AI is helping scientists explore possibilities that are too numerous to examine one by one. It can predict a molecular structure, propose a new protein, search for a mathematical construction or suggest an experiment. The scientific advance comes from the collaboration: researchers define a worthwhile question, AI expands the search, and evidence determines which ideas survive.
This guide is for curious readers, students and researchers who want to understand that partnership. Follow the contribution of each participant, then ask what was actually checked. The examples below span research published from 2023 through 2026; they illustrate different methods, rather than a ranking of tools.
How the collaboration works
Scientific AI includes several kinds of systems. A specialized model can learn relationships between structures and properties. A generative model can propose candidate designs. A language model can help connect papers or write analysis code. An automated laboratory can connect a selection algorithm to instruments. These systems contribute at different points in a research project.
- Define the question. Scientists choose the phenomenon, constraints and evidence needed. “Find a useful material” becomes a measurable objective with a reference material and explicit test conditions.
- Represent what is known. Researchers assemble data, methods and prior results. They check units, provenance, missing observations and whether the examples represent the intended problem.
- Generate or prioritize possibilities. AI proposes structures, explanations, programs or candidate experiments. The team can ask for alternatives and examine where the model is uncertain.
- Test outside the proposal process. Experiments, held-out data, mathematical checks or other appropriate measurements evaluate the candidates. The evaluator needs its own scrutiny.
- Learn from the outcome. People interpret successes and failures, refine the question and decide what to try next. Keeping unsuccessful attempts makes the search easier to understand and repeat.
In active learning, a model helps choose the next observations that would be useful to obtain, then is updated with those results. A valuable experiment may reduce uncertainty even when it produces a disappointing candidate. The practical opportunity is a more informative cycle of research. Our research design guide explains controls, replication and fair comparisons.
Five ways AI is contributing to discovery
1. AlphaFold 3: turn molecular information into a structural hypothesis
The 2024 AlphaFold 3 study describes a model that predicts joint structures involving proteins, nucleic acids, small molecules and other molecular components. Its evaluations compare predictions with reference structures. For a scientist, a proposed arrangement can suggest which molecular contacts deserve investigation or help frame a structural question.
The division of work: the model supplies a structural prediction; researchers choose a biological question and assess its relevance with appropriate evidence. A confidence estimate describes the model’s assessment. Binding strength, biological activity and behavior under different conditions require their own measurements. A predicted complex is a starting point for those questions.
2. RFdiffusion: design proteins and test the designs
The 2023 RFdiffusion paper introduced a generative approach to protein backbones, including designs intended to bind other proteins. The researchers experimentally characterized selected designs. A cryo-electron microscopy structure of a designed binder attached to influenza haemagglutinin closely matched its design model.
The division of work: researchers specify a design problem, computation generates candidates, and laboratory work establishes properties of selected physical proteins. This is a concrete connection between a generated proposal and experimental evidence. Readers should still track the selection funnel: how many designs were generated, made, tested and successful. Agreement for one characterized design has a more specific meaning than success across every candidate or application.
3. FunSearch: search programs whose results can be checked
FunSearch, published online in December 2023, combined a language model with a systematic evaluator. It searched for programs producing mathematical constructions, including larger cap sets: collections of points satisfying a particular combinatorial restriction. The study also explored heuristics for online bin packing, where incoming items must be assigned to limited-capacity bins.
The division of work: people formulate the task and evaluator; the model proposes programs; computation checks their outputs; researchers interpret the resulting constructions. The cap-set contribution improved known constructions within the studied problem. It did not settle every aspect of the general problem. This example shows why a reliable, inexpensive checker can make a large search productive.
4. Co-Scientist: develop hypotheses with researchers
The 2026 Co-Scientist paper describes a Gemini-based system whose agents generate, critique and refine hypotheses around scientist-provided goals. Its biomedical evaluations included drug repurposing, target discovery and mechanisms related to antimicrobial resistance. The authors reported laboratory support for selected proposals, including in vitro tests of candidates and combinations for acute myeloid leukaemia.
The division of work: scientists steer the research and perform the empirical checks; the system broadens and revises the candidate explanations. Agreement between its agents is an internal ranking signal. Experimental evidence supplies a separate test. Results in cells establish a laboratory finding with defined conditions; patient benefit and clinical use require further evidence. The article’s methods and validation sections matter more than the assistant’s conversational fluency.
