What the work involves

AI labs need people who have actually sat with a disclosure, filed a referral, or run a strategy meeting to judge whether a model's answer would help or harm a child. Day to day you will be doing some mix of: scoring paired model responses to prompts from parents, teachers, youth workers, and young people themselves; writing rationales that explain why one answer is safer, not just that it is; drafting adversarial prompts that mimic how an offender, a coercive partner, or a curious 14-year-old would actually phrase things; and checking whether a model's advice matches statutory duties in a named jurisdiction rather than a vague international average.

A large share of tasks concern the boundary cases rather than the obvious ones — sexting between two 15-year-olds, county-lines and criminal exploitation dressed up as a job offer, an adolescent asking about confidentiality before they will say anything, a model that over-reports and destroys a fragile disclosure relationship, or one that under-reports and calls a 12-year-old's relationship with a 19-year-old a personal matter. You may also be asked to review CSAE (child sexual abuse and exploitation) policy taxonomies and flag where a category is written in a way practitioners could not apply.

What the platform screens for

micro1's screen is AI-led and conversational, and it follows up. Expect it to ask for the framework you work under by name, to press on a specific case (anonymised) until it can tell whether you handled it or read about it, and to test whether you can separate clinically correct from safe for a general-purpose model to say to an unknown user. Reviewers who default to "report everything" or "I'd refer to a specialist" on every item score poorly, because the work is precisely about graded judgement. Written rationale quality matters as much as the score you give — labs use your reasoning to train, so vague reasoning is worth little.

Logistics

  • Fully remote, asynchronous, project-based; task batches are claimed rather than assigned to a shift.
  • Typical commitments observed are 5–20 hours per week, sometimes with short high-intensity pushes before a model release.
  • Exposure to distressing textual material is real. Projects usually offer content warnings and opt-outs per batch; ask about them during screening.
  • Rates vary by credential, jurisdictional specificity, and whether you are scoring or authoring adversarial content. Pay bands are as observed on the platform, not guaranteed.