: Theo / @theopriestley.com : Check LinkedIn and other job boards, they're awash with these types of freelance gigs now paying min. wage to spend hours writing prompts and assessing the outputs to “improve” AI. [embedded post] @hypervisible : “Freelancers are encouraged to ‘stay creative’ as they test AI with prompts about torture or animal cruelty, leaked training documents obtained by Business Insider show.”
Context & Ripple Effects
The leak adds detail to a long-running shift in AI training from generic labeling toward judgment-heavy work. Scale AI and peers had already recruited writers and poets for model-improvement tasks, while journalists were more recently being recruited for fact-checking and prompt drafting as AI-training freelancers.
It also makes the safety workforce itself part of the product story: earlier reporting documented outsourced labeling of violent and toxic material to improve ChatGPT. The documents identify a defined boundary—CSAM is excluded—even as contractors are asked to devise other harmful test cases.
First-order effects
- Outlier and Scale AI freelancers are tasked with creating and assessing adversarial prompts across sensitive harms, making human judgment and exposure central to stress-testing workflows.
- The stated CSAM prohibition sets an operational limit on what these contractors may generate, while leaving other high-risk categories within the testing remit.
Second-order effects
- Training-data vendors will need task design, reviewer guidance, and quality controls that can elicit difficult edge cases without turning open-ended creativity into uncontrolled harmful-content production.
- As specialized freelancers become a larger input to model evaluation, providers competing for capable contributors may face greater pressure to differentiate on task safeguards and work conditions.
Third-order effects
- If this approach scales, AI safety evaluation will increasingly depend on a distributed human labor layer rather than model-only testing—an expanding AI enforcement surface with recurring questions about accountability for contractor welfare and output quality.
- The contrast between expanding high-skill prompt work and earlier toxic-content labeling suggests a more segmented training-data labor market, where expertise and exposure are priced and managed differently.
The trend: AI developers are industrializing adversarial evaluation through specialized freelance work, turning safety testing into a managed supply chain of human judgment.