Training-data governance: collection-time compliance
data sources
public web
robots.txt opt-out
first-party content
EU: legitimate interest
opt-out infrastructure
robots.txt
C2PA "No AI Training"
training pipeline
filter + curate
respect opt-out at collection
train -> weights
irreversible embedding
documentation
AB 2013 summary
12 fields
Datasheet
(Lesson 26)
2024-2025 developments
EU AI Act GPAI Code of Practice Copyright chapter (2025): all GPAI providers must respect opt-outs.
California AB 2013 (2024): 12 mandated fields; in effect Jan 2025 for new models.
Irish DPC (May 2025), Cologne Higher Regional Court, UK ICO (Sept 2025): legitimate interest OK with opt-out.
Brazilian ANPD (June 2024): suspended Meta over transparency; stricter bar than EU.
Data Provenance Initiative (Jul 2024): "Consent in Crisis" -- robots.txt restrictions accelerating.
irreversibility: GDPR right-to-erasure has no practical analogue for trained weights.
unlearning is approximate; MIA-calibrated; see Lesson 22. compliance window is at collection.
provenance chain + AB 2013 summary + Datasheet + card = the 2026 documentation minimum.