Process Beats Output: The New Turing Test

🔥 What happened Princeton and Stanford researchers dropped a "Process Turing Test" that judges AI by how it solves problems, not what it answers. Across decision-making, working memory, and planning tasks, their process-based classifier hit an AUC of 0.88 — even when task performance was identical. 💡 Why it matters Old CAPTCHAs and Turing tests break the moment a model gives the right answer. Process features still expose it: Claude Sonnet 4.5, GPT-5, and Gemini 2.5 Pro all get caught when you watch the path, not the result. Fine-tuning on 10.7M human decisions makes agents more human-like — but that edge vanishes under cross-task transfer. ⚡ Our take Judging AI by output alone is security theater. The real bottleneck isn't the answer — it's the process, and that's exactly where we need to aim.