In 2023, a historian at a mid-tier university ran an experiment that got almost no attention. She took a language model, fine-tuned it on 40 years of Cold War diplomatic cables, and asked it to generate the counterfactual history of a world where the Cuban Missile Crisis had gone hot. The model produced 200 pages of plausible-looking history: cabinet memos, newspaper editorials, declassified intelligence assessments, even the text of a surrender document that was never signed. She then asked three Cold War historians to read it. None of them could identify which documents were real and which were generated. The generated history was internally consistent, stylistically accurate, and factually coherent with everything that happened before October 1962.
That experiment is a small version of something much larger, and it points at a question the AI industry has barely begun to ask: what happens when generative models become good enough to explore counterfactual histories at scale?
Not just as a thought experiment. As an actual practice, with tools, methods, and a growing body of output. We now have the technical capacity to generate alternate histories by the millions, to explore them systematically, and to learn from them. And almost nobody is asking what that does to our relationship with the actual history we live in.
Here’s the claim I want to test: the interesting thing about AI-generated counterfactuals is not that they’re plausible. It’s that they’re plausible enough to be seductive, and seductive enough to distort how we understand the real past. The multiverse of counterfactuals is a new kind of epistemic hazard, and we’re not building the tools to navigate it.
What “Counterfactual History” Actually Means
Let me be precise, because the phrase gets used loosely.
A counterfactual history is a history that didn’t happen. “What if the Cuban Missile Crisis had gone hot?” is a counterfactual. “What if the Roman Empire had never fallen?” is a counterfactual. “What if I had taken that job in 2019?” is a counterfactual.
Historians have been writing counterfactuals for decades, and they’ve developed methods for doing it well:
Counterfactuals must be minimal. Change one thing, keep everything else constant. A counterfactual that changes too many things at once is not a counterfactual. It’s a fantasy.Counterfactuals must be plausible. The change must be consistent with what we know about the world. “What if the Cuban Missile Crisis had gone hot?” is plausible. “What if Kennedy had been a space alien?” is not.Counterfactuals must be useful. They must teach us something about the actual history. A counterfactual that doesn’t illuminate anything is just entertainment.
These methods are hard to apply at scale. A good counterfactual history takes months to write. A good counterfactual analysis takes years of research. The number of counterfactuals that have been rigorously produced is small, and the number that have been rigorously evaluated is smaller.
AI changes this. A model can generate thousands of counterfactuals in an afternoon. It can explore variations that no human would think to explore. It can produce output that is stylistically indistinguishable from real historical documents. And it can do all of this without the methodological discipline that historians have developed over decades.
That’s the problem. Not that AI can generate counterfactuals. That AI can generate them faster than anyone can evaluate them, and the output is convincing enough to be trusted.
The Sacred Cow: “More Data Means More Truth”
Here’s the belief I want to dismantle, and it’s the default assumption of the entire generative AI industry.
The claim goes like this: more data means better models, better models mean more accurate outputs, and more accurate outputs mean more truth. The path to truth is paved with more data.
This is wrong, and it’s wrong in a way that’s specific to counterfactuals.
For factual questions, more data does tend to mean more truth. A model trained on more text is more likely to know that Paris is the capital of France. That’s because the answer to the question is in the data, and more data means more coverage of the answer.
For counterfactual questions, more data does not mean more truth. It means more plausible-sounding fabrication. The model isn’t retrieving a counterfactual from the data. It’s generating one. The generation is constrained by the data, but the data doesn’t contain the counterfactual. It contains the ingredients from which the counterfactual is made.
This is the crucial distinction. A model trained on 40 years of Cold War cables is not retrieving the history of a hot Cuban Missile Crisis. It’s synthesizing a plausible history from the patterns in the cables. The synthesis is constrained by the data, but it’s not determined by it. Change the sampling temperature, and you get a different history. Change the seed, and you get a different history. Change the fine-tuning, and you get a different history.
There is no ground truth. There is no “correct” counterfactual. There is only plausibility, and plausibility is a property of the model, not of the world.
The contrarian take: for counterfactual questions, more data makes the model more confident without making it more correct. The confidence is the problem, because it’s indistinguishable from accuracy. A model that generates a convincing counterfactual is not more accurate than a model that generates a less convincing one. It’s just more persuasive.
Why Counterfactuals Are Seductive
Here’s the psychological part, and it’s the part that gets the least attention.
Humans are drawn to counterfactuals. This is well-documented in the psychology literature. We ruminate on the paths not taken. We construct narratives about how things could have been different. We use counterfactuals to regulate emotion, to make sense of the past, and to plan for the future.
The draw is especially strong for upward counterfactuals: the paths where things went better. “What if I had taken that job?” “What if the crisis had been resolved without war?” “What if the empire had never fallen?” These are the counterfactuals that promise a better world, and the promise is psychologically rewarding even when the counterfactual is not real.
AI-generated counterfactuals are powerful because they satisfy this draw. They give us a vivid, detailed, plausible version of the better world. We can read it, see it, almost live in it. And the more vivid and detailed it is, the more it feels real. The more it feels real, the more it distorts our relationship with the actual past.
