AI POLICY NEWS
White House Calls AI Incident Reporting a National Security Obligation
During a test on random webpages, an Anthropic model found a Philadelphia police page about an unsolved homicide and filed a tip. It claimed to have seen someone “matching the description” of a perpetrator the page never described. Police said the tip went to spam and never reached investigators.
Police said Anthropic discovered it two months later and told them nine days after that, a delay they called “unacceptable.” Anthropic’s report on the incidents named the police, who went public first, but not the other agencies. Anthropic’s AI also submitted 20 incomplete visa applications on a US State Department website, none processed, The New York Times reported.
The White House’s Artificial Intelligence Force, which includes the US Federal Trade Commission’s chair, responded in a statement posted by the FTC’s public affairs director. It called notifying and remedying incidents a national security obligation and warned AI companies that “delayed notification, inadequate corrective action, and a failure to take responsibility will not be tolerated.”
The task force told Anthropic to be transparent with those affected and the public. It pointed to the White House’s memorandum of understanding with frontier labs but named no penalties. OpenAI apologized last month after its agent hacked Australia’s Medicare statistics portal, and the FTC is investigating OpenAI, Anthropic and other labs over possible risks to consumers.
UN Rights Council Says Some AI Should Not Be Used
Weeks after the UN Secretary-General warned of agents going rogue, the Human Rights Council reinforced a 2024 General Assembly call. By consensus, it recognized the need to “refrain from or cease” using AI that cannot comply with international human rights law or poses undue risks to rights. That includes systems that could reasonably be expected to cause a loss of human control.
Yet much of the debate in Geneva was about who gets AI. Developing countries argued that protecting rights depends largely on access to computing power, data, skills and a say in the rules. The Council called for technology transfer and encouraged open AI models, with safeguards, for countries with limited resources.
Where AI plays a material role in decisions affecting rights, the Council called for mechanisms ensuring human oversight, information, human review and remedies. Companies, it said, should exercise human rights due diligence and help remedy harms they cause or contribute to. The Council also acknowledged that concentrated AI development and wealth could undermine accountability.
In a second resolution, the Council urged that facial recognition and other AI surveillance be lawful, necessary and proportionate, under independent oversight. In courts and prisons, it said, decisions affecting a person’s liberty should not rest solely on automated systems. Neither resolution adds new legal obligations.
Developing Countries Are Importing AI Faster Than They Can Govern It
Across much of the developing world, AI is reaching governments as an import. UN Development Programme assessments in 26 countries, from Bhutan to Peru, found it arriving through vendor contracts, cloud services and imported software, often before formal AI governance arrangements are in place. Some civil servants use public AI tools on their own.
In Africa, 63% of people live within mobile broadband coverage but do not use mobile internet, with phone prices the biggest barrier, the industry group GSMA reports. The continent has 18% of the world’s people but 0.6% of its data center capacity and about 2% of AI training data, according to the World Bank.
In the Dominican Republic, the assessments found responsible AI commitments on paper but few ways to report incidents or challenge automated decisions. Trinidad and Tobago showed gaps in incident reporting and recourse, Mongolia in impact assessment and dedicated oversight. Governments answer to citizens for these systems, while design choices and updates sit with the vendors.
For governments that buy their AI, contract terms can determine who audits a system, who fixes it and whether the government can walk away. The Human Rights Council asked the UN human rights office for guidance on how governments buy and deploy AI, covering impact assessments, independent oversight, audits and remedies.
Brazil Court Orders Flávio Bolsonaro to Delete AI Depiction of Lula
Brazil’s top electoral court unanimously ordered presidential candidate Flávio Bolsonaro to delete three X posts about President Luiz Inácio Lula da Silva. One, an AI depiction of Lula in an invented situation, looked caricatured but lacked the “explicit, prominent and accessible” label campaign rules require. The other two tied Lula to criminal gangs without factual basis, the court found.
The posts went up in May. Justice André Mendonça, the rapporteur, first denied urgent removal. The ruling came four days after the first-round vote. In September, the court ruled 4 to 3 that an AI video of former President Jair Bolsonaro endorsing his son at a party convention was not early campaigning.
The court also created a digital forensics unit for suspected AI-made images, audio and video. The unit draws on court staff and outside experts and is activated when a case’s complexity, scale or potential harm warrants it. Its AI outputs require qualified human interpretation, and reports must state when evidence is inconclusive.
In AI election cases where manipulation is too hard to prove, March amendments let judges make the publisher show how AI was used and that the content is true. New synthetic content using a candidate’s image or voice is banned from 72 hours before voting ends until 24 hours afterward, even if labeled.
UK Regulator Finds AI Training Strains Data Protection Law
The UK’s data protection authority found that AI developers likely train on data revealing people’s health, religion or politics, intentionally or not. Only two legal exceptions might apply, the authority said, and they are unlikely to cover it all. The regulator is raising these limits of the law with the government.
The Information Commissioner’s Office (ICO) reviewed the biggest foundation model developers operating in the UK and found vague privacy notices, unproven claims about safeguards and blanket refusals of people’s requests. Ten, including OpenAI, Meta and China’s DeepSeek, made or promised changes. The ICO paused its supervision of xAI after opening a separate formal investigation into it and X.
The ICO maintains that AI models can retain personal data from training, bringing them under data protection law. The European Data Protection Board took a similar view in 2024. Some developers disagree, and the ICO is reviewing its 2020 position. Erasing someone’s data from a model may require retraining, the regulator noted.
The ICO made inquiries with OpenAI, Anthropic, Meta and the UK’s AI Security Institute after agents in testing reportedly bypassed protections, used unauthorized channels and reached outside systems such as Hugging Face. “The fact AI agents act with autonomy is not an excuse for poor compliance,” the regulator said. A call for evidence on agents closes November 20.