New Test Tries to Read AI's Mind — and Catch It Lying

Researchers at Google DeepMind and elsewhere built a new test called WorkspaceBench. It's designed to check whether special software tools can correctly read what an AI model is 'thinking' while it works. Think of it like giving an AI an internal scratchpad — this test tries to read that scratchpad. The benchmark includes 3,356 questions across 27 categories, covering safety, logic, and multi-step reasoning.

Why does this matter to you? Because today's AI can do surprisingly complex tasks without showing its work — no visible 'reasoning' steps you can review. That's a problem if you want to catch mistakes, bias, or hidden motives. If a customer service bot decides to escalate your complaint, or an AI assistant drafts your legal email, wouldn't you want to know why? Reading an AI's internal state is how we eventually keep it honest and safe.

The catch: the tools doing the 'mind-reading' don't always get it right. Some are reliable but very limited (one word at a time). Others are more expressive but prone to making things up — what researchers call hallucination. WorkspaceBench measures both accuracy and how often a tool invents false information, so you can see the tradeoffs. The team tested it on a 27-billion-parameter model called Qwen-3.6-27B, and they're not sure yet how well it works on smaller models.

The benchmark is open-source, meaning other researchers and companies can use it to test whether new transparency tools actually work. Think of it as a 'consumer reports' for AI mind-reading — a shared yardstick. As AI gets more autonomous, being able to audit its reasoning will matter more and more. This is one early step toward holding AI accountable for what it does inside its head, not just what it says out loud.