Is willpower a battery? Ego depletion after the failed replications

"I've used up all my willpower today." It sounds like common sense, and for about fifteen years it was also mainstream psychology: self-control was a limited resource, every act of it drew down the same tank, and a hard morning left less for the afternoon. The name for it was ego depletion.

Then the large replications came in, and the effect mostly disappeared. This piece goes through what the original studies found, in their own numbers, what happened when dozens of labs ran them again, what is left standing (tiredness after hours of hard thinking is real; a battery that one choice drains is not), and how to plan hard work without the battery model.

Radishes and cookies: the 1998 study

Roy Baumeister, Ellen Bratslavsky, Mark Muraven and Dianne Tice asked 67 students to skip a meal, then sat them in a room that smelled of freshly baked chocolate chip cookies. One group had to eat radishes and leave the cookies alone. A second group could eat the cookies and chocolates. A third got no food at all. Then everyone worked on puzzles that, unknown to them, could not be solved. How long they kept trying was the measure of self-control left over.

Radish group (resisted the cookies): 8.35 minutes on average, 19.4 attempts
Chocolate group: 18.90 minutes, 34.3 attempts
No food: 20.86 minutes, 32.8 attempts

The radish eaters gave up in 60% less time than the no-food group (8.35 against 20.86 minutes). The authors' reading: resisting the cookies had used up something, and the puzzle got less of it. It was a large, vivid effect, and it launched hundreds of studies with the same two-step design: one task that takes self-control, then a second one to see if people do worse.

How the idea grew

By 2010 there were enough studies for a meta-analysis. Martin Hagger and colleagues pooled 83 studies with 198 tests and found an average effect of d = 0.62, which they called medium to large. In everyday terms, d = 0.62 means the average "depleted" person would do worse than about 73% of people who weren't depleted.

There was a mechanism too. In 2007 Matthew Gailliot, Baumeister and colleagues published nine studies suggesting self-control used up blood glucose and that a sugary drink restored it. The battery had a fuel.

The cracks

Three problems showed up between 2010 and 2016.

Publication bias. Evan Carter and Michael McCullough re-ran the 2010 analysis with statistical corrections for small studies and missing null results. Adding unpublished studies and correcting for bias (2015), the pooled effect fell from g = 0.43 to g = 0.24 by one method and to an effect "indistinguishable from zero" by another.
The glucose story fell apart. Robert Kurzban (2010) argued the brain's energy use barely changes with effortful tasks and that the glucose data didn't show the drop. Daniel Molden and colleagues (2012) found that just rinsing the mouth with a sugar drink, without swallowing, improved self-control as much as drinking it, with no change in blood sugar: a signal of reward, not fuel. A 2016 analysis by Miguel Vadillo and colleagues found the glucose literature showed signs of reporting bias and advised drawing no conclusion from it.
Too good to be true. All nine of the 2007 glucose studies were significant, a pattern that critics calculated had less than a 1% chance with samples that size.

The big replications

The decisive tests were preregistered: many labs agree on one protocol and one analysis before collecting data, and every result gets reported. Three of them tested ego depletion:

2016, Registered Replication Report (led by Martin Hagger and Nikos Chatzisarantis): 23 labs, 2,141 people. A hard version of a letter-crossing task, then a test of attention (the multi-source interference task). The original study it copied had found d = 0.69. Pooled result: d = 0.04, 95% interval −0.07 to 0.15. Only 3 labs found an effect, and one of those went the wrong way.
2021, Dang and colleagues: 12 labs, 1,775 people, a Stroop task then an antisaccade task. Result: d = 0.10, small but statistically significant (0.16 after removing people who may have answered at random).
2021, Vohs, Schmeichel and colleagues, led by Kathleen Vohs, one of the theory's own researchers: 36 labs, 3,531 people, two different protocols. Preregistered result: d = 0.06, not significant. A Bayesian test found the data about 4 times more likely if there were no depletion effect at all than under the effect the team expected.

What those numbers mean

Put them on the same scale as the 2010 meta-analysis. At d = 0.62 the average depleted person does worse than 73% of the comparison group. At d = 0.06 it is 52%, and at d = 0.04 it is 52% as well: a coin flip. Even the best case, d = 0.10, is 54%. On the radish study's own scale, where puzzle times varied by about 7 minutes from person to person, d = 0.04 is about 17 seconds out of 21 minutes, where the original found 12 and a half minutes: invisible on any job.

