In 1958, Frank Rosenblatt stood before cameras at the Cornell Aeronautical Laboratory and made a claim that would echo for decades: his Perceptron, a machine roughly the size of a wardrobe, could learn to recognize images the way a human brain does. The press ran with it. Headlines promised a machine that would soon walk, talk, and be conscious of its existence. Within ten years, Marvin Minsky and Seymour Papert had published Perceptrons, a book that mathematically demonstrated the device’s severe limitations — it couldn’t even solve XOR, a trivially simple logical problem. Funding evaporated. The first AI winter began.

The standard lesson people draw from this episode is about hype: promise less, deliver more. But that reading is too comfortable, and it misses what actually went wrong and why it still matters now.

The deeper problem wasn’t Rosenblatt’s enthusiasm. It was a category error baked into the framing from the very start. The Perceptron was presented — and widely understood — as a model of biological cognition. It wasn’t. It was a linear classifier with a clever learning rule. The gap between “inspired by neurons” and “works like a brain” was treated as negligible, a gap in degree rather than kind. When the mathematical shortcomings became undeniable, the entire enterprise of neural computation got tarred with the same brush. Researchers threw out a genuinely powerful computational idea because it had been sold as something it was never equipped to be.

That conflation of mechanism and metaphor is not a relic. It is the structural flaw that resurfaces in every generation of AI progress. When deep learning arrived decisively in the early 2010s — ImageNet, AlexNet, the whole cascade — the framing again reached for biological analogy. “Hierarchical feature learning, just like the visual cortex.” Technically defensible in a narrow sense. Strategically reckless in every other sense. Because the moment you anchor a technology to a biological claim, you invite evaluation on biological terms, and no artificial system built so far survives that evaluation intact.

Today’s large language models carry the same inherited debt. Describing them as systems that “understand” or “reason” borrows cognitive vocabulary that implies much more than the underlying mechanism — autoregressive next-token prediction — actually does. This isn’t purely semantic fussiness. It shapes research priorities, regulatory frameworks, and public risk assessment in concrete ways. If a model “understands,” then failures feel like aberrations. If a model is a sophisticated statistical pattern-matcher, failures feel like the expected behavior of a known system operating near its limits. These are not equivalent stances. One produces surprise; the other produces engineering.

Rosenblatt himself was not a fool or a fraud. His actual technical contribution — the perceptron learning theorem — is real and foundational. The problem was that the machine got embedded in a story that outran it. Minsky and Papert’s demolition job, while mathematically precise, also benefited from the overclaim: it’s much easier to disprove a brain than to disprove a classifier.

What the Perceptron episode should have taught us, and manifestly did not, is that biological metaphor in AI is not neutral description — it is a bet. It raises the stakes, compresses the timeline for expected results, and makes failure feel more total than it is. Each time we reach for the vocabulary of cognition, consciousness, or understanding to describe systems built from gradient descent and matrix multiplication, we are making Rosenblatt’s press conference mistake again, just with better hardware.

The Perceptron worked. The story told about it didn’t. We’re still sorting out which lesson to learn from that.