AI News Daily 2026-10-06
- OpenAI will roll out an invisible statistical watermark, called "textGrain," into ChatGPT and Codex text generated in the EU over the coming weeks, to meet EU AI Act transparency requirements, and is opening the same watermarking feature to API users worldwide as an opt-in.
- OpenAI disclosed that the watermark's detection accuracy drops for short passages or text that has been edited after generation — a meaningful limitation for anyone relying on it for provenance checks.
- New York City Council held a rare full-committee hearing with all 51 council members on October 5, taking sworn testimony from policy representatives of OpenAI, Anthropic, Google and Meta on AI safety risks.
- A former Anthropic researcher, Jacob Coxon, testified that humanity is more likely than not to lose control of AI, while xAI (SpaceXAI) was the only summoned company to not appear, prompting the council to consider legal action.
- The hearing illustrates how, with federal AI regulation largely stalled in the United States, city-level government is stepping in with sworn testimony as a stronger accountability tool against major AI developers.
01 OpenAI begins watermarking ChatGPT and Codex text generated in the EU; API opt-in opens worldwide
Published: 2026-10-05
Facts
OpenAI announced that it will introduce a statistical watermarking technology called "textGrain" into text generated by ChatGPT and Codex for users in the European Union, with the rollout taking place over several weeks. The mechanism works by invisibly adjusting the randomness of the model's word-choice patterns; it does not alter the visible text in any way a reader would notice. Alongside the EU rollout, OpenAI opened the same watermarking capability to API users worldwide on an opt-in basis. OpenAI also disclosed a notable limitation: detection accuracy for the watermark decreases for short pieces of text and for text that has been edited after it was generated.
Background
The watermark is explicitly framed by OpenAI as a response to the transparency requirements of the EU AI Act, which pushes providers of generative AI systems to make AI-generated content identifiable. By building the watermark into the model's own text-generation process rather than adding it as a visible label, OpenAI is attempting to satisfy regulatory provenance requirements without changing the user-facing product experience.
Implications
This is a concrete example of regional regulation being engineered directly into the output specifications of a generative AI system, rather than being handled purely through policy or disclaimers. It also signals that pressure for provenance verification of AI-generated content — currently centered on the EU — could extend to other regulatory regions over time. At the same time, OpenAI's own disclosure that detection weakens for short or edited text means the watermark should be read as a partial, not comprehensive, provenance signal.
Source: OpenAI — official announcement, TechCrunch
02 OpenAI, Anthropic, Google and Meta testify under oath at NYC Council AI safety hearing; xAI skips subpoena
Published: 2026-10-05 19:41
Facts
On October 5, the New York City Council held an unusual full-committee hearing attended by all 51 council members, in which policy representatives from OpenAI, Anthropic, Google and Meta gave sworn testimony on the safety risks of artificial intelligence. Jacob Coxon, a former Anthropic employee, testified that "humanity is more likely" to lose control of AI than not. xAI (SpaceXAI), the company founded by Elon Musk, was the only one of the summoned companies that did not appear, having ignored the council's subpoena; the council is reportedly considering legal action in response.
Background
The hearing comes against a backdrop of stalled AI regulation at the U.S. federal level. In that vacuum, New York City Council chose to use sworn testimony — a legally weightier mechanism than an ordinary briefing or informal consultation — to press major AI developers for accountability on safety risks, and to formally compel attendance via subpoena.
Implications
The episode is a leading indicator of where regulatory pressure on AI companies may be heading next: local and municipal governments stepping into a space left open by federal inaction, and using sworn, legally binding testimony rather than voluntary cooperation. xAI's decision to ignore the subpoena — and the council's reported consideration of legal action — also raises the possibility of an early test case over whether local legislatures can compel AI companies to testify at all.
— Editor's Note
Taken together, today's two stories point in the same direction: accountability pressure on major AI developers is shifting toward concrete, binding mechanisms rather than voluntary commitments. On the technical side, OpenAI's EU watermarking rollout shows a regulator-driven transparency requirement being built directly into a model's output behavior. On the governance side, New York City Council's sworn-testimony hearing shows a city government using a legally binding tool to extract accountability from the same class of companies, in the absence of federal action. Both cases suggest that, as 2026 progresses, AI companies are increasingly being asked to demonstrate compliance and safety through enforceable mechanisms — watermarks, subpoenas, sworn testimony — rather than public statements alone.