Is “decision fatigue” real, and does an AI assistant actually fix it?

“Decision fatigue” as most people use the phrase means a specific theory: willpower is a tank, every choice drains it, and by late afternoon you make worse decisions because the tank is low. That theory, called ego depletion, is the shakiest part of the story. Two large multi-laboratory replications found effects statistically indistinguishable from zero, and the famous hungry-judges study that anchors every popular explainer has a plausible mundane explanation. People do get tired, and long stretches of choosing are unpleasant. But “an AI can make your small decisions so you keep your willpower for the big ones” is a pitch built on a mechanism the evidence does not support. Sources were accessed on 11 September 2026.

The original claim, stated fairly

The idea comes from Baumeister and colleagues’ 1998 paper, which proposed that “Choice, active response, self-regulation, and other volition may all draw on a common inner resource.” The best-known experiment had people who “forced themselves to eat radishes instead of tempting chocolates” subsequently quit faster on unsolvable puzzles than people who had not resisted anything. The authors concluded that “the self’s capacity for active volition is limited and that a range of seemingly different, unrelated acts share a common resource.”

The decision-specific version arrived in 2008. Vohs and colleagues “tested the hypothesis that making many choices impairs subsequent self-control”, and reported that “Making choices led to reduced self-control (i.e., less physical stamina, reduced persistence in the face of failure, more procrastination, and less quality and quantity of arithmetic calculations).” That paper is the academic root of the phrase “decision fatigue”.

It was a productive theory. By 2010 a meta-analysis of published studies put the effect at a medium size, d = 0.62. That number is where the trouble started.

What the replications found

In 2014, Carter and McCullough re-analysed that meta-analysis with methods for detecting small-study effects. Their conclusion: “we found very strong signals of publication bias, along with an indication that the depletion effect is actually no different from zero.” A literature that looks like a strong finding can be a filter that only lets positive results through.

The definitive test was a 2016 Registered Replication Report in Perspectives on Psychological Science, led by Hagger and Chatzisarantis. “Multiple laboratories (k = 23, total N = 2,141) conducted replications of a standardized ego-depletion protocol”, using a procedure agreed in advance with the theory’s proponents. The pooled result “was small with 95% confidence intervals (CIs) that encompassed zero (d = 0.04, 95% CI [-0.07, 0.15]”. For reference, the original claim was fifteen times larger.

Proponents argued the task was not depleting enough, so a second, larger project used procedures closer to the classic ones. Vohs, one of the theory’s original authors, led it. Published in Psychological Science in 2021, it ran across 36 laboratories with 3,531 participants. “Confirmatory tests found a nonsignificant result (d = 0.06)”, and Bayesian analysis “found that the data were 4 times more likely under the null than the alternative hypothesis. Hence, preregistered analyses did not find evidence for a depletion effect.”

That is not the whole story, and the honest reading has to include the rest of the abstract. When the authors ignored their preregistered exclusion criteria, exploratory analyses “found a statistically significant effect (d = 0.08)”. And the “depletion effect was larger for participants who reported more fatigue but was not moderated by trait self-control, willpower beliefs, or action orientation.” So: with the best available design, the effect is either absent or tiny, with the largest hint of it among people who said they felt tired. That is a long way from a tank that empties.

Two things are worth holding apart. These are laboratory tests of a specific mechanism (one act of self-control impairing the next). They do not show that people never tire. They show that the willpower-as-fuel model, tested as its authors specified, does not produce the effect it predicts.

The hungry judges

Every explainer reaches for the same graph. In a 2011 PNAS paper, Danziger and colleagues analysed Israeli parole rulings and reported that “the percentage of favorable rulings drops gradually from ≈65% to nearly zero within each decision session and returns abruptly to ≈65% after a break.” It is widely cited as proof that mental depletion warps expert judgement.

The critiques are less famous and more important. Within months, Weinshall-Margel and Shapard published a letter in the same journal after interviewing court personnel, “learning that case ordering is not random”. In particular, “unrepresented prisoners usually go last and are less likely to be granted parole than prisoners with attorneys. Using the same decision rules as Danziger et al., our data indicate that unrepresented prisoners account for about one-third of all cases, but they prevail only 15% of the time, whereas prisoners with counsel prevail at a 35% rate.” They also noted the original analysis “lumped together decisions rejecting parole and cases that were deferred to a later date”.

