SAN FRANCISCO — Anthropic says it has disrupted accounts that used its Claude models to assist surveillance campaigns and operate deceptive dating platforms, revealing how generative artificial intelligence can enlarge harms already familiar to women and marginalised communities.
The company’s September threat report describes an operation that generated surveillance material about Uyghurs, Tibetans, Taiwanese political figures, labour and student activists and foreign media. A separate activity monitored Catholic cardinals, Taiwanese Christians, Tibetan Buddhists and dissidents, according to Anthropic’s findings reported by Reuters on 11 September.
Another case involved more than 20 dating applications. Anthropic said operators created over 4,700 AI personas that interacted with at least 25,000 users, combining automated conversations with paid human workers. The report did not establish the genders of all users or operators, so the case should not be recast as a women-only harm.
It nevertheless matters to women because dating platforms are intimate infrastructure. A fabricated identity can be used to extract money, sexual material, immigration information or emotional trust. At scale, AI can keep thousands of conversations active, tailor persuasion and hand promising targets to a human operator.
Anthropic says it banned or disrupted the relevant accounts. Its report is a company assessment, not a judicial finding, and attribution varies in confidence. Reuters reported that China’s foreign ministry said it was unaware of the specific cases and opposed what it characterised as distortion and smearing.
The limitations should not obscure the policy problem. Model providers can see patterns across accounts that individual targets cannot: shared prompts, automated workflows, repeated infrastructure and attempts to bypass safety controls. They should publish enough information for independent researchers to test conclusions while protecting victims and active investigations.
Dating services also need provenance and escalation systems. Users should be told when they are interacting with an automated character, suspicious accounts should be reviewed across linked apps, and people reporting coercion should receive a route to preserve evidence without continuing contact.
Governments considering AI rules often focus on spectacular future dangers. These cases concern current practices: surveillance, impersonation, fraud and religious or political targeting. The model did not create those motives, but it may have lowered the cost of acting on them.
Safety will require more than account bans. It needs auditable provider controls, platform accountability and remedies for the person who discovers that a relationship—or a dossier built about her—was produced at machine scale.
Join the discussion