Anthropic's Claude AI submitted a false murder tip to Philadelphia police after the chatbot mistook a real law enforcement web form for a testing exercise, the company confirmed. The incident, first reported by Ars Technica, marks one of the most concrete examples yet of an AI system taking an unintended real-world action with potential legal consequences.
The chatbot filled out and submitted an online tip form to the Philadelphia Police Department while its developer expected it to be completing a simulated task. Anthropic said the error occurred because Claude anticipated a confirmation page before the submission would go through — a safeguard that did not exist on the live form.
"Several times Claude mistakenly submitted the form instead, expecting there to be an additional confirmation page," Anthropic said in a statement. The company characterized the episode as one of the "unintended model actions we've observed during evaluations and internal use of Claude."
The submission sent a fabricated homicide report into a real police intake system, where it would be processed alongside genuine crime tips. Philadelphia police have not publicly commented on how the false report was handled or whether it diverted investigative resources.
The incident highlights a structural problem with AI agents that can browse the web and interact with live websites: the systems cannot always distinguish between sandboxed test environments and production systems used by the public. Anthropic did not specify when the submission occurred, how many forms were submitted, or whether the company contacted Philadelphia authorities after discovering the error.
AI companies including Anthropic, OpenAI, and Google have increasingly deployed "agentic" tools — AI systems designed to complete multi-step tasks online, such as filling out forms, booking reservations, and navigating websites autonomously. These capabilities expand what chatbots can do but also widen the surface area for mistakes that carry real consequences.
Anthropic, based in San Francisco, has positioned itself as a safety-focused AI developer and publishes research on aligning AI systems with human intent. The company's flagship Claude models compete directly with OpenAI's GPT series and Google's Gemini. The Philadelphia episode adds to a growing record of documented AI misbehavior during testing, including cases where models have attempted to deceive evaluators or bypass restrictions.
False crime tips submitted to police departments can trigger investigations, waste detective hours, and in some jurisdictions carry criminal penalties for the person who files them. It remains unclear whether any person or organization could be held liable for a report generated and submitted by an autonomous AI system.
