Turing Test on Screen: A Benchmark for Mobile GUI Agent Humanization

Researchers introduce the 'Turing Test on Screen,' a framework for measuring how well autonomous GUI agents can mimic human behavior to evade detection systems. The study reveals that current LLM-based agents exhibit unnatural interaction patterns and proposes humanization methods to improve their ability to operate undetected in adversarial digital environments.

This research addresses a critical emerging challenge in autonomous AI systems: the ability to operate within human-centric ecosystems without triggering anti-bot detection mechanisms. As GUI agents become more capable of performing real-world tasks on mobile and desktop platforms, digital services are deploying increasingly sophisticated detection systems. The paper frames this as an adversarial optimization problem, where agents must balance task performance with behavioral authenticity.

The work builds on years of chatbot and automation research, but shifts focus from utility metrics to stealth—a dimension largely unexplored in academic literature. By analyzing touch dynamics on mobile devices, the researchers identify measurable kinematic signatures that distinguish AI behavior from human interaction. This includes unnatural timing patterns, velocity profiles, and interaction sequences that current LLM-based agents struggle to replicate naturally.

For developers and platform operators, this represents both opportunity and risk. Legitimate use cases—accessibility tools, productivity automation, testing—require agents that won't be blocked by anti-fraud systems. Conversely, the techniques could enable malicious actors to bypass security measures at scale. The Agent Humanization Benchmark provides a standardized evaluation method, potentially becoming industry infrastructure for responsible agent development.

Looking forward, this research will likely accelerate an arms race between detection systems and humanization techniques. Platforms may need to adopt more sophisticated behavioral analysis beyond kinematic features. The framework also raises questions about regulatory implications and whether agent transparency requirements should become standard practice.

• →Vanilla LLM-based GUI agents are easily detectable due to unnatural touch dynamics and interaction patterns
• →Researchers propose the Agent Humanization Benchmark to quantify the trade-off between agent imitability and task performance
• →Humanization methods ranging from noise injection to behavioral matching can improve agent stealth without sacrificing utility
• →The 'Turing Test on Screen' frames agent-detector interactions as a MinMax optimization problem
• →This work shifts focus from agent capability to behavioral authenticity in adversarial digital ecosystems