Almost every text about the future of AI has one thing in common: it guesses. A year, a percentage, a curve pointing somewhere definite. Everyone has a timeline. That is understandable — uncertainty is uncomfortable, and a number feels like knowledge. But a forecast is a strange thing to demand from a field that changes so quickly that yesterday's measurement points become saturated before they have time to stabilize.
There is an alternative that sounds softer but is harder in thought: the scenario. And it is worth being exact about what a scenario is, because the word is abused. A scenario is not a forecast. Nor is it a desired future. A scenario is a consistent story about how already observable uncertainties can combine. Its purpose is not to tell us which course of events will occur, but to make choices testable under several possible courses.
The difference is not academic. It determines what one can actually do with the text.
Consider what a forecast requires. To predict, one must choose: a speed of capability development, an adoption rate, a capital cycle, a geopolitical direction. Each of these is uncertain. Choosing a value for each and multiplying them together yields a single story that looks precise but rests on four guesses stacked on top of each other. Worse: it hides the real problem for policy — that decisions have to be made before the outcome is known. A forecast pretends the uncertainty is gone. It is not gone. It is merely hidden.
A scenario does the opposite. It makes uncertainty visible. If one builds three stories — say one where development is fast, one where it slows, and one where it becomes uneven — and tests each against the same set of variables, something appears that a single forecast can never show: which questions recur regardless of scenario, and which change depending on how events unfold. The first is gold. A measure that is wise in all three stories is robust. A measure that works in only one of them is fragile — and one knows that before betting on it.
Here is the discipline that separates honest AI writing from the rest: the distinction between what is evidenced and what would be forecasted. Evidenced is what can be measured and referenced — model results on established tests, companies' actual use, capital flows, electricity demand, expert assessments with names and dates. Forecasted is a guess about outcomes. Most texts blur them: a measurement and a speculation are presented in the same breath, with the same confident tone. Anyone who wants to think honestly keeps them separate throughout, and says plainly which is which. "This is evidenced." "This would be a guess." That is not excessive caution. It is a condition for the reader being able to trust the rest.
There is one more thing scenario thinking forces into the open, and it is uncomfortable: scenarios must not be values. If one names three stories and quietly loads one as "the good one" and another as "the bad one", one has not built scenarios — one has built an argument disguised as analysis. A fast trajectory is not "success" just because the technology spreads quickly; it can stress institutions with long lead times. A slow trajectory is not "safety"; it can conceal passivity. The point is not to choose which scenario one prefers. The point is to build preparedness that holds while the outcome is still unknown.
This is harder than it sounds, because it goes against a strong instinct. A forecast feels like knowledge and gives a pleasant sense of control. A scenario initially feels like indecision — like refusing to answer the question. But it is exactly the opposite. The forecast is often borrowed confidence; it sounds firm but is a guess in a fine suit. The scenario is a method: it forces the question "which decisions are wise regardless of how this goes?" — and that question has actual answers.
That is where the gain lies. Scenario thinking replaces a question that cannot be answered with one that can. "When will AGI arrive?" cannot be answered honestly; no one knows, and those who pretend to know are guessing. "What is wise to prepare that does not depend on when AGI arrives?" can be answered perfectly well. Put the indicators early in the chain, so change is seen when it happens rather than long afterward. Take reversible steps where possible, and early steps only where lead times are so long that waiting becomes more expensive than being wrong. It is not spectacular. It is manageable, which is the whole point.
The temptation to guess will not disappear. It is built into how we talk — we want a number, a year, a verdict. And sometimes a guess is all there is. But one should know what one is doing when one guesses, and one should not sell the guess as knowledge. The problem with prophecy is that it is satisfying and usually wrong. The value of a scenario is that it survives being wrong. The goal has never been to know the future. The goal is not to be surprised by it.
Method note
The method in this text is the same one behind the scenarios on this site: three stories tested against the same variables, no value-loading, and a consistent distinction between what is evidenced and what would be a guess.