From Alchemy to Science—and from Science to AI?

Essay 2 of 4 — AI and the Future of Science. The first essay asked whether AI might turn research into a new alchemy. Here is the other possibility: perhaps the same technology will help reorganize science itself.

It is tempting to draw a straight historical line: alchemy gave way to modern chemistry, and human-led science will now give way to AI-driven discovery. In this story, the decisive change is that machines search larger hypothesis spaces, conduct experiments, and write papers faster than people.

But history does not arrange itself so neatly. Alchemy included experimental practices, skilled observation, and useful techniques. Modern science inherited some of them while developing new norms for evidence, explanation, and public scrutiny. Likewise, a machine performing a research task is not automatically evidence of a new scientific paradigm.

The stronger question is this: what would have to change in the production and justification of knowledge before AI-based research counts as a different kind of science?

Three kinds of change, not one

AI can change science at three distinct levels.

At the first level, AI-assisted science, people choose questions and judge conclusions while AI accelerates literature search, programming, and analysis. This is an important change in productivity, not necessarily in the rules for establishing a claim.

At the second level, AI-driven science, agents propose hypotheses, select experiments, learn from negative results, and recommend follow-ups. Research labor is reorganized; accountability and independent validation still matter.

At the third, hypothetical level, AI-native science, an AI system generates concepts and models humans would be unlikely to formulate. Some may be difficult to explain in familiar terms but still yield predictions that survive independent challenges. This level raises questions about the standards by which a scientific community recognizes knowledge.

Preparing diagram.

These levels need not arrive in order, and different research fields may occupy different levels at the same time. Nor is the third inherently better. A transparent, human-led experiment may produce more reliable knowledge than an autonomous but poorly checked pipeline.

What a paradigm shift would actually mean

Thomas Kuhn’s account of scientific revolutions is often summarized as a sudden replacement of one worldview by another. His historical argument is more demanding: scientists share exemplary problems, ways of working, values, and standards for judging results. A revolution can reorganize that shared practice. [1]

We could therefore look for changes in three places.

Questions. Does AI make it feasible to pursue problems that were previously beyond the community’s conceptual or practical reach?

Methods. Do cycles of machine-generated proposals, experimental criticism, and revision become recognized forms of scientific inquiry—with auditable records of why each experiment was selected?

Standards. Do researchers accept findings that cannot yet be expressed as satisfying human-readable theories, provided that independent experiments establish their predictive and intervention value?

Preparing diagram.

If only the first two change, we might be looking at a new instrument or division of labor. For a stronger paradigm claim, we would need evidence of changes in the third and of sustained adoption by a scientific community.

What current systems demonstrate

The Robot Scientist Adam, reported in 2009, already automated aspects of generating and testing scientific hypotheses. [2] This is a useful reminder that automation alone is not new.

A 2023 review in Nature surveyed how AI supports scientific inquiry across modeling, experimentation, and discovery. [3] The 2026 study of the AI Scientist went further: the authors built systems that generated research ideas, executed computational experiments, and wrote manuscripts. In a deliberately controlled workshop-review exercise, one of three submitted AI-generated papers met a workshop acceptance threshold, while the authors explicitly identified methodological and implementation limitations. That does not establish general scientific reliability. [4]

The 2026 Co-Scientist study showed another configuration: multiple agents generated, criticized, and refined hypotheses, with experimental work reported in selected biomedical cases. The authors themselves described the results as initial validation rather than proof that automated discovery needs no human scrutiny. [5]

These examples support a significant shift in what AI systems can attempt. They do not, on their own, show that the norms of science have undergone a Kuhnian revolution.

The central objection

Perhaps calling this a new science is unnecessary. Astronomy depends on instruments; statistics uses models whose behavior is not always intuitive; modern laboratories already distribute knowledge across many specialists. Science has never required a single person to understand every stage of a result.

That objection has force. We should not confuse what humans cannot immediately explain with what science cannot evaluate. Nor should we claim that prediction has suddenly replaced explanation: scientists have debated their relationship for generations.

A stronger case for a paradigm shift would show that AI opens durable new forms of inquiry and that scientific communities establish new, trustworthy ways of accepting or rejecting their products.

An uncertain threshold

I would not declare that science has already become AI-native. My narrower proposition is that the critical frontier lies between the automated generation of discoveries and a socially and technically reliable system for justifying them.

This also explains the seeming paradox from the first essay. The same tools can create a new alchemy—plausible claims without adequate tests—or a new science in which difficult-to-imagine claims are rigorously examined.

The difference is not the presence of AI. It is what happens when an attractive hypothesis meets the possibility of being wrong.

The next essay asks what happens when the supply of hypotheses grows faster than our ability to test them.

---

References

1. Thomas Kuhn. Stanford Encyclopedia of Philosophy, revised 2025.
2. King, R. D., et al. (2009). The automation of science. Science, 324, 85–89.
3. Wang, H., et al. (2023). Scientific discovery in the age of artificial intelligence. Nature, 620, 47–60.
4. Lu, C., et al. (2026). Towards end-to-end automation of AI research. Nature, 651, 914–919.
5. Accelerating scientific discovery with Co-Scientist (2026). Nature, 655, 487–496.

This is an exploratory essay, not a peer-reviewed philosophical thesis. Its three-level framework and paradigm criteria are proposals. Generative AI assisted in developing and editing this text; all such proposals remain open to criticism and revision.