What would count as a scientific discovery by AI?

What would count as a scientific discovery by AI?

A proposed answer starts the work. A discovery also needs a question that matters, a test suited to the claim, and a result that changes what can be understood or done. Those requirements differ across the sciences.

When does a pattern become a proof?

Checking examples can suggest a mathematical statement. A proof must explain why it holds throughout its stated domain. The step from a convincing pattern to a general argument changes the standard of evidence.

What can a protein’s shape tell us?

A predicted structure can help explain molecular interactions and guide further work. A living system also depends on motion, partners and conditions. Confidence in a structure does not settle every claim about biological function.

Which observation should come next?

A model can help choose a candidate or an experiment. The observation then feeds back into the search. The useful question is what the next measurement can distinguish, including outcomes that would expose a mistake.

A test that could change our mind

Verification is specific to the claim. A proof checker tests a formal argument against given rules. Independent data can test a prediction. A physical experiment can reveal behavior missing from a model. Success at one task does not substitute for all the others.

What is the search rewarding?

Optimisation improves a score. The score represents a choice about what to reward, and it can leave out something that matters. A stronger search does not repair an inadequate objective by itself.

The next odd number completes the next square

An n by n square grows into an (n + 1) by (n + 1) square by adding a row and a column. Counting their shared corner once gives 2n + 1 new cells. Starting from one cell, this construction gives 1 + 3 + … + (2n − 1) = n² for every positive integer n.

A structure prediction tested against structures

AlphaFold’s 2021 research paper reports its structure-prediction method and evaluation. It provides a concrete place to examine what was predicted, what was compared, and which questions lie outside that comparison.

Primary sourceHighly accurate protein structure prediction with AlphaFold (Nature, 2021) ↗

What becomes possible after the result?

A result can make another question tractable, reveal a mechanism, or enable a useful intervention. Asking what follows from a result connects the technical achievement to its scientific consequence.