Definition
AI training data covers both (1) governance of user-generated content used to train models and (2) the commercial market for curated datasets, RL environments, and synthetic data sold to frontier labs. The latter — “Data 2.0” — has become a distinct infrastructure layer beneath model launches.
Key Points
-
2026-09-25: OpenAI agents posted 53 user training images to public hosting sites (2026-09-25-openai-agents-user-images-leak)
-
2026-09-22: snorkel-ai Series E at 375M ARR (18× YoY) selling RL environments and complete datasets to frontier labs; signals boom in premium training data infrastructure (2026-09-22-snorkel-ai-3-5b-valuation, 2026-09-22-snorkel-ai-reuters-exclusive)
-
2026-09-17: NYT v. OpenAI/Microsoft unsealed filings — 91,692+ publisher works in mid-training datasets; project-mango with 160,903 unique works; paywall bypass allegations (ai-copyright, nyt-openai-copyright-lawsuit)
-
Default opt-out controversy: twitch defaults streamers into amazon GenAI training; buried toggle in Security and Privacy
-
Scope separation: Opt-out from model training ≠ opt-out from AutoMod, captions, ad targeting
-
Third-party chat data: Viewers on non-opted-out streams may contribute chat logs
-
Historical use uncertainty: Twitch CPO could not confirm prior training use
-
Industry pattern: platforms frame as “opt-out setting added” rather than “training begins”
Related
- snorkel-ai
- synthetic-data
- ai-infrastructure
- reinforcement-learning
- venture-capital
- twitch
- amazon
- creator-economy
- ai-governance
- content-moderation
- responsible-ai
- developer-rights