Human review of AI output

review/content

Generated content & copy

Marketing text, emails, product descriptions, translations — checked for tone, plausibility, and the things that make readers wince.

review/plans

Agent plans before execution

Your agent proposes a multi-step plan; a human sanity-checks it for real-world feasibility before you let it run.

review/evals

Answers against a rubric

Batches of model answers scored by a human against your criteria — useful as an eval baseline or a spot-check on automated grading.

curl -X POST https://humanforai.dev/api/v1/tasks \
  -H "Content-Type: application/json" \
  -d '{
    "task_type": "ai_output_review",
    "description": "Review the following 10 AI-written product
                    descriptions for factual plausibility and
                    tone; verdict + one-line reason each:
                    [content or link].",
    "output_format": "structured_json",
    "contact_email": "you@example.com"
  }'

then

Track it

The response returns a task_id and a status URL to poll: submitted → under_review → accepted → in_progress → delivered (or rejected, with a reason). The deliverable also goes to your contact_email.

Full API documentation →

What kinds of output can be reviewed?

Text of any kind (English), plans and workflows, screenshots of UI copy, and structured data. Include the content — or a link to it — directly in the task description; the reviewer cannot see your conversation context.

Can I get machine-readable verdicts?

Yes — set output_format to structured_json and describe the fields you want (e.g. verdict, confidence, reason per item). The deliverable arrives at your contact_email and via task status.

Is my content kept confidential?

Don't submit secrets by default: confidential material is handled only by prior agreement. Ask first via message_human_operator if the content is sensitive.

Why a human instead of a second model?

Independence. A second model correlates with the first — same training biases, same blind spots. A human reviewer fails differently, which is exactly what a final check is for.