In a 2001 paper, the statistician Leo Breiman described two cultures of working with data. One assumes a stochastic model of how the data arise and estimates its parameters, a practice he said described 98 percent of statisticians. The other builds an algorithm that predicts outcomes and judges it by accuracy on new data, a group he put at 2 percent of statisticians, many working in other fields. The split is a useful way to see what separates analysis meant to explain from systems built to predict, which is much of what readers mean when they set data science against AI.
Key takeaways
- Breiman named two goals of data analysis, prediction and information, and two cultures: data modeling and algorithmic modeling.
- Data models are checked with goodness-of-fit tests and residuals, while algorithmic models are checked by predictive accuracy.
- In his chlorine project, a decision tree reached 95 percent accuracy on a 5,000-compound test set; in his ozone project, the predictor failed because its false alarm rate was too high.
- Many different models can fit equally well yet tell different stories about which variables matter, which limits what a model can explain.
- The paper is from 2001 and never uses the label AI, so mapping it onto current job titles is an interpretation.
The two goals and the two cultures
Breiman, then a professor of statistics at the University of California, Berkeley, pictured nature as a black box that turns input variables x into responses y. He listed two goals for analysis. Prediction means saying what the responses will be for future inputs. Information means learning how nature associates the responses with the inputs.
The data modeling culture fills the box with an assumed formula, such as linear regression, logistic regression, or the Cox model, then estimates parameters from data. The algorithmic modeling culture treats the inside of the box as complex and unknown and looks for any function f(x) that predicts y well, using tools such as decision trees and neural nets. His complaint was that the first culture had crowded out the second, leading to what he called irrelevant theory and questionable conclusions. The paper in Statistical Science, volume 16, issue 3, pages 199 to 231, is open access.
| Feature | Data modeling | Algorithmic modeling |
|---|---|---|
| Starting assumption | A stochastic model generated the data | The mechanism is unknown |
| Typical tools | Linear and logistic regression, Cox model | Decision trees, neural nets |
| How it is validated | Goodness-of-fit tests, residual examination | Predictive accuracy on held-out data |
| Breiman's estimate of its share | 98 percent of statisticians | 2 percent, many in other fields |
A prediction problem that worked: chlorine in mass spectra
Breiman's consulting work gave him two contrasting projects. In the mid-1970s the Environmental Protection Agency wanted to know whether the presence of chlorine in a compound could be predicted from its mass spectrum, because chemists could not examine every spectrum by hand. The database held known structures and spectra for 30,000 compounds. It was split at random into a 25,000-compound training set and a 5,000-compound test set.
Linear and quadratic discriminant analysis were hard to adapt to spectra of varying length. A decision tree built on 1,500 yes-no questions, designed from knowledge of how chlorine shows up in spectra, reached 95 percent accuracy on both chlorine-containing and chlorine-free compounds. The model was judged on data it had not seen, and nobody needed it to explain chemistry.
A prediction problem that failed: next-day ozone
The other project aimed to predict Los Angeles Basin ozone levels 12 hours ahead so that alerts could be issued earlier. The data covered more than 450 meteorological variables over seven years, with the first five years used for training and the last two held out for testing. Large linear regressions with variable selection, then added quadratic terms and interactions, were the methods of the time. Breiman wrote that the project was a failure because the false alarm rate of the final predictor was too high.
The account is useful because it shows the test-set discipline cutting both ways. Held-out data revealed that the model did not work, which a good in-sample fit would have hidden.
Why a good model may not explain much
Breiman called the next problem the Rashomon effect. With 30 variables, about 140,000 five-variable subsets compete in a linear regression, and many of them come within 1.0 percent of the lowest error. Each can tell a different story about which variables matter. Picking the single best equation and reading it as the truth about nature is therefore risky. The same multiplicity appears for decision trees and neural nets.
For anyone reading a feature-importance chart or a coefficient table, the point stands: a model that predicts well supports a claim about prediction, and a claim about cause or mechanism needs more.
An illustrative example: missed clinic appointments
This example is illustrative, not drawn from a real clinic. Suppose a clinic wants to reduce missed appointments.
- Information goal: does distance from the clinic change the chance of a miss? An analyst would fit a simple model, report an estimate with its uncertainty, and discuss whether other factors could explain it.
- Prediction goal: which patients should get an extra reminder call? A team would train a model on past records, test it on later months, and care mostly about how often its flags are right.
The same data could serve both, but the outputs and the checks differ, and a team that mixes the goals may read a prediction tool as proof of a cause.
Questions that identify which culture a task belongs to
- Is the deliverable an explanation a person will act on, or a function that will run on new cases?
- Is the check a fit test on the data at hand, or accuracy on data held out?
- Would a different model with equal accuracy change the story told?
- Are the future cases like the past ones, or has the world changed since the data were collected?
FAQ
Does Breiman's paper say data science and AI are different fields?
No. The paper is about statistics and does not use the term AI. The split between explaining and predicting is a framework that readers can apply to the two labels, and it does not settle how institutions define them.
Are the 98 percent and 2 percent figures measured?
They appear as estimated culture populations in the paper, which reads as the author's own estimate of the situation in 2001, and no survey is cited for them. They are not current numbers.
Is prediction always better than explanation?
The paper argues for using both tools, not for dropping one. Prediction suits cases where a reliable forecast is the goal, while explanation suits cases where a decision depends on understanding a mechanism.
Has the field moved on since 2001?
Yes. David Donoho's 2017 review credits Breiman with calling for an emphasis on prediction over inference, and later data science programs combine statistics, machine learning, and computing. The two cultures remain a way to ask what a project is for.
Related reading
- AI in scientific research
- Designing AI systems
- Alternatives to artificial intelligence
- How AI affects modern society
More in AI Basics.
Last verified: October 2026
Sources
- Statistical Modeling: The Two Cultures, Breiman L, Statistical Science 16(3), 2001
- 50 Years of Data Science, Donoho D, Journal of Computational and Graphical Statistics 26(4), 2017