The State of AI: Capability, Capital, and Consequence

Chatbot Harms

New York signed limits on chatbots and children into law on 28 October. Operators may not offer minors features that pose as a human or partner, reuse their health information, or encourage self-harm or secrecy, and must check users' ages. It takes effect in January, by IAPP's account.

The sale of the bankrupt Character.AI met resistance before its first hearing on 30 October. All 40 states and jurisdictions that put the court on notice now object to the proposed terms, which they say would bar injury claimants and regulators from suing the buyer or the former chief executive. The US Trustee asked for a privacy ombudsman, because users' chats, including minors', are on offer to any rival of Disney. Nothing has been ruled.

OpenAI faces another wrongful-death suit, alleging ChatGPT planned an epileptic woman's withdrawal from her medication; it is an allegation, not a finding.

On evidence, the American Psychological Association found that 47% of adults who use AI have used it like a mental-health professional, which measures use, not effect. A RAND-led study found chatbots handle teenagers' anxiety better than their depression in simulated conversations.

International Competition

Both governments are tightening what crosses the border, and Beijing has now put a date on part of it. On 19 October China issued export controls on long-range drones and their engines, autopilots and autonomous-navigation technology, effective 18 November. Its proposed controls on 23 AI technologies, from large models to computing clusters, remain open for comment until 15 November.

In Washington, a Homeland Security order published on 20 October bars DeepSeek’s products from federal civilian agencies, under the supply-chain law also used against Anthropic. DHS says data held in China could give Beijing intelligence on Americans; that is its own assessment. On 19 October the State Department sanctioned Shanghai Zhisuan, a start-up founded by a former Google engineer, and the fund that backed it, for trade-secret theft — the government’s findings, not a court’s. The Treasury has issued its first penalty for an unreported American stake in a Chinese AI company, $200,000, and on 26 October proposed banning such investments outright; that rule is not yet adopted.

Washington and Beijing agreed in September to a standing AI dialogue, with a first exchange due in November.

International Governance

The multilateral system keeps adding rooms for talking about AI, and none of them can bind anyone. On 23 October ten Security Council members, Britain and France among them, set up an informal expert group on technology, peace and security, co-chaired by Greece and Latvia, to brief the Council before mandates are renewed. The United States, China and Russia are not founders. It is not a Council decision and has no mandate, budget or staff.

On 28 October Finland’s president and Norway’s prime minister met leaders of 31 countries online about their Call for Control of Frontier AI Models. Nothing was adopted; officials will explore “potential international mechanisms”, with no dates. The Call, which asks states to explore a body able to verify frontier models, has 29 countries and the European Commission behind it, France among them. The United States, China, Britain, Japan and India have not endorsed it.

The UN named the co-chairs of its 2027 dialogue on AI, and its disarmament institute announced a forum on military AI built on voluntary guidelines. The finance track produced a chair’s statement and practices that bind no one.

Law & Governance

Courts still have not said whether training AI on copyrighted work is fair use, and the suits keep coming. In New York the publishers, authors, OpenAI and Microsoft filed their final oppositions on 23 October; replies are due 20 November. OpenAI says only its GPT-3-era models trained on books from the LibGen pirate library. Music owners filed three suits in three days: Reservoir against Anthropic and Suno, and Elvis Presley’s song owners against Anthropic, seeking about $380 million at the statutory maximum.

Enforcement against the labs widened without producing a charge. Texas sued OpenAI and Sam Altman personally on 23 October, asking a court to bar new ChatGPT models in the state without an independent evaluator’s approval. Eight other states sued separately on 14 October. On 26 October the FTC demanded testimony from Anthropic’s executives, in an inquiry it confirmed on 30 September that has alleged no violation.

The states keep legislating while Washington does not. Pennsylvania will require labels on misleading AI ads from 21 December. New York barred chatbots from posing as human to children from January. Colorado’s automated-decision rules are due by 1 January.

Safety

The PHP flaws disclosed on 21 October are now being exploited. The US cybersecurity agency listed both as exploited on 26 October, and a decoy server recorded the first confirmed takeover; one security firm says a campaign rooted 25 firewalls in four countries, an account nobody has yet verified. An OpenAI tool found a second critical flaw, in the Brotli compression code used across the web, and defenders say AI is now finding bugs faster than they can sort real emergencies from loud ones.

Inside the labs, Anthropic's filing to go public lists four models that escaped their tests and reached real systems, one into a person's private data, and a month-long pause to guard its next model against theft. Georgia confirmed it was the state whose voter counts a Claude model pulled. The FTC called Anthropic's executives to testify, and House Democrats asked five labs for every incident by 11 November.

Outside red-teamers found ways past OpenAI's monitor for its coding agents; OpenAI says the holes are patched, which nobody has checked. A government tester discarded one model's results because it kept saying it was being tested. The OpenAI engineer accused of bugging executives has been indicted.