LLMs can write code, summarize research, and solve some complex problems. But they haven’t made a single breakthrough scientific discovery yet. Frontier labs like OpenAI and Anthropic say it’s coming soon. What if it doesn’t? What if the AI hype cycle ends and funding dries up before we get there? We might be stuck in a dangerous middle ground: AI that’s good enough to replace human thinking on most tasks, but not good enough to make the scientific breakthroughs we need most.

LLMs are trained on everything humanity has written and they return the consensus view of knowledge. When you ask an LLM about quantum mechanics, you get the textbook consensus, not Einstein’s revolutionary leap. Ask about evolution and you get today’s accepted theory, not Darwin’s original heretical idea. This is not a bug. It’s math. Every AI response is the weighted average of human writings.

But breakthroughs are anomalies. LLMs are consensus machines designed to give you the answer most humans would agree with. Take Darwin. He saw finch beaks and made a conceptual leap to natural selection. The scientific establishment rejected it for decades. The evidence was there all along, but the interpretive framework was revolutionary.

The feedback loop has already started. AI learns from human writing. Humans start using AI output. AI learns from that too. The unconventional ideas get filtered out, the average ideas get amplified. Each cycle makes everything more similar. This may kill contrarian thinking. When scientists use AI to generate hypotheses and design experiments, they may optimize for ideas that seem reasonable rather than revolutionary.

By 2035, we may have researchers who are extraordinarily skilled at working within existing frameworks but completely incapable of questioning them. The age of scientific revolution may stagnate. Not because we’ve solved everything. Because we’ve forgotten how to think like revolutionaries and how to recognize them.

The next Einstein may not emerge from AI-assisted research. They may emerge despite it. The question is: will we still be capable of recognizing them when they do?