Preface: AI and the question of cause

Key points

Large language models are trained to learn probabilities: how likely each next word is. That is seeing, the bottom rung of the ladder of causation.
Causal inference starts from the same probabilities and asks a different question: what happens if we change something?
It began on farm plots with Neyman and Fisher, and came together between 1974 and 2000 with Rubin, Holland, Robins and Pearl.
Judea Pearl grew Bayesian networks, maps of probabilities used in AI, into causal maps, and won the 2011 Turing Award.

Chapters

Key terms

Large language model: An AI system trained to learn probabilities: how likely each next word is, given the words before, across huge amounts of text. That teaches it what usually goes together, the bottom rung of the ladder of causation. It can repeat causal knowledge it has read, but on its own it can’t say what will happen if something is changed.
Ladder of causation: Judea Pearl’s three rungs of causal questions. Seeing: given what I’ve seen, what should I expect? Doing: what happens if I intervene? Imagining: what would have happened instead? Most AI learns on the bottom rung, and data from it can’t climb to the next without an experiment or a causal model.
Bayesian network: A map of probabilities: boxes for the things that vary and arrows for which depend on which, so a computer can work out how likely one thing is given another. Judea Pearl named and developed them for artificial intelligence in the 1980s, then turned them into the causal maps of this series by adding a rule for doing.
Rubin Causal Model: The potential-outcomes framework for causal effects: each unit has an outcome under each treatment, and the effect is the difference, though only one is ever seen. Jerzy Neyman wrote it down for farm plots in 1923; Donald Rubin extended it in 1974 to data nobody controlled, and Paul Holland gave it this name in 1986.
Structural causal model: A causal map in which each box is set by the boxes with arrows pointing into it, with no fixed formula: what statisticians call a non-parametric structural equation model. Judea Pearl linked these models to potential outcomes and gave them a rule for doing: to model a change, remove the arrows into the thing being changed.

Transcript

Ask an AI assistant a farming question, and you'll get a fluent answer in seconds. Today's large language models are trained to learn probabilities: how likely each next word is, given the words before. That's seeing: the bottom rung of the ladder of causation. The training learns patterns, not causes. Causal inference starts from the same probabilities, but asks: what happens if we change something?

That question began on farm plots. In 1923 Jerzy Neyman wrote down the yield each plot would give under each treatment, though only one could be seen. Ronald Fisher showed that letting chance choose each plot's treatment makes the comparison fair. So for experiments, cause has been part of statistics for a century.

But most farm data comes from decisions nobody controlled. In 1974 Donald Rubin carried the same 'what would have happened' thinking to data like that. In 1986 Paul Holland named it the Rubin Causal Model, and helped bring causal questions back into mainstream statistics. That year James Robins extended it to treatments that change over time.

Meanwhile, in AI, Judea Pearl had built Bayesian networks: maps of probabilities. In the 1990s he turned them into causal maps, with a rule for doing, not just seeing. By 2000 the pieces had come together: a modern synthesis. Pearl's work on probability and cause won the 2011 Turing Award.

Now each needs the other. Causal inference gives AI the 'what if', so it's foundational to AI we can trust to act. And machine learning lets causal inference cope with the thousands of measurements in real farm data. If AI is going to shape farming and food for decades, causal inference belongs in its foundations, not bolted on. This course uses it for one practical question: which farming practices change the nutrition in our food.

Questions

Do AI language models understand cause and effect?

Not on their own. They are trained to learn probabilities: how likely each next word is, given the words before. That captures what goes together, the bottom rung of the ladder of causation. They can repeat causal knowledge they have read, but saying what will happen if something is changed needs a causal model or an experiment.

When did causal inference become part of statistics?

For experiments, in the 1920s: in 1923 Jerzy Neyman wrote down the yield each farm plot would give under each treatment, and Ronald Fisher showed that letting chance allocate the treatments makes a fair comparison. For data nobody controlled, from 1974, when Donald Rubin extended the same thinking; Paul Holland named it the Rubin Causal Model in 1986.

What was the modern synthesis of causal inference?

The coming together, between 1974 and 2000, of potential outcomes and causal maps. Donald Rubin applied “what would have happened” thinking to data nobody controlled; James Robins extended it to treatments that change over time; and Judea Pearl, who had built Bayesian networks for AI, turned them into causal maps with a rule for doing rather than seeing.

Sources

Splawa-Neyman, J. (1990). On the application of probability theory to agricultural experiments. Essay on principles. Section 9 (translated by D.M. Dabrowska and T.P. Speed from Neyman's 1923 Polish essay). Statistical Science, 5(4), 465–472. doi:10.1214/ss/1177012031
Fisher, R.A. (1926). The arrangement of field experiments. Reprinted in Kotz, S. and Johnson, N.L. (eds) (1992), Breakthroughs in statistics (pp. 82–91). New York: Springer. doi:10.1007/978-1-4612-4380-9_8
Rubin, D.B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66(5), 688–701. doi:10.1037/h0037350
Holland, P.W. (1986). Statistics and causal inference. Journal of the American Statistical Association, 81(396), 945–960. doi:10.1080/01621459.1986.10478354
Robins, J. (1986). A new approach to causal inference in mortality studies with a sustained exposure period: application to control of the healthy worker survivor effect. Mathematical Modelling, 7(9–12), 1393–1512. doi:10.1016/0270-0255(86)90088-6
Pearl, J. (1988). Probabilistic reasoning in intelligent systems: networks of plausible inference. San Mateo, CA: Morgan Kaufmann. doi:10.1016/C2009-0-27609-4
ACM (2011). A.M. Turing Award: Judea Pearl, “for fundamental contributions to artificial intelligence through the development of a calculus for probabilistic and causal reasoning.” amturing.acm.org/award_winners/pearl_2658896.cfm
Pearl, J. (1995). Causal diagrams for empirical research. Biometrika, 82(4), 669–688. doi:10.1093/biomet/82.4.669
Pearl, J. (2000). Causality: models, reasoning and inference. Cambridge: Cambridge University Press (2nd edn 2009). bayes.cs.ucla.edu/BOOK-2K/
Emsley, R. (2013). What is… causal inference? Methods@Manchester “What is…” series, University of Manchester, 14 February 2013 (slides). hummedia.manchester.ac.uk/institutes/methods-manchester/docs/CausalInference.pdf

All 31 films · Glossary · The Vitagri Academy