There is a small, stubborn puzzle in the study of how people choose under uncertainty: the same odds can produce very different behaviour depending on how they are described, and depending on when they are described. A decision that looks reckless at 2pm can look sensible at 8pm, even when nothing about the underlying numbers has changed. The question worth sitting with is not whether people are rational, but whether the format and timing of probability information quietly rewrites the decision before we have a chance to notice.

The framing effect is older than the internet

Daniel Kahneman and Amos Tversky spent the 1970s and 1980s demonstrating that logically equivalent descriptions of the same risk produce different choices. In their classic "Asian disease" problem, participants chose between programmes described in terms of lives saved versus lives lost — identical outcomes, opposite preferences. This is the framing effect, and it is one of the most replicated findings in behavioural economics.

What matters here is a narrower version of that insight. Probabilities can be expressed as percentages ("a 20% chance"), as fractions ("one in five"), as frequencies ("20 out of 100 people"), or as odds ("4 to 1 against"). Each format carries slightly different cognitive baggage. Percentages are abstract and numeric. Frequencies feel concrete and countable. Odds are the native language of racing and markets, but they invert the intuition most people carry — longer odds mean lower probability, which trips up anyone reading quickly.

The consistent finding across the literature is that frequency formats tend to produce more accurate risk estimates than percentages, particularly among people with lower numeracy. Gerd Gigerenzer and colleagues have argued for years that this is not a trivial presentation quirk but a fundamental mismatch between how human minds evolved to reason (about counts of things) and how modern institutions communicate (about abstract ratios).

So a percentage is already doing something to the decision. The interesting question is what happens when you add a clock.

Why 8pm might be different from 2pm

There is a well-documented daily rhythm to risk-taking and self-control. Research on time-of-day effects — sometimes filed under "morning morality" or "ego depletion" — suggests that self-control resources are not flat across the day. Roy Baumeister's work on willpower, later complicated and partially contested, proposed that acts of restraint draw on a limited pool that depletes with use. Even where the strong depletion model has been challenged, the broader pattern holds: late-day decision-making tends to be more impulsive, more affect-driven, and less inclined toward careful deliberation.

There is also a circadian dimension. Many people experience a late-afternoon dip in alertness and a subsequent evening period where arousal is higher but executive function is lower. This combination — more emotional pull, less top-down control — is precisely the state in which a vivid, concrete description of risk can dominate a colder, more abstract one.

Now layer the framing effect on top. If percentages encourage a more deliberative, calculating mode of thought, and frequencies encourage a more intuitive, experiential mode, then the same format may land differently depending on the time of day. A percentage at 8pm, when deliberative capacity is reduced, might not be processed as a percentage at all. It might be processed as a vague signal of "risky" or "not risky," and the cautious choice becomes the default because the cognitive machinery needed to weigh it properly is offline.

That is the plausible mechanism behind the claim in the title: probabilities shown as percentages double cautious choices at 8pm. Whether the effect size is exactly double in every context is less important than the direction and the interaction. Format and timing are not independent variables. They compound.

The reward loop underneath the caution

Here is where it gets more interesting, and where the behavioural psychology of reinforcement enters.

B.F. Skinner's work on schedules of reinforcement identified the variable-ratio schedule — where a reward arrives after an unpredictable number of responses — as the most resistant to extinction. This is the engine behind a great deal of compulsive behaviour, and it is not confined to any one activity. The key property is unpredictability: the organism cannot predict when the payoff comes, so it keeps responding.

Cautious choices under uncertainty are, in a sense, the mirror image of this. Where variable-ratio reinforcement produces persistence, loss aversion produces avoidance. Kahneman and Tversky's prospect theory established that losses loom roughly twice as large as equivalent gains. A cautious choice is often a refusal to accept a symmetric bet because the downside feels heavier than the upside.

At 8pm, both systems are active. The reward loop is still running — the pull toward the uncertain option is not gone — but the loss-aversion brake is also engaged, and the deliberative capacity to weigh the two is reduced. A percentage format, which requires more cognitive effort to translate into felt risk, may tip the balance toward the cautious default simply because the effortful path is less available.

This is not a story about people becoming more sensible in the evening. It is a story about the decision being made by a different system, with different inputs, and arriving at a different output.

What this means for how we present risk

The practical implication is not that percentages are bad. It is that the format of a probability is a design choice with behavioural consequences, and that choice interacts with context in ways most communicators never model.

Consider a concrete example from public health communication. During the COVID-19 pandemic, UK messaging used a mix of formats: "1 in 3 people" for asymptomatic prevalence, "95% effective" for vaccines, "the risk of hospitalisation is X%". These were not equivalent in how they landed. The "1 in 3" framing was widely criticised for being alarmist; the "95% effective" framing was criticised for overpromising. Neither criticism was about the underlying data. Both were about the format.

Now add timing. A public health message delivered at 8pm, when people are tired and scrolling, is processed differently from the same message delivered at 9am. The percentage format may produce more caution in the evening — which, depending on the desired behaviour, could be a feature or a bug.

For anyone designing decisions — whether in product design, financial communication, or health messaging — the lesson is to test format and timing together, not separately. A/B tests that vary only the wording, or only the hour, will miss the interaction. The effect may well be in the combination.

Where to look next

The forward-looking question is not whether this effect is real — the framing literature is robust enough that it almost certainly is — but how large it is in naturalistic settings and how it varies across individuals. Numeracy is the obvious moderator: people with higher numeracy are less susceptible to format effects, and possibly less susceptible to time-of-day effects on those same decisions. Age, sleep quality, and stress are all plausible moderators that have not been cleanly tested in this specific interaction.

The more useful research direction, though, may be to stop treating "the decision" as a single event and start treating it as a function of when it is made and how the options are described. If a percentage at 8pm reliably doubles cautious choices, then the timing of a decision is not a neutral container. It is part of the decision. And that has implications for everything from when to schedule a difficult conversation to when to send a renewal notice to when to ask someone to commit to something they might regret in the morning.

The clock is not background noise. It is an input.