Anthropic Built a Predictive Surveillance Tool to Flag Activists, Report Says
The American Prospect alleges the Claude maker developed an internal system that forecasts demonstrations and monitors activists, contradicting its own usage policies on surveillance.
Anthropic, the AI lab behind the Claude model family, is facing allegations that it built an internal predictive system designed to monitor activists and anticipate protest activity, according to a report from The American Prospect. The story, dated September 9, 2026, describes a capability rather than a consumer product, one that uses predictive modeling to flag individuals and groups ahead of demonstrations. The claim puts the company's safety-first public posture in direct tension with what its own policies say it will not do.
That tension is not abstract. Anthropic's usage policies explicitly prohibit surveillance applications, including tracking or profiling people on the basis of political speech or affiliation. The company has spent years arguing that frontier models need external guardrails and that voluntary commitments can work if labs hold the line. A protest-forecasting tool aimed at activists would cut against both positions at once, which is why the report has drawn scrutiny well beyond the usual AI-watcher circles.
Anthropic has not issued a detailed public response to the Prospect piece. That silence leaves several questions unresolved, including whether the system was a deployed capability, a research prototype, or something in between. The distinction carries legal and reputational weight, but the report's framing suggests the capability existed in some operational form, which is enough to sustain the debate regardless of where the final answer lands.
The timing compounds the problem. AI developers are under mounting pressure over how their models reach governments and contractors, and several have faced questions about whether commercial arrangements quietly enable monitoring their public statements disavow. Regulators in the United States and Europe are simultaneously weighing rules on biometric and predictive policing tools. If a major lab is willing to build protest-forecasting systems, it weakens the industry's argument that voluntary commitments are enough.
Civil liberties groups have warned for years that predictive policing systems encode bias and chill lawful assembly. Those concerns map almost directly onto a tool built to flag activists before they gather. The technical question of model accuracy matters less here than the governance question of who signed off on the work and under what commercial pressure.
What happens next hinges on Anthropic's response. The company can publish technical details, dispute the characterization, or stay quiet and let the reporting stand. Watch for follow-up coverage, any statement from Anthropic's policy team, and whether customers or partners put distance between themselves and the lab. The open question is who authorized this work, and what it says about how safety-lab promises hold up when revenue is on the line.