The Week AI Rules Asked Newsrooms Three Different Questions

Three rule-setters moved this week, and each one asked journalism for something different.

Twenty countries, two absentees. At the UN General Assembly in New York, twenty countries and the European Union signed a statement promising to keep artificial intelligence under human control. The United States and China — the two countries that build most of the world's AI — did not sign it. Al Jazeera and ABC News Australia both reported the declaration.

Delhi: the rules arrive as a purchase order. India's public broadcaster, Prasar Bharati, published a draft notice seeking studios and start-ups to make AI anchors, synthetic presenters, dubbing and archive restoration for its television, radio and streaming services. Inside it, the broadcaster wrote the country's synthetic-media rules into its own purchasing conditions: labels on AI-generated content, C2PA provenance metadata, on-screen disclosure of an AI presenter, no deepfake of a living public figure, elected official or religious figure without prior approval, no voice cloning without consent, and documented human review before anything is broadcast-ready. Comments on the draft closed on Monday.

New York: what the reporting is worth. Filings in The New York Times' copyright case against OpenAI and Microsoft were unsealed last Thursday. Digiday reported what they contain: OpenAI's head of ChatGPT, Nick Turley, writing that publishers face an existential threat from OpenAI's products and calling those products "largely substitutive, period"; Microsoft's own measurement of click-through rates from its search engine to the Times and the Daily News running 83 to 93 percent lower inside its Copilot chatbot than in ordinary search; and Satya Nadella's testimony that chatting with a bot had substituted going to the publisher's website. Publishers' lawyers argue the combination eviscerates the fair-use defence. Both sides have asked Judge Sidney Stein to rule without a trial.

Brussels: a rule that already has dates. The European AI Act's transparency rules have applied since 2 August 2026. Deepfakes have to be labelled, as does AI-written text about matters of public interest that nobody reviewed or edited, and providers have to embed machine-readable marks in what they generate. Enforcement is running, with fines reaching 15 million euros or 3 percent of global turnover. Systems already on the market in August have until 2 December 2026 to meet the marking obligation. For a newsroom, the line worth knowing: fixing spelling and grammar does not count as generating content, while publishing an audio clip in a voice that sounds like a real person does. The Commission's transparency guidelines set out how the obligations are read.

None of the three answered the question underneath them: who pays for journalism when an assistant answers the question instead of linking to the reporting.