"Willpower is a tank that one hard choice drains" has had its fair test, in more than 70 labs and with more than 7,000 people. What came back was close to nothing.

Michael Inzlicht, one of the field's best-known researchers and a co-author of the 2021 study, wrote in 2025 that ego depletion "wasn't real" and that he had "felt like a fraud" teaching it. Defenders of the theory argue that the replication tasks were too mild to deplete anyone. That is possible; but if a 15-minute demanding task doesn't do it, the idea doesn't tell you much about a normal workday either.

Decision fatigue and the hungry judges

The best-known cousin of ego depletion is decision fatigue, and its most famous evidence is a 2011 study of Israeli parole boards. Shai Danziger, Jonathan Levav and Liora Avnaim-Pesso looked at 1,112 rulings by eight judges over 50 days. The share of favorable rulings started at about 65% after each food break and fell to nearly zero before the next one, then jumped back to 65%.

It is a striking chart, and it has strong critics:

The order of cases isn't random. Keren Weinshall-Margel and John Shapard (2011) pointed out that prisoners without lawyers tend to be heard last in each session, and lawyers arrange their own cases, so the end of a session holds the weaker cases.
A rational judge would do it too. Andreas Glöckner (2016) showed that favorable rulings take longer (7.4 against 5.2 minutes in the study's own data), so a judge who avoids starting a long case just before a break produces the same curve, with no fatigue at all.
Too big to be plausible. Daniël Lakens noted that a drop from 65% to nearly zero is far larger than almost any known psychological effect.

The parole study doesn't prove that people decide worse as a session wears on. It shows that the pattern is easy to produce in other ways.

Beliefs about willpower

One offshoot looked more promising. Veronika Job, Carol Dweck and Gregory Walton (2010) found that only people who believed willpower is limited showed depletion; and in a 2015 study of 153 students over a term, those who believed willpower is not limited procrastinated less and earned higher grades under a heavy course load.

This, too, has weakened. A preregistered replication by Carruth, Ramos and Miyake (2023, 187 people) found no depletion effect and no effect of beliefs, and in the 36-lab study of 2021 beliefs about willpower made no difference either. Their 2023 review called the evidence "weak at best". So believing willpower is unlimited isn't a proven trick, though it does no harm to stop telling yourself you've run out.

What is left: tiredness, time and breaks

None of this means people never get mentally tired. What holds up is less dramatic and more useful:

Hours of hard thinking leave a mark. In a 2022 study in Current Biology, Antonius Wiehler, Mathias Pessiglione and colleagues gave two groups a working day of either demanding or easy computer tasks. By the end, only the hard-task group had more glutamate in the lateral prefrontal cortex, smaller pupil responses, and a shift towards choices that needed less effort and paid off sooner. The authors read it as a reason fatigue makes us want to stop; it is one study, with a small sample, and it measured a correlation.
Later in the day is a little worse, and breaks help. Hans Sievertsen, Francesca Gino and Marco Piovesan (2016) used about two million Danish national test results, where the time of the test was set by the school timetable. Scores fell by 0.9% of a standard deviation for every hour later in the day, and a 20 to 30 minute break before the test raised them by 1.7%. On a test with a standard deviation of 100 points, that is about 1 point an hour, and about 2 points back from a break: real, small, and steady.
Short breaks lift energy more than output. A 2022 meta-analysis of breaks of 10 minutes or less (Albulescu and colleagues, 22 samples, 2,335 people) found they raised vigor (d = 0.36) and cut fatigue (d = 0.35), but improved performance only slightly (d = 0.16, not significant) and mainly on routine tasks. Demanding work may need longer breaks.

A different model fits these findings better than the battery. Kurzban, Duckworth, Kable and Myers (2013) proposed that the feeling of effort is a signal about opportunity cost: the mind weighing whether the current task is still worth doing against everything else it could do. Effort feels harder when a task is long, dull or unrewarding, not because a tank is empty.

Planning hard work without the battery

If willpower isn't a fuel gauge, the plan for a day changes in a few concrete ways.

