Scientists May Finally Be Able to Read AI's Mind

Why do even the people who build AI struggle to control it? Because they don't fully know what happens inside it. They can see what goes in and what comes out, but the middle, the billions of connections inside it, has been a black box.

On October 15, 2026, Nick Cammarata, a former OpenAI researcher now at a company called Goodfire, published an essay titled Neural Networks Are Already Clean. His claim might be the most significant AI research of this year.

The old belief: AI is a messy junk drawer

An AI model is made of billions of tiny units called neurons. Each one switches on when it spots something. For years, researchers believed these neurons were a jumble, with each one mixing many unrelated ideas, like a drawer stuffed with keys, batteries and old receipts. In 2023, an OpenAI study found that only about 1.7% of neurons in a model could be clearly explained. The rest looked like noise. Cammarata worked on that early research himself, which makes his change of mind worth paying attention to.

The new finding: most neurons do one clear job

Using the latest AI tools to study other AI models neuron by neuron, Cammarata's team examined Qwen3-32B, a popular open AI model from the Chinese company Alibaba. They found that about 90% of its neurons do one clear job. Around 7% handle two closely related jobs, and only 2 to 4% remain a mystery.

The "mess" came from how researchers studied AI, not from the AI itself. Older tools showed only a few examples of what made a neuron switch on, so researchers named neurons by what they saw on the surface, like "cats" or "cars", and missed the shared pattern underneath. Older AI designs also made neurons harder to read than today's models.

His examples make this easy to see. One famous neuron switched on for both cat faces and sports cars, which seemed random. It turned out to be a V-shape detector. A cat's pointed ears and chin form a V, and so does a sloping car hood. Another neuron switched on for random computer code and Bengali text. Both contained the letters "PR". The neuron was simply a very good "PR" spotter.

The proof: switching off a memory

The team switched off a handful of neurons in a model. After that, the AI said Psalm 23, one of the best-known prayers in the Bible, was not a known text. When asked for the US Declaration of Independence, it made up a document about a country that does not exist. A few neurons held those memories, and removing them removed the knowledge.

Why bigger AI may be easier to read

You might expect the biggest AI models to be the hardest to understand. Cammarata suggests the opposite. A small model has little room, so each neuron has to cover many related ideas at once. A large model has more room, so each neuron can focus on one narrow job, much like a large hospital has specialists while a village clinic has one doctor who does everything. His team saw the same clean pattern in models with over a trillion parameters. If this holds, our ability to understand AI could grow as AI grows.

Why this matters to you

Today, we check if AI is safe by asking it questions and watching its answers. That is like judging a person's honesty only by what they say. If we can read neurons, we can look for the parts that cause lying, flattery or made-up facts, and fix them before the AI reaches your phone, your bank or your doctor. It is the difference between guessing and taking an X-ray.

The caution

Cammarata calls this a working idea, not a final answer. The findings are not peer-reviewed yet, and other researchers still need to test them. Still, the message is hopeful. AI may be less of a mystery than we feared, and that gives us a real chance to make it safer.