5. ERA: help build better tools for scientific analysis
The 2026 Empirical Research Assistance (ERA) study explored AI-generated scientific software. A language model and tree search iteratively improved programs against a measurable objective, incorporating research ideas. The reported applications included single-cell data analysis, geospatial analysis and neural activity prediction in zebrafish.
The division of work: researchers define the task, data and evaluation; AI searches implementations; scientists assess what the improved analysis means. Better software can enable discovery by making a previously difficult analysis more useful. A higher benchmark score establishes performance under that evaluation. Scientific interpretation also depends on data quality, a fair comparison and performance on genuinely new observations.
Match the claim to the evidence
A headline may compress several stages into the word “discovered.” Read the paper to identify the stage reached. The table is a reading framework: different fields use different tests, and a mathematical result may be established through proof or exact checking rather than a laboratory experiment.
On a narrow screen, focus the table and use the arrow keys to scroll.
| Reported result | Evidence to look for | Useful next question |
|---|---|---|
| A promising prediction | A stated model, inputs, uncertainty and comparison with relevant reference data. | Does it hold for cases outside the data used to develop or select the model? |
| A new candidate or hypothesis | An explicit proposal and a check against prior work. | What observation would distinguish it from competing explanations? |
| A computational advance | Inspectable methods, sound checks and evaluation separated from optimization. | Is the gain robust to new inputs, alternative baselines and realistic costs? |
| An experimental finding | Measurements, controls, sample counts, uncertainty and repeat trials. | Can another group reproduce the result, and under which conditions? |
| A useful application | Testing in the intended setting with relevant practical constraints. | Does performance survive scale, cost, reliability and deployment requirements? |
Watch for data leakage: information from the evaluation set entering training, prompts or candidate selection. Also check the denominator. Ten successful designs out of ten tested means something different when thousands were screened first. Comparisons should account for the entire process, including computation, human review and unsuccessful experiments.
For claims of faster discovery, identify the clock being measured. Seconds to generate a hypothesis, weeks to validate it and years to make an application dependable are distinct intervals. A fair speed comparison explains the task, baseline and resources included.
A worked example: choose the experiment that teaches you more
This example shows two useful goals: find a strong candidate and learn where the model’s assumptions fail. Researchers decide how to balance them. The basic research guide explains why deeper understanding and immediate product performance can lead to different experiments.
Make the partnership inspectable
Useful collaboration leaves a trail that another researcher can examine. Keep the research question, model and data versions, prompts where relevant, generated candidates, selection decisions, analysis code and measured outcomes together. Record when people changed a proposal and why. Our reproducible research workflow guide explains how to organize those materials.
Human responsibilities include choosing meaningful questions, judging whether the data fit them, designing discriminating tests and deciding how far a conclusion reaches. AI can help draft an alternative explanation or analysis, while a person remains responsible for checking citations, units, assumptions and results. More elaborate agent discussions can still repeat a shared error.
For a manageable first use, ask AI to help compare explanations for an already documented result or review an analysis against a known reference case. Specify the evidence it may use, ask it to expose assumptions, and check each consequential output. Use research data only in systems approved for that data and preserve the project’s confidentiality requirements.
Read the next discovery with five questions
- What question did the researchers ask? Write the scientific objective in one sentence.
- What did AI contribute? Identify the prediction, design, hypothesis, code or experiment choice.
- What did people establish? Find the measurements, mathematical checks or independent evaluation.
- How broad is the result? Record conditions, comparison methods, unsuccessful attempts and remaining uncertainty.
- What happens next? Look for replication, released code and data, corrections and testing in a different setting.
Download the editable AI-discovery reading record and use one copy per paper. Follow the primary papers linked above to their methods, supporting information and update notices. Revisit the claim when independent work tests it or the authors issue a correction. A university or company announcement can help locate a study; the study supplies the details needed to evaluate it.
AI’s contribution becomes clearest when the story preserves this chain: a human question, a computational suggestion, a test, and a revised understanding. That is how a faster search can become better science.
Continue with the Science guide collection or AI research, evaluation and standards in China.