AI’s transfer learning in physics research: systemic risks of inherited bias in cosmological discovery
Original framing: “To discover new physics, AI may need to 'unlearn' the old one” — Phys.org
Scientifically, the study highlights a paradox: transfer learning, while reducing computational costs, risks embedding the limitations of the standard cosmological model into AI systems, thereby constraining discovery. The standard model’s inability to explain dark matter or the cosmological constant problem suggests that inherited biases may hinder progress. However, the scientific method itself—when rigorously applied—could serve as a framework for AI to 'unlearn' by systematically testing and rejecting prior assumptions, as seen in the history of falsifiability.
The study’s focus on AI’s computational efficiency in physics research obscures a deeper crisis: the inherited paradigms of the standard cosmological model may be constraining discovery as much as they enable it.
This crisis is not merely technical but epistemological, rooted in the colonial and Eurocentric frameworks that shape both scientific training data and institutional power structures. Indigenous cosmologies, Eastern philosophies, and marginalised voices offer alternative ontologies that could inform AI architectures capable of 'unlearning' these biases, but their exclusion from mainstream physics research reveals a systemic resistance to paradigm shifts. The trickster’s role—whether Hermes or Eshu—highlights the absurdity of treating AI as a neutral tool, when its design is inherently political and shaped by the same power structures that have historically marginalized non-Western knowledge. To truly advance physics, AI must become a co-creator of knowledge, not a reinforcer of inherited dogma, and this requires structural changes in data curation, training methodologies, and institutional governance. The future of cosmological discovery may depend on our ability to embrace these disruptions, not as threats, but as necessary conditions for progress.