To Discover Something New, AI May First Have to Forget
We keep being told AI will accelerate discovery. A new cosmology study points to an uncomfortable twist: the smarter an AI is about what we already know, the more likely it is to overlook what we don't.
The paper — "Transfer Learning Beyond the Standard Model," published in the Journal of Cosmology and Astroparticle Physics in June 2026 — tackled one of physics' most expensive problems. To test theories of the universe, cosmologists run enormous numbers of simulated universes, each one computationally costly. The researchers asked whether "transfer learning," the same trick behind modern AI, could cut that bill: train a neural network first on the well-understood standard model of cosmology (called ΛCDM), then fine-tune it on more exotic physics.
It worked, spectacularly. In some cases the shortcut reduced the number of expensive simulations needed by more than a factor of ten. Then it backfired. When a signature of new physics — the mass of ghostly particles called neutrinos — closely resembled a pattern the model had already learned from the standard model, the network forced the new effect into the familiar mold. The team calls this "negative transfer." Their own summary is blunt: pre-training "can accelerate inference, but may also hinder learning new physics."
Here the views split. Optimists see teething trouble. The cost savings are real, the failure mode is now documented, and the authors already found architectural fixes that blunt it. Every powerful tool has quirks you learn to engineer around. Skeptics counter that the problem is structural, not incidental. "The negative transfer is not random. It is driven by underlying physical degeneracies in the model," says co-author Veena Krishnaraj. An AI that learns from the known is, by construction, biased toward the known — and the rarest, most valuable discoveries are precisely the ones that don't fit.
That tension reaches far beyond cosmology, because this is the architecture behind almost every AI you already touch. The chatbot drafting your email, the model screening a stack of CVs, the system flagging an unusual medical scan — all are transfer-learning machines that pattern-match the present against the past. They are brilliant on the familiar and quietly weak on the genuinely novel. As these systems mediate more of what gets hired, funded, diagnosed, and discovered, the same pull applies: toward the average, away from the new. The risk isn't that AI gets things wrong loudly; it's that it nudges us toward the conventional so smoothly we don't notice.
The researchers' own word for the fix is the takeaway: to discover, an AI sometimes needs to "unlearn." The next leap in AI for science may come less from models that know more, and more from models that know when to set their knowledge aside.