A Counterfactual Is Not an Alternate Reality

After an outcome, the mind produces alternatives with suspicious ease. If the meeting had started earlier, if the treatment had been different, if the warning had been heard, if the model had chosen the other action—then the result would have changed. Sometimes that statement is well supported. Sometimes it is a story assembled from hindsight. In neither case have we observed a second reality.

A counterfactual asks what would have happened if a specified part of the actual situation had been different while other relevant features followed an explicit model. It is one of the most useful forms of reasoning available. Science uses it to define causal effects. Courts use it when considering responsibility. Engineers use it in failure analysis. Ordinary people use it every time they ask whether a different choice would have mattered.

A counterfactual is a comparison, not a place

The grammar can mislead. “In the world where the driver braked earlier” sounds as though researchers have located another complete universe and inspected its crash report. Usually the phrase is shorthand. It identifies a comparison case produced by assumptions about how the relevant system works.

In potential-outcomes language, a unit can have an outcome under treatment and an outcome without treatment. For the same unit at the same moment, only one is observed. This is often called the fundamental problem of causal inference. The missing outcome must be estimated from experiments, comparable groups, repeated structure or a causal model. The counterfactual is therefore not free evidence. It is the quantity the design is trying to identify.

Structural causal models express the same discipline differently. They describe relationships among variables, then represent an intervention by changing part of that structure and calculating what follows. The result can be powerful because the changed condition is explicit. It remains conditional on the model, the data and the assumptions that connect them.

Association cannot answer every “what if”

If people who take an action tend to have a better outcome, the action may help. It may also be selected by people who differ in ways that matter. Healthier patients may follow a treatment more consistently. Better-resourced schools may adopt a program earlier. Careful boaters may buy more safety equipment and also make more conservative decisions. The observed association mixes the possible effect with selection and context.

A randomized experiment can make groups comparable on average before treatment, which strengthens a causal contrast. It still does not reveal both outcomes for the same individual. Nor does randomization automatically solve noncompliance, missing data, measurement error, limited samples or transfer to a different population.

Observational studies can support causal claims when their design and assumptions are adequate. The crucial move is to name what must be true: which confounders were measured, how treatment is defined, whether compared cases could realistically receive either condition and whether the measured outcome corresponds to the question. Saying “the model controlled for everything” is not an assumption list.

Prediction, intervention and counterfactual are different jobs

A predictive question asks what is likely given what has been observed. An interventional question asks what is likely if an action is deliberately imposed. A counterfactual question asks what would have happened to a particular case under a different action than the one actually taken. These questions can use some of the same data while requiring different information.

A weather forecast can predict rain without showing that carrying an umbrella causes rain. A model can estimate the effect of offering a training program without determining whether one particular worker would have succeeded without it. A system that performs well at prediction does not automatically possess the causal structure needed for reliable intervention or individual counterfactuals.

This distinction matters for AI explanations. A model may generate a fluent story about why an output changed after one input was edited. That is not necessarily evidence about the internal mechanism that produced the original answer. An AI Explanation Is Not the Mechanism makes the same point at the model level: a plausible account and a faithful causal explanation are separate achievements.

Possible-world language is useful—and easy to overread

Philosophers have long analyzed counterfactuals through possible worlds: roughly, consider the closest situations in which the antecedent is true and ask whether the consequence is also true. The framework helps expose questions about similarity and background conditions. Which facts stay fixed? Which laws or regularities remain? How much change is allowed to make the antecedent true?

But semantic usefulness does not settle ontology. A theory can represent alternatives as possible worlds without claiming that every alternative is a physically existing universe. Maps represent roads not taken; spreadsheets represent budgets not adopted; simulations represent storms that did not occur. Representation does not manufacture its referent.

The same caution applies to the simulation hypothesis. A computer can render several trajectories from the same initial data. That shows the system can encode alternatives. It does not show that our unrealized choices continue as accessible physical histories. The Simulation Hypothesis, Without the Hype separates computational possibility from evidence about the world we inhabit.

Hindsight quietly edits the model

After a failure, people know which details became important. They can build the alternative around those details and leave everything else conveniently stable. “If we had launched thirty minutes earlier, we would have avoided the storm” may ignore that an earlier departure would also have changed traffic, crew readiness, fuel decisions and the forecast information available at the time.

A good counterfactual changes the intervention and then allows all causally downstream consequences to change. It does not hold favorable effects fixed while deleting inconvenient ones. This is why a decision journal is valuable: it preserves the options, probabilities and evidence available before the outcome teaches you which narrative feels obvious.

Responsibility also requires a realistic alternative. A person cannot be blamed for failing to take an option that was unavailable, unknowable or outside their authority. Conversely, uncertainty about the exact alternate outcome does not erase every causal judgment. Evidence may support a probability shift or a range even when it cannot identify a certain individual result.

Sometimes the honest answer is a bound

Counterfactual quantities are not always point-identifiable. Available data and assumptions may support only a lower and upper bound. That is not a defective answer. A bound can show that an effect is too small to matter, large enough to justify caution or still wide enough that a decision turns on values and risk tolerance.

The discipline described in Uncertainty Is Information, Not Failure applies directly. Separate sampling uncertainty from uncertainty about measurement, model structure, confounding and transport to the present case. A narrow confidence interval around the wrong estimand is not precision about the question you meant to ask.

Use the seven-question counterfactual test

1. Specify the actual outcome. What happened, to whom, over what period and according to which measure?
2. Name one intervention. Replace “if things were different” with a change that could be implemented or clearly imagined.
3. Set the comparison. What remains fixed, and why is that stability defensible?
4. Draw the causal path. Which consequences would the intervention change directly, and what else would change downstream?
5. Identify the missing outcome. Which part is observed, estimated, extrapolated or assumed?
6. State the evidence design. Randomization, natural variation, matched comparison, mechanism, expert model or speculation are not interchangeable.
7. Report the range that survives. Give a probability, bound or conditional conclusion rather than a cinematic certainty.

For decisions, add one more question: would more counterfactual analysis change the action? The Reversibility Test prevents a low-stakes choice from becoming an elaborate alternate-history project while protecting choices whose consequences are hard to repair.

What counterfactuals can genuinely do

They can clarify whether a proposed cause makes a difference under specified conditions. They can reveal which assumption carries a conclusion. They can compare policies before deployment, diagnose failures after deployment and improve future decisions. They can also expose moral luck: two equally reckless actions may produce different outcomes even when the underlying decision quality was similar.

What they cannot do is provide direct access to the unobserved case. Detail does not change that boundary. A thousand-run simulation can map implications of a model more precisely; it cannot validate the model by running itself. Validation still comes from contact with observed data, successful interventions, out-of-sample performance and the failure of credible alternatives.

The bottom line

A counterfactual is neither fantasy nor footage from another reality. It is a disciplined comparison between what happened and what a stated causal model says would follow from a changed condition. Its strength comes from the evidence and assumptions that support the comparison—not from how vividly the alternative can be narrated.

Use counterfactuals to learn. Do not confuse the branch on the page with another world you have observed.

Research and framework notes

• Judea Pearl, Causal Inference in Statistics: An Overview (UCLA-hosted paper), on association, intervention and counterfactual reasoning.
• Miguel Hernán and James Robins, Causal Inference: What If, a freely available Harvard-hosted text on causal contrasts, identification and study design.
• Stanford Encyclopedia of Philosophy, Counterfactual Theories of Causation, on possible-world and causal interpretations.
• For the site's evidence standard, see What Counts as Evidence for a Simulated Reality?.