Working from that record, I've contributed MCAB, a cognitive-autonomy benchmark, to ImpactBench at the MIT Media Lab, built three instruments you can open and use, and designed the study that will run that benchmark with people for the first time.
973 conversations, coded and classified into a five-phase longitudinal record, read as data, not memory.
A writing probe for the gap between how you felt you answered an AI (accepting, revising, resisting) and how far your writing actually stayed inside its frame.
The Modulated Cognitive Autonomy Benchmark. It asks one thing of an AI response. Does it compress the user's thinking, preserve it, or expand it? Built from the record. It works on anyone's.
An instrument for the moment before a thought becomes a conclusion. You place things. It reads where you put them.
Four AI advisors, built as a family, each biased on purpose. One rushes. One wants numbers. One argues with whatever looks obvious. One skips straight to what you already wanted.
Image and video guardrails for Samsung.com's move into generative AI. Building the review pipeline across 70+ markets is where I watched users lose track of which decisions were still their own, at a scale my own record could not reach.
Autonomy is not how often you leave the frame. Stay inside a good frame and that may be judgment. Bolt from it and that may be reflex. What matters is whether you can tell which one you just did, and whether you leave when the frame is wrong. The instrument for that is built. I haven't run it yet, because running it takes more than one person and a lab where the question is worth asking.