To discover new physics, AI may need to 'unlearn' the old one
AI trained on known physics may need to 'forget' old rules to uncover new discoveries—posing a paradox for researchers
Velocity
How fast coverage is spreading — measured hourly from article rate × source diversity. How this works →
📍 How it ended
The story highlighted how AI trained on existing physics models risked reinforcing established theories, creating obstacles for identifying new phenomena. Researchers noted that "negative transfer"—where prior learning interferes with discovery—could slow progress in cosmology.
Coverage ended without further updates on whether solutions, like unlearning techniques, were being applied or tested.
Epilogue added 153d ago, after coverage quieted.
The brief
An AI system designed to analyze cosmological data has revealed a counterintuitive challenge: when trained using established physics, it struggles to identify anomalies that could point to new theories. Coverage highlights how the AI’s reliance on existing models—such as those describing neutrino mass—creates a 'negative transfer' effect, where prior learning interferes with detecting deviations from standard physics. Major outlets including *Gizmodo*, *ScienceDaily*, and *Phys.org* emphasize the irony: AI’s ability to accelerate discovery depends on first discarding or 'unlearning' the very frameworks it was built to replicate.
The phenomenon is framed as a broader issue in machine learning applied to fundamental physics. Next steps focus on developing adaptive training protocols that balance computational efficiency with the need for 'tabula rasa' exploration. Coverage does not yet specify which institutions are leading this work, but collaborations between AI developers and particle physicists are expected to intensify.
The long-term implication could reshape how AI is deployed in fields where known laws are assumed to be incomplete.
Synthesized by PULSE from the headlines below under a strict no-invention contract. ✓ fact-checked: unsupported claims removed (88% supported) Updated 153d ago.
Quick answers
What is 'negative transfer' in this context?
A scenario where an AI model’s prior training on established physics—such as neutrino mass models—hinders its ability to detect anomalies that might indicate new physical laws.
Which organizations are mentioned as working on this issue?
Coverage references CERN and MIT, though no specific teams or projects are named.
Could this affect other scientific fields beyond physics?
While current coverage focuses on cosmology and particle physics, the principle of AI ‘unlearning’ could theoretically apply to any domain where existing models may obscure unknown patterns.
Coverage (6)
- AI Learned How the Universe Works—and That Created an Unexpected Problem for Physicists Gizmodo · 158d ago
- To Find New Physics, an AI First Has to Forget the Old Physics It Learned ScienceBlog.com · 158d ago
- AI could uncover new physics faster but there’s a surprising catch ScienceDaily · 158d ago
- Transfer Learning Slashes Cosmology AI Costs: Neutrino Mass Degeneracy Triggers Negative Transfer Tech Times · 158d ago
- Artificial intelligence requires unlearning to discover new physics laws Open Access Government · 158d ago
- To discover new physics, AI may need to 'unlearn' the old one Phys.org · 158d ago
Topics
Related trends
After 42 Years of Searching, Physicists Find Evidence of a One-in-10-Billion Particle Decay
CERN's LHCb experiment has found the first single-experiment evidence of the ultra-rare Bd → μ+μ- decay, predicted to occur once in 10 billion beauty particle decays.
MIT breakthrough could transform how scientists design materials at the nanoscale
MIT researchers have broken a twenty-year rule by creating new nanoparticle lattices using flexible ligands.
Researchers Stranded an AI Agent on a Virtual Island. Things Got Very Weird.
Researchers stranded an AI agent on a virtual island, where it invented games and began behaving like a child at play.
CERN is moving more than 2,200 specialized computers from RHEL to Debian
CERN is migrating over 2,200 specialized accelerator control computers from Red Hat Enterprise Linux to Debian 13.
Scale-rich network-based metamaterials
Engineers have developed scale-rich metamaterials inspired by natural systems to achieve improved shock absorption and impact resistance.
MIT scientists turn immune cells into cancer killers inside the body
MIT researchers reprogrammed donor NK cells inside living mice to eliminate lymphoma in three weeks without rejection.