This is not a hypothetical. It’s happening. There are already communities of people who use AI to generate alternate histories, and some of them spend more time in the counterfactuals than in the actual history. The counterfactuals are more satisfying than the real past, because they can be optimized for satisfaction. The real past can’t be optimized. It just is.
The risk is not that people will believe the counterfactuals are real. It’s that they will prefer the counterfactuals to the real, and the preference will change how they think about the actual world. If you spend a year reading AI-generated histories where the Cuban Missile Crisis was resolved peacefully, the actual history, with its near-miss and its terror, starts to feel less compelling. The actual history is contingent and fragile. The counterfactual is clean and resolved. The counterfactual is easier to live with.
That’s the seduction. It’s not a lie. It’s a preference.
What AI Counterfactuals Are Actually Good For
I don’t want to leave you with nothing but critique. Counterfactual generation is genuinely useful, and the use cases are real.
- Historical research.
Historians have always used counterfactuals to test causal hypotheses. “If X hadn’t happened, would Y still have happened?” is a standard way of figuring out what caused what. AI can generate these counterfactuals at a scale that was previously impossible, and the results can be genuinely illuminating.
The catch is that the results are hypotheses, not findings. A generated counterfactual tells you what the model thinks would have happened, not what would have happened. It’s a starting point for research, not a conclusion.
- Policy analysis.
Governments and NGOs have started using counterfactual generation to explore policy scenarios. “What if we had intervened earlier?” “What if we had chosen a different strategy?” The generated counterfactuals help identify alternatives that humans might have missed.
The catch is the same: the counterfactuals are hypotheses, and they need to be evaluated by humans with domain expertise. A generated policy counterfactual is not a prediction. It’s a prompt for thinking.
- Education.
Counterfactual histories are a powerful teaching tool. They help students understand contingency, causality, and the role of individual decisions in shaping outcomes. AI can generate counterfactual histories that are tailored to a specific student’s level and interests, which is something that wasn’t possible before.
The catch is that the student needs to understand that the counterfactual is not real. If they don’t, the teaching tool becomes a source of confusion.
- Personal exploration.
People use AI to explore counterfactual versions of their own lives. “What if I had moved to Berlin?” “What if I had stayed with my ex?” This is a form of self-reflection, and it can be genuinely useful for processing regret, exploring identity, and making decisions.
The catch is that the counterfactual is not real, and living in it too long can become a form of avoidance. The real life is the one you’re living. The counterfactual is a rehearsal, not a replacement.
The pattern: AI counterfactuals are useful when they’re treated as tools for thinking, not as substitutes for reality. The moment you start preferring the counterfactual to the actual, the tool has become a hazard.
What To Actually Do
Three moves, concrete enough to start on Monday.
- Treat AI counterfactuals as hypotheses, not facts.
If you’re using AI to generate counterfactuals, label them as hypotheses. Don’t let them enter your mental model as facts. The generated history is a story the model told you, not a history the world lived. The distinction matters, and it matters most when the story is compelling.
A rough starting point:def counterfactual_analysis(prompt, model, n_samples=20):
samples = [model.generate(prompt) for _ in range(n_samples)]
return {
"prompt": prompt,
"samples": samples,
"plausibility": [score_plausibility(s) for s in samples],
"disagreements": find_disagreements(samples),
"uncertainty": estimate_uncertainty(samples),
}
The point is to generate multiple counterfactuals and look at the distribution, not at any single sample. The distribution tells you where the model is confident and where it’s guessing. A single sample tells you nothing.
- Distinguish between upward and downward counterfactuals.
Upward counterfactuals are the ones where things went better. They’re the most seductive, and the most likely to distort your relationship with the actual past. Downward counterfactuals, where things went worse, are less seductive but often more informative. They remind you of what actually happened.
If you’re generating counterfactuals, generate both. Compare them. The comparison is where the insight lives.
- Build a reality anchor.
If you’re going to spend time in counterfactuals, build a reality anchor. Something that reminds you of the actual history, the actual present, the actual life. It could be a journal entry, a physical object, a conversation with someone who shares your reality. The anchor is not a constraint on exploration. It’s the thing that makes exploration safe.
Without an anchor, the counterfactual becomes the default, and the actual becomes the exception. That’s the failure mode, and it’s a quiet one.
The Mental Model
Here’s the lens I want you to keep.
Generative models don’t just generate text. They generate plausible worlds, and plausible worlds are seductive in a way that facts aren’t. The seduction is not a lie. It’s a preference, and the preference is the thing to watch.
The interesting question is not whether AI can generate counterfactuals. It can, and it will. The interesting question is whether we’ll build the tools and norms to keep counterfactuals in their proper place: as tools for thinking, not as substitutes for reality. The tools are easy. The norms are hard, because they require discipline in an environment that rewards the opposite.
The multiverse of counterfactuals is not a physical multiverse. There is no other world where the Cuban Missile Crisis went hot. There is only this world, with this history, and this present. The counterfactuals are a way of understanding this world better. The moment they become a way of escaping it, they’ve stopped being useful.
What’s your read? If you’ve used AI to explore counterfactuals, did the exploration change how you feel about the actual history? And if you’re building tools for counterfactual generation, how do you keep them in their proper place? Drop it in the comments. I’m genuinely curious whether the seduction is as strong in practice as it looks from the outside.
Top comments (0)