Fluent Text Is Not One Step from General Intelligence

A model writes smooth prose. It debugs code. Someone concludes: general intelligence must be one step away. Michael I. Jordan, drawing on an older idea from Hubert Dreyfus—the first-step fallacy—says that leap is the mistake. Doing a few dazzling tasks well is not the same as being close to doing everything.

What the first-step fallacy looks like

The pattern is familiar. Early AI could play narrow games or solve tidy puzzles; enthusiasts treated each win as proof the finish line was near. Today the demos are larger—chat, code, images—but the logic is the same: impressive first steps get read as almost-arrival.

Jordan’s blunt description of current systems: large statistical boxes. They map inputs to outputs. That can be enormously useful. It is still not system-level thinking about an open world of goals, institutions, and trade-offs.

Fluency tricks the eye

Language that sounds human is especially persuasive. If a system explains itself in polished paragraphs, people grant it intentions, plans, even a kind of mind. Jordan has long resisted that anthropomorphism. Smooth text is evidence of strong pattern matching on language—not proof that the machine shares human understanding of work, risk, or society.

That matters for product judgment. A chatbot that feels like a colleague can still fail the moment the task leaves the training distribution—or the moment “good enough prose” is mistaken for “good enough decision.”

Engineering fields have theory; hype often skips it

Jordan contrasts today’s AI culture with electrical, chemical, and mechanical engineering. Those fields rest on deep theory—Maxwell, Newton, and the practice of engineers who know constraints. Much AI talk, he argues, looks more like clever programmers stacking data and compute by intuition, while the public conversation drifts toward science fiction.

• A strong demo shows a capability, not a finish line
• Input–output mapping can be useful without being “general mind”
• Clear limits are part of serious engineering, not a lack of ambition

Keep the capability; drop the mirage

None of this asks readers to dismiss large models. Jordan himself notes that neural scaling properties are real and likely to last. The ask is narrower: describe what the statistical box does, build the surrounding systems that make it safe and useful, and stop treating fluency as a ticket to AGI. Naming the first-step fallacy is a way to protect judgment—especially for younger researchers told that either everything is already solved or nothing they do matters.