Then in 2016 Glöckner showed, by simulation, that “the observed influence of order can be alternatively explained by a statistical artifact resulting from favorable rulings taking longer than unfavorable ones. An effect of similar magnitude would be produced by a (hypothetical) rational judge who plans ahead minimally and ends a session instead of starting cases that he or she assumes will take longer directly before the break.” If granting parole takes longer to write up than denying it, and a judge avoids starting a long case just before lunch, you get the cliff in the graph without any hunger at all.

So the hungry-judges result may be real as a pattern and wrong as an explanation. It is also an odd anchor for a theory about your afternoon: taken at face value, a swing from 65 percent to near zero is far larger than anything the laboratory literature has found.

What is actually known about fatigue and choice

Strip away the disputed mechanism and a few things remain defensible. People report feeling tired and depleted after long stretches of demanding work; the 2021 replication found that the small effect it did detect was concentrated among people who reported more fatigue. Post-lunch dips in alertness are a separate, better-supported phenomenon with physiological rather than willpower explanations (see why do I get tired after lunch). And choosing takes time and attention that could go elsewhere.

What is not established is the part the pitch needs: that decisions draw down a shared reserve, that the reserve is what runs out by 4pm, and that reducing the count of small decisions protects the quality of large ones. Where the evidence runs out, it runs out there.

What the AI pitch is selling

The productivity-newsletter pitch for these tools follows a standard structure: invoke the hungry-judges study as if it were settled, add a large and usually uncited figure for how many decisions you supposedly make each day, then present the assistant as the remedy. It is a tidy story, and its foundation is the ego-depletion account that the replications did not support.

If ego depletion were robust, the logic would hold. It is not, so the pitch is solving a problem that may not exist as described. That does not mean assistants are useless. A tool that drafts three email options saves time, which is real. But those are time savings, and they should be sold as time savings, not as protection for a willpower reserve that replications of 2,141 and 3,531 participants could not find.

There is a subtler cost. Handing routine judgement to a tool does not obviously preserve your capacity for hard judgement; the evidence on cognitive offloading suggests it changes how you assess your own abilities, a problem we cover in cognitive offloading and the AI confidence gap. If you delegate the small calls because you believe you are protecting the big ones, and the belief is false, you have paid a real price for an imaginary benefit.

What to do instead

Treat “decision fatigue” as a description of how a long day feels, not a mechanism to engineer around. If you choose worse late in the day, the plausible drivers are ordinary: hours since you ate or moved, sleep, and time on task. Those respond to breaks, food and sleep, not to fewer decisions. Use an AI assistant for what it demonstrably does, which is save time on drafting and lookup, and judge it on that. And when a product claims to rescue your willpower, ask which study it means. If the answer is the judges, you now know the rest of the story. For how we handle theories that failed replication, see how we research.

Common questions

Is decision fatigue a real thing?

As a feeling, yes: people report being tired of choosing after long stretches of work. As the specific theory that decisions drain a shared willpower reserve (ego depletion), the evidence is weak. Two preregistered multi-lab replications, with 2,141 and 3,531 participants, found effects of d = 0.04 and d = 0.06, both statistically indistinguishable from zero.

What did the ego depletion replications find?

The 2016 Registered Replication Report across 23 laboratories found d = 0.04 with a confidence interval spanning zero. The 2021 paradigmatic replication across 36 laboratories found d = 0.06, nonsignificant, with Bayesian analysis favouring the null by a factor of four. Exploratory analyses found at most a tiny effect (d = 0.08), larger among participants who reported feeling fatigued.

Is the hungry judges study true?

The pattern was observed: favourable parole rulings fell across each session and rebounded after breaks. The depletion explanation is contested. A published critique found case ordering was not random, with unrepresented prisoners heard last and succeeding less often, and a 2016 simulation showed the pattern can arise from favourable rulings taking longer and judges avoiding long cases before a break.

Can an AI assistant reduce decision fatigue?

It can reduce the time you spend on some tasks, which is a genuine benefit. The claim that it preserves a willpower reserve for important decisions depends on ego depletion theory, which large replications did not support. There is no controlled evidence that offloading small decisions to an assistant improves the quality of later large ones.

Why do I make worse decisions late in the day?

Likely causes include hours since eating, poor sleep, and sustained time on task, all of which affect alertness through mechanisms unrelated to a willpower tank. The 2021 replication found the small depletion effect it detected was larger in people who reported more fatigue. Breaks, food and sleep are the evidence-aligned responses; reducing your decision count is not.