Artificial intelligence can help scientists forecast the weather, study images, and predict the shapes of molecules. These are useful achievements, but a correct prediction does not always explain why something happens. The difference matters when deciding what a scientific result tells us.
Here, scientific understanding means having an explanation of how or why something happens, supported by evidence. That explanation helps us reason about related situations and ask further questions. Understanding can be partial: scientists may explain one part of a process while still investigating another.
Consider a pendulum. A simple model can estimate how long it takes to complete a swing while leaving out friction and air resistance. That may be enough to study the timing of its movement. To explain why the pendulum slows down and stops, however, those missing details become essential.
A scientific model is a simplified description of something in the world. It might be an equation, a computer program, or a physical object. Scientists use models to study problems that would otherwise be too complicated to handle.
Every model leaves something out. A model of a liquid might assume that it is evenly mixed. A model of a moving object might ignore friction. These choices can make the work easier, provided the missing details do not seriously affect the answer being sought.
The question therefore comes first. When the purpose of an investigation changes, details that once seemed unimportant may become central. Scientists need to choose a model that includes enough information for the question they are asking.
In a traditional mathematical model, some assumptions are easy to see. An equation might leave out a force or treat a changing quantity as constant. Yet the surrounding work also involves choices about instruments, measurements, and which observations to include.
Many AI systems learn by finding patterns in data, such as photographs or recorded measurements. Researchers choose the inputs, meaning the information given to the system, and decide what it should predict. Training adjusts the system to reduce errors in those predictions.
Those choices matter. The examples may cover only certain places, conditions, or kinds of samples. Measurements may contain errors. A system can learn the patterns present in its training data without learning everything needed for the intended scientific task.
Imagine that two kinds of biological samples are photographed with different microscopes. An AI system might learn to separate the images partly by recognizing differences between the instruments. It could achieve a high score while learning less about the samples themselves than researchers expected.
The recorded answer used for training matters too. A dataset may attach a label, meaning a category or recorded answer, to each example. Researchers therefore need to consider how well those labels represent the property they want to study.
Predicting something means estimating what will happen or what an observation will show. Explaining it means giving an account of how or why it happens. These tasks can support each other, but success at one does not automatically establish success at the other.
Tools for examining AI predictions can help researchers study a system's behavior. For example, they may highlight parts of an image that appear to influence its answer. Researchers still need to check whether the explanation accurately reflects how the system produced that answer.
An explanation also needs to be useful to the person reading it. A clear display can help someone follow a result, but researchers still need to establish whether it faithfully describes the computation. They also need to identify conditions in which the system's answers may be unreliable.
There is a further question about cause. In the microscope example, an instrument's image pattern may help the system sort samples without explaining their biological differences. Changing an input in the computer shows how the program responds. Establishing what causes a change in the samples themselves requires a different investigation.
A useful prediction can still move science forward. It may reveal an unexpected pattern, suggest an experiment, or make a difficult calculation possible. In their 2022 work on scientific understanding with AI, Mario Krenn and colleagues examine how these tools can also help researchers develop new ideas and concepts.
Proteins are molecules that perform many tasks in living cells, including speeding up chemical reactions and transporting substances. Their shapes help determine how they work. Predicting those shapes gives researchers a useful starting point for studying their behavior.
John Jumper and colleagues described AlphaFold2 in a 2021 research paper. In a blind assessment, it predicted many protein structures with accuracy close to structures determined experimentally. That achievement gave scientists a powerful way to investigate proteins whose shapes had not yet been established.
Consider a hypothetical experiment involving a protein that speeds up a chemical reaction. Its predicted shape might suggest which part helps it interact with another molecule. Scientists could alter that part and compare the changed protein's activity with the original. They would also need to check whether the alteration disrupted the protein more generally before drawing conclusions about that particular interaction.
The prediction helps researchers choose a specific question to test. This is one practical way that predicted structures can guide experiments. The laboratory work then helps establish whether the proposed explanation fits what happens.
The structure is only part of the picture. A predicted shape does not fully explain how a protein folds, how it moves, or everything it does in a living cell. Each of those questions concerns a different aspect of its behavior.
AlphaFold2 also provides a confidence score for different parts of a predicted structure. That score estimates how accurate the shape of each part is likely to be. It does not establish that a proposed explanation of the protein's behavior is correct.
Predictions can guide experiments, and experiments can help check predictions. If the results disagree, scientists need to investigate the reason. The prediction may be wrong, the experimental evidence may be uncertain, or the two methods may represent the protein under different conditions.
AlphaFold3 extended this work to predict arrangements involving proteins and other molecules. This helps researchers study how different molecules fit together. It predicts structures rather than simulating how the molecules move over time.
