In 2007, Netflix looked vulnerable.

If you had asked people then whether Netflix, still widely known as a DVD-by-mail company, would outcompete Blockbuster and become one of the most influential media companies in the world, many would have been sceptical. There were plenty of reasons to believe the future would unfold differently.

But once it happened, the story changed.

Suddenly, the signs seemed obvious. Blockbuster was too slow. Netflix understood technology better. Consumer behaviour was shifting. Broadband adoption was increasing. The economics of physical stores were unsustainable. The future was visible to anyone paying attention.

A decade later, the outcome seemed almost self-evident, even though the explanation arrived after the fact.

Nassim Nicholas Taleb explores this tendency in The Black Swan. Before a rare and consequential event occurs, the future is uncertain, ambiguous, and difficult to predict. Afterwards, we reconstruct a narrative that makes the outcome feel far more predictable than it really was. We identify the causes and build a coherent explanation that suggests the result was always there to be seen.

This is called post-rationalisation.

Humans misunderstand uncertainty because we experience the world forwards, but explain it backwards. Our explanations may contain some truth, but because we already know the outcome, we underestimate how many other outcomes were possible at the time.

The story feels inevitable because we know the ending.

What makes this observation particularly interesting is that the same pattern can be reversed.

Instead of explaining the past, we can start to speculate about the future in the same way. We take a conclusion about what might happen, construct a logical argument for why it will happen, and then gradually treat the argument as though it supports the predictive claim. The conclusion feels stronger because the reasoning around it is coherent.

This is what I call the pre-rationalisation fallacy.

Pre-rationalisation is the mistake of supporting a predictive claim with logic while leaving the assumptions behind that logic unevidenced.

The important distinction is between logic and evidence. Logic can show that a conclusion follows from a set of assumptions. Evidence helps us decide whether those assumptions reflect reality.

When these things are conflated, a prediction can feel grounded because the argument is logically coherent, even though the assumptions carrying the argument have not been tested (i.e. there is no evidence to support the claim).

Logic tells us what conclusions follow from a set of assumptions. If those assumptions are true, complete, and weighted correctly, logic can help us reason about what may follow.

Unfortunately, reality often contains variables absent from the model, and while a logical argument without evidence may correctly identify causal mechanisms, it may also exaggerate or underestimate the importance of the involved factors.

A logical argument can therefore be internally consistent and still be a poor predictor.

On the other hand, evidence is information from reality that should change our confidence in a claim. Evidence is gathered through repeatable scientific experiments which test whether the assumptions behind an argument reflect reality.

Logic connects assumptions and conclusions. Evidence tests assumptions against reality.

A predictive claim may sound convincing because the reasoning is coherent, but its strength depends on whether the assumptions behind the reasoning reflect reality. Without that, the argument may only show what would happen in a world matching the claimant’s model.

The world is rarely identical to the model.

The rise of generative AI provides a useful case study because the uncertainty and stakes around claims are high, while the incentives to own the narrative are enormous.

AI is one of the defining disruptions of the decade, and people who seem to understand where things will gain influence. This does not necessarily imply bad intentions – many people are acting in good faith – but incentives shape behaviour. When a major disruption emerges, there is a reward for explaining what it means before there is necessarily enough evidence to know with confidence.

People rush to build their explanation of where AI is going and how we should react. Those explanations are often constructed around existing worldviews, incentives, biases, and hopes. Confirmation bias strengthens the effect: once we have a preferred explanation, we become more likely to notice supporting examples and less likely to inspect the assumptions that would weaken the argument.

As a result, AI becomes a canvas onto which people project their existing beliefs.

The result is a flood of competing narratives. AI will centralise power. AI will decentralise power. AI will eliminate specialists. AI will increase the value of specialists. AI will replace software engineers. AI will make software engineers more productive.

Many of these arguments are internally consistent, but they cannot all be true simultaneously. Logical consistency alone only tells us what might follow if a particular set of assumptions turns out to be correct.

This is why reading AI-related content can become disorienting. The reader is often choosing between several coherent arguments that begin from subtly different assumptions. The argument that sounds most confident may simply be the one that hides its assumptions most effectively.

As larger claims depend on more assumptions, pre-rationalisation becomes more prevalent as the size of the disruption increases.

When change is small, the conditions of the claim are easier to state. If a team introduces AI into a customer support workflow, they can define the context and measure whether the change helped. The claim is narrow enough for evidence to correct it.

Large disruptions create broader claims. When someone says that AI will redefine how every organisation works, the conclusion depends on many more conditions. A claim may be stronger in some settings than others, but the broader phrasing hides those differences.

Broad claims feel more widely applicable, attract more attention, and position the speaker as having a higher-level perspective. The claim sounds strategic because it is broad, but the cost is precision.

As the claim becomes broader, the conditions under which it is true become harder to see, and when the assumptions disappear, the argument becomes easier to repeat and harder to test.

This is where pre-rationalisation becomes most prevalent. A broad predictive claim can be supported by a coherent argument, while the assumptions required for that claim remain hidden. The reader or listener is left with a feeling of insight, but without the ability to judge the prediction.

Broad claims also become difficult to disprove.

If someone claims that AI will reduce response time in a specific workflow, the claim can be tested. The conditions are visible enough for reality to correct the argument. If response time does not improve, quality drops, or rework increases, the claim has been weakened.

A claim such as “AI will transform every organisation” is different. If the transformation does not appear in one case, the claimant can point to the wrong implementation, culture, or timing. Some of those explanations may be valid, but the original claim did not state its conditions clearly enough to be tested against them.

A claim that never states its conditions can always escape into them later.

This is why sweeping AI commentary can remain convincing even when individual examples disappoint. The claim is broad enough to absorb contrary evidence. Success in one context can be treated as confirmation, while failure elsewhere can be explained as a mismatch between the claim and the context.

That makes the claim less useful. A prediction that cannot be weakened by reality gives us little guidance about what to believe, what to measure, or what to do next.

During periods of uncertainty, we need models, predictions, and experiments to help us reason about what might happen next. Logic, evidence, and speculation all play a part, but we should distinguish between them.

When reading or writing AI content, or any content about an uncertain but high-stakes topic, the following questions help put claims into context.

1. Did the problem come before the solution? A claim that starts with AI as the assumed answer may be reasoning backwards from a preferred conclusion.

2. Does the claim clearly outline its assumptions? The strength of a prediction depends on the conditions that must hold for the argument to work – they should be described precisely.

3. Where does the claim apply? A claim that is true in one context can become misleading when quietly generalised to a much broader one.

4. Is the claim disprovable? A claim that can survive almost any outcome is protected from the very evidence that should discipline it.

5. Does the claimant understand the counterpoints? A serious argument should understand the conditions under which it might fail and the reasoning behind counterclaims.

6. Who benefits if the claim is believed? Incentives can shape which assumptions are emphasised, which are ignored, and how confidently the conclusion is presented.

This is a non-exhaustive list, but it helps separate a coherent argument from a well-supported one.

Pre-rationalisation is the mistake of supporting a predictive claim with logic while leaving the assumptions behind that logic unevidenced. Naming the mistake helps us notice when a conclusion is borrowing more confidence than it has earned.

AI may be one of the most significant technological shifts in decades, so learning how to think critically when big claims are made will prove more important than ever. When someone makes a claim like "30% of the workforce will be made redundant by AI in the next three years", we should examine not only if their argument is logical, but also ask what the assumptions are and whether there is evidence to support them.