1. Plan around hours, not around choices. Fatigue follows long stretches of demanding work. Put the task that needs the most thought in the first two or three hours, which the time-of-day data mildly supports; there is more on this in Eat the frog?.
2. Break long stretches with real breaks. After a couple of hours of hard thinking, take 20 to 30 minutes, the size of break that moved the Danish scores. Five-minute breaks help mood and energy; don't expect them to rescue output on a hard task. The research on timed work blocks is in Does the Pomodoro technique work?.
3. Read "I've got nothing left" as a signal to switch, not to stop. If effort is a cost signal, the fix is often to make the next step smaller or more rewarding, not to wait to recharge. The two-minute start turns a stalled task into a first step small enough to do now.
4. Remove temptations instead of resisting them. The radish study's lesson survives in a different form: the people who do best at self-control mostly arrange their day so they need less of it. Decide the day's list the evening before, keep it in one place, put the phone out of reach for the hard block.
5. On long shifts, watch hours, not willpower. A crew on a 12-hour day gets tired because of the hours and the work; the fix is breaks, rotation of the demanding jobs and handovers that don't rely on memory, not a pep talk.

In tick-tock-done

tick-tock-done keeps the day's list in one place, so the choice of what to do next is made once, in the morning, not every time a task ends. Put the hard task at the top with today's date, tick it off when it's done, and let the rest of the list wait in order. For crews, shifts and the time clock show who has been on for how long, which is the fatigue signal that actually holds up.

What the research does not show

It doesn't show self-control is unlimited. The replications tested short lab tasks of 5 to 15 minutes. They don't rule out that a long, grinding day of resisting impulses has effects; they show that the short-task version, the one the theory was built on, doesn't replicate.
The fatigue study is small. The 2022 glutamate study is a single study with a small sample. Exact group sizes and percentages quoted in press coverage are not repeated here because we could not check them against the paper.
The time-of-day effects are small and come from schoolchildren taking tests, not adults at work.
The parole debate is not settled. The original authors replied that their effect holds after the critics' adjustments; a newer analysis reportedly finds a weaker version of the pattern. What is clear is that the 65%-to-zero figure can't be taken as a measure of fatigue alone.
The planning steps are ours. They follow from the findings above; they were not tested as a package.

Sources

Baumeister, Bratslavsky, Muraven and Tice (1998). Ego depletion: is the active self a limited resource? Journal of Personality and Social Psychology 74(5). PDF
Hagger, Wood, Stiff and Chatzisarantis (2010). Ego depletion and the strength model of self-control: a meta-analysis. Psychological Bulletin 136(4). Record
Carter and McCullough (2014), Frontiers in Psychology. Article · Carter, Kofler, Forster and McCullough (2015), Journal of Experimental Psychology: General 144. PMC
Hagger, Chatzisarantis et al. (2016). A multilab preregistered replication of the ego-depletion effect. Perspectives on Psychological Science 11(4). Full text
Dang et al. (2021). A multilab replication of the ego depletion effect. Social Psychological and Personality Science 12(1). Article
Vohs, Schmeichel et al. (2021). A multisite preregistered paradigmatic test of the ego-depletion effect. Psychological Science 32(10). Record
Molden et al. (2012). Motivational versus metabolic effects of carbohydrates on self-control. Psychological Science 23(10). Record · Vadillo, Gold and Osman (2016), Psychological Science 27(9). Record
Danziger, Levav and Avnaim-Pesso (2011). Extraneous factors in judicial decisions. PNAS 108(17). PDF · Weinshall-Margel and Shapard (2011), PNAS 108(42). Record · Glöckner (2016), Judgment and Decision Making 11(6). Article
Job, Walton, Bernecker and Dweck (2015). Implicit theories about willpower predict self-regulation and grades in everyday life. JPSP 108(4). PDF · Carruth, Ramos and Miyake (2023), PLOS ONE 18(6). PMC · Miyake and Carruth (2023), Frontiers in Psychology. PMC
Inzlicht (2025). The collapse of ego depletion, republished by 3 Quarks Daily. Essay
Kurzban, Duckworth, Kable and Myers (2013). An opportunity cost model of subjective effort and task performance. Behavioral and Brain Sciences 36(6). PMC
Wiehler, Branzoli, Adanyeguh, Mochel and Pessiglione (2022). A neuro-metabolic account of why daylong cognitive work alters the control of economic decisions. Current Biology 32(16). Paris Brain Institute summary
Sievertsen, Gino and Piovesan (2016). Cognitive fatigue influences students' performance on standardized tests. PNAS. HBS Working Knowledge
Albulescu et al. (2022). "Give me a break!" A systematic review and meta-analysis on the efficacy of micro-breaks. PLOS ONE 17(8). Summary