By Evan Vega
For years, the people who build artificial intelligence have admitted something unsettling: they can’t fully say what goes on inside it. An AI model is billions of tiny parts and the numbers connecting them, and the working belief has been that those parts are a tangled mess, each one holding a jumble of unrelated jobs. That’s why AI gets called a black box.
This week one of the researchers who helped build that belief said the evidence for it was misread. In an essay published on 15 October 2026 with the AI lab Goodfire, Nick Cammarata argues that most of an AI’s inner parts can be read after all. If he’s right, it changes how AI can be checked for safety, how it can be fixed, and how much anyone can trust what it says. What follows is what it means in plain terms, and where to be careful.
For years AI researchers said the inside of an AI model is an unreadable black box. Now Nick Cammarata, writing with the AI lab Goodfire, says most of an AI’s inner parts can be read after all. A plai
Safety. Today AI is mostly judged by what it says, and a model that has learned to mislead could say all the right things. Cammarata argues that a model’s own wiring has no reason to hide anything, so reading the inside could catch what the answers don’t show.
Fixing AI. When a model picks up a bad habit, the usual response is more training and hoping it sticks. If the parts can be read, a bad habit can be traced to where it came from and changed with small, careful fixes. For anyone asked to explain an AI’s decision to a customer, a regulator or a court, that’s a very different conversation.
The web. One part the researchers examined responds to specific signs on pages of health SEO spam, and the essay notes that a well-known research page opened with a giant embedded image that AI reading tools ran out of room inside. How a page is built shapes what AI takes from it.
Picture a filing cabinet where every drawer holds a random jumble: tax returns, a recipe and a love letter in one drawer, because the cabinet ran out of room. That was the picture of an AI’s insides, and it’s why researchers built second AI systems just to untangle the first one. Cammarata says the drawers are mostly labelled. In one free, open AI model, he estimates about nine in ten of the parts he checked have one main job. He calls it a working hypothesis and puts himself at about 85% sure.
His best-known example is a part inside an image AI that lit up for cat faces and for the fronts of cars. For years it was cited as proof that AI is a mess. Cammarata, who helped write some of that research, now says it was looking for a V shape all along: pointed cat faces, the V-shaped front of a car, V-shaped flames. Another part, written off by an automated tool as junk, turned out to track Colorado. When researchers turned it up, the AI started planning a hike.
The pushback is real. Google DeepMind’s Neel Nanda called the work interesting but not a big update, and researcher Stephen Casper pointed out that knowing what most parts do isn’t the same as knowing what the whole machine does. Goodfire raised $150 million on the idea that AI can be read, and the detailed paper isn’t out yet.
So: the black box may be less black than we’ve been told. Ignore headlines that say AI is “solved” or the danger is over; the author doesn’t say that. And watch for the paper, and for someone without a financial stake to check the work. The model he studied is free to download, so they can.
Every number, every source and the full reaction are at alreadyclean.novcog.us.com, including the cat-and-car story, what the experts said and what isn’t proven yet. The essay: Neural Networks Are Already Clean (Goodfire, 15 October 2026).
Part of the Frontier Watch Series: Read the previous investigation
More Coverage:
→ Read this investigation on North Denver Tribune
→ Coverage from Daily Colorado News
