When will AI make a real scientific discovery?

Question answering is not discovery

An AI model passing scientific question-answering benchmarks demonstrates knowledge and reasoning under test conditions; it does not establish a new natural phenomenon. Cortist separates solving existing problems, generating hypotheses and independently reproducing a novel empirical discovery. Each stage requires different evidence.

What Stanford's 2026 science report indicates

The Stanford AI Index documents fast-growing scientific AI publications and strong results on some chemistry questions. Yet research replication and end-to-end scientific tasks remain difficult: leading agents are markedly below expert baselines on several research workflows. That gap is central to this forecast's uncertainty.

How an autonomous research loop works

Scientific discovery may involve literature mapping, hypothesis formation, experiment selection, data acquisition, interpretation and replication. For the proposed F-006 outcome, the AI must perform the substantive hypothesis and analysis work. Humans can provide safety approval, laboratory equipment and execution, but their intellectual and operational contributions must be recorded rather than hidden.

Why independent reproduction matters

Peer review, patents, a news release or a published paper can be useful indicators but are not sufficient to resolve this question. Independent researchers need to repeat the central empirical result, not merely say that the AI paper sounds plausible. Publication bias, contamination from existing work and experimental leakage require explicit review.

Examples of promising directions

AI-guided scientific discovery spans molecular design, material science, physics, astronomy and environmental science. The Stanford AI Index documents work on AI co-scientists and automated papers, but a fully generated workshop paper is not automatically a replicated scientific breakthrough. We will distinguish the provenance and result of each project.

What could speed or slow progress?

Better laboratory interfaces, agent planning, scientific world models and high-quality instrument feedback could accelerate the research loop. Weak experimental robustness, reproducibility failures, unsuitable benchmarks and limited access to equipment could delay a qualifying case. The forecast is about independently reproduced scientific knowledge, not a count of AI-written publications.

The 40% draft and timeline

The provisional F-006 estimate is 40% by 2035. The year slider displays subjective cumulative scenario anchors, not a validated time-to-discovery model. When credible replication evidence appears, Cortist must retain this first version and explain any future probability adjustment with specific primary sources.

Independent event review before scoring

The meaning of predominantly autonomous research, acceptable human roles and scientific novelty must be prospectively approved. A new forecast ID would be required if future reviewers materially changed the question. Until then the displayed value remains an unscored editorial draft.

How evidence will change this forecast

Cortist monitors original datasets, official measurements and independent evaluations rather than treating manufacturer, researcher or policymaker claims as settled outcomes. Any material reassessment must disclose the previous value, the new evidence, counterarguments and the reason for changing the subjective assessment. The older snapshot will remain at its original versioned URL.

Before this question becomes eligible for formal scorekeeping, independent reviewers must agree on the boundary of the event, adjudication rules, data sources and any necessary fallback measurement. If an essential definition changes, a fresh forecast question will be published instead of rewriting an old one.