Scientists Built a Lie Detector Test for AI's Inner Thoughts

When you ask an AI a question, you usually see only its answer. But inside, the model passes through a series of half-formed ideas before settling on a reply — researchers call this internal scratchpad the "global workspace." A team including Neel Nanda has built WorkspaceBench, a test suite that grades tools which try to read that scratchpad and turn it into plain English.

The timing matters. Newer AI systems can now do surprisingly complicated things without "showing their work" — that is, without writing out the step-by-step reasoning we could previously read and check. If the reasoning is invisible, the only way to know whether a model is being honest, hiding something, or just confused is to look directly at what is happening inside it. Think of it as moving from reading someone's diary to needing a brain scan.

The test is tough to build fairly. Researchers cannot simply open up an AI and see what it is "really thinking," so there is no answer key. They worked around this by designing 3,356 questions — covering safety, logic, and multi-step puzzles — where the model almost certainly has to use one specific idea to get the answer right. They also track a second score: how often a reading tool invents details that were never there. A tool that confidently makes things up is worse than useless for safety work.

Results so far show a trade-off rather than a winner. Simple tools are reliable but limited; more expressive ones say more but tend to fill gaps with fiction. The benchmark, built on an open-source model called Qwen-3.6-27B, is now public so anyone releasing a new mind-reading tool can be scored against the same yardstick. That is the real point: turning a vague promise into a number anyone can compare.