AI oversight tightens

U.S. regulators are moving from AI warnings to AI enforcement, with new takedown powers arriving as chatbots spread false medical claims and major services stumble. (cyberscoop.com) The Federal Trade Commission said it is preparing “robust enforcement” of the TAKE IT DOWN Act, which makes platforms create a notice process and remove nonconsensual intimate images, including AI fakes, within 48 hours. The law was signed in May 2025, and the Federal Trade Commission says it will enforce that notice-and-removal system. (ftc.gov, cyberscoop.com) CyberScoop reported on April 20 that the agency is also looking harder at voice-cloning scams, after Federal Trade Commission officials said at recent events that AI impersonation and synthetic media are climbing the priority list. Commissioner Mark Meador said the agency was “actively spinning everything up” for enforcement once complaints can be filed. (cyberscoop.com) At the same time, evidence is piling up that consumer-facing chatbots still fail basic safety tests in medicine. A BMJ Open audit of five public chatbots found 49.6% of answers to 250 health prompts were “somewhat” or “highly” problematic, with only two refusals to answer. (bmjopen.bmj.com) The researchers tested Gemini, DeepSeek, Meta AI, ChatGPT, and Grok across cancer, vaccines, stem cells, nutrition, and athletic performance. BMJ Group said half of the answers could plausibly steer lay users toward ineffective treatment or harm if followed without professional guidance. (bmjgroup.com, bmjopen.bmj.com) A separate Nature report showed how easily those systems can absorb bad information once it appears online. Researchers invented a fake condition called “bixonimania,” seeded it in bogus papers, and Nature reported on April 7 that chatbots then warned users about the fictional illness as if it were real. (nature.com) Reliability has become part of the same policy debate. On April 20, OpenAI acknowledged an outage affecting ChatGPT and Codex, and outside outage trackers said the disruption spread across web, mobile, and application programming interface access before service recovered by April 21. (9to5mac.com, status.openai.com, statusgator.com) The enforcement case is getting more concrete, not just more theoretical. CyberScoop reported that the Department of Justice won its first conviction under the new federal law this month when James Strahler of Columbus, Ohio, pleaded guilty in a case involving AI-generated deepfake nudes used to harass at least six women. (cyberscoop.com) The result is a narrower, tougher agenda for AI in 2026: remove illegal synthetic images fast, treat cloned voices like fraud tools, and stop presenting shaky chatbot answers as dependable help in health or other high-risk settings. (ftc.gov, bmjgroup.com, cyberscoop.com)