CAPTCHAs can still detect AI agents
4 hours ago
- #Cognitive psychology
- #AI detection
- #Process Turing Test
- AI systems match or surpass humans in many tasks but differ in cognitive processes, enabling detection of AI agents.
- CAPTCHAs are not reliable for distinguishing humans from AI because AI solves them differently, with distinct error patterns and behavioral features.
- The Process Turing Test measures whether AI processes are indistinguishable from humans, going beyond output-based Turing Tests.
- CogCAPTCHA30 combines a CAPTCHA with 29 cognitive tasks to evaluate both output and process equivalence between humans and AI.
- Smaller AI models like Qwen and Centaur are more human-like in process than larger frontier models like GPT, Claude, and Gemini.
- Process-based discrimination is robust when AI lacks full access to evaluation criteria but can be bypassed with complete information.
- Simulating human cognitive psychology remains a significant challenge, and AI progress does not inherently make AI more human-like.
- The Process Turing Test offers a more durable method for human verification compared to traditional methods like passwords or CAPTCHAs.