A test should match the claim being made. If researchers want to know how a system performs on new samples, they need samples that were kept apart from training and from the choices used to develop the system. Otherwise, the test may make the result look stronger than it is.
This is one common form of data leakage. Information from the test data influences the system before the final evaluation. That influence can enter through training, data preparation, or decisions about which model to use.
Suppose researchers take several photographs of each sample. If photographs of the same sample appear in both training and testing, the system has already encountered that sample. To check how well it sorts unfamiliar samples into the right categories, researchers should keep all photographs of each sample together in either the training data or the test data.
A system intended for use in another laboratory should be checked against data from that setting. A system intended for future use needs evaluation that considers changes over time. Researchers should examine where errors occur and whether a simpler method performs just as well, since an average score can hide important weaknesses.
If a pattern is presented as a discovery, researchers should check how much it depends on their choices. A result that disappears after a reasonable change in the data or analysis needs further investigation. These checks help distinguish a promising finding from a pattern that may be fragile.
Claims about cause need evidence about what happens when a condition changes. Experiments and comparison groups can help provide that evidence, as in the hypothetical protein experiment. Studies based on existing observations may also contribute, but their conclusions depend on whether the analysis adequately accounts for other possible explanations.
Repeating an analysis is another useful check. Running the same code can reveal whether the result can be reproduced, while collecting new observations can test whether it holds up with fresh evidence. Both approaches have limits if they repeat the same mistaken assumption.
AI often learns from information produced through earlier scientific work. Its training material may already combine observations with calculations from existing models. Examining how that material was produced helps explain what the learning system has available to draw on.
GraphCast, an AI weather forecasting system, was trained on a record of past atmospheric conditions called ERA5. That record combines observations with a numerical weather model. Understanding its origin helps researchers assess what the AI has learned and which features of earlier methods it may carry forward.
Some approaches also use equations from physics to guide learning. This can help rule out some implausible answers. The equations must suit the conditions being studied, while new observations can reveal where the chosen physical explanation needs improvement.
The availability of a tool can also influence which questions receive attention. A problem that AI can analyze quickly may seem more attractive than one that requires difficult measurements or lengthy experiments. Scientists need to keep important questions in view even when their tools cannot yet answer them easily.
A powerful tool may also encourage people to believe they understand more than the evidence supports. That is a risk in how a tool is used, rather than an unavoidable result of using AI. Researchers can address it by stating clearly what has been predicted, what has been explained, and what remains unknown.
AI can make scientific work faster and open new lines of inquiry. In the protein example, a predicted shape suggests where to look, an experiment tests what happens, and an explanation connects the findings to the protein's behavior. Those are distinct contributions that can work together. Recognizing what each contributes helps scientists turn a useful prediction into a better understanding of the world.
Roman Frigg and Stephan Hartmann, Models in Science, Stanford Encyclopedia of Philosophy.
P. Jonathon Phillips and colleagues, Four Principles of Explainable Artificial Intelligence, NIST, 2021.
Bernhard Schölkopf and colleagues, Toward Causal Representation Learning, Proceedings of the IEEE, 2021. Author preprint.
Mario Krenn and colleagues, On scientific understanding with artificial intelligence, Nature Reviews Physics, 2022. Author preprint.
John Jumper and colleagues, Highly accurate protein structure prediction with AlphaFold, Nature, 2021. Open full text.
EMBL-EBI, pLDDT: Understanding local confidence, AlphaFold training materials.
Thomas C. Terwilliger and colleagues, AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determination, Nature Methods, 2024; first published online November 30, 2023.
Josh Abramson and colleagues, Accurate structure prediction of biomolecular interactions with AlphaFold 3, Nature, 2024.
Sayash Kapoor and Arvind Narayanan, Leakage and the reproducibility crisis in machine-learning-based science, Patterns, 2023 Authors' research page.
Genevera I. Allen, Luqin Gan, and Lili Zheng, Interpretable Machine Learning for Discovery: Statistical Challenges and Opportunities, Annual Review of Statistics and Its Application, 2024.
Remi Lam and colleagues, Learning skillful medium-range global weather forecasting, Science, 2023. Author preprint.
George Em Karniadakis and colleagues, Physics-informed machine learning, Nature Reviews Physics, 2021.
Lisa Messeri and M. J. Crockett, Artificial intelligence and illusions of understanding in scientific research, Nature, 2024.
For corrections, permissions, privacy requests, or accessibility assistance, use the contact form or call +1 (650) 427-9358. For details, see Legal, Privacy and Accessibility.
Disclosure: AI tools assist with research, summaries, drafting, and editorial review. Their use varies by article and stage of the writing process. Ardan Michael Blum writes, revises, and updates the content, shapes its arguments and interpretations, and makes the final editorial decisions.
© 2026 Ardan Michael Blum.