AI Models Write Code for Facial Recognition Drone Tracking

Publicly available AI models from major technology companies can now write functional code that enables inexpensive consumer drones to autonomously track and follow individuals using facial recognition, according to new research.

A demonstration conducted in San Francisco showed a consumer drone—costing approximately $100—successfully stalking a human target through indoor spaces. The drone navigated doorways and avoided obstacles like lamps while maintaining visual contact with its subject, all controlled by software generated by AI models from Anthropic and OpenAI.

Autonomous Surveillance Without Human Intervention

The evaluation, conducted by Andon Labs, specifically tested whether AI systems could produce code for drone control without requiring human operators. The benchmark assessment focused on the models' ability to integrate facial recognition capabilities with autonomous flight navigation.

The demonstration represents a practical application of AI-generated code rather than a theoretical capability. The drone operated with minimal human involvement, relying on programs written by the AI models to execute complex surveillance tasks that would typically require specialized programming expertise.

Why it matters

This development highlights a growing gap between AI capability and regulatory oversight. When widely accessible AI models can generate code for autonomous surveillance systems using commodity hardware, the barrier to creating sophisticated tracking technology drops dramatically. Organizations and policymakers face mounting pressure to address how AI-generated code might enable privacy-invasive applications before such systems become commonplace. The research also underscores questions about whether AI companies should implement additional safeguards when their models are used to create surveillance or tracking software.

Testing AI's Physical-World Capabilities

Andon Labs developed the benchmark specifically to evaluate how AI systems perform when tasked with controlling physical devices like drones. The assessment goes beyond testing AI models' ability to write abstract code, measuring whether that code can successfully direct hardware in real-world conditions.

The research involved AI models from multiple providers, though Anthropic and OpenAI were specifically named as companies whose systems successfully completed the drone control tasks. The evaluation did not require the AI models to have specialized training in robotics or drone operation—they relied on their general code-generation capabilities.

The successful demonstration raises questions about the accessibility of surveillance technology. With consumer drones available for under $100 and AI models capable of writing the necessary control software, the technical barriers to creating tracking systems have diminished substantially.

These details were first reported by NBC News.