Did an Ig Nobel Study Really Prove Rich People Steal Candy From Children?

No. A real 2012 experiment found that UC Berkeley students temporarily induced to see themselves as relatively higher in social rank reported taking more candy than students induced to see themselves as lower-ranking. But the participants were not recruited as rich and poor people, they were explicitly told they could take candy, and no children were present at the candy jar.

The result itself was real as reported: the higher-rank group averaged 1.17 candies, compared with 0.60 in the lower-rank group. That is nearly twice as much, but an absolute difference of 0.57 candy per participant. The researchers treated taking more as an unethical behavior because the candy was described as intended for children in another laboratory.

The study became viral again after Paul Piff and his coauthors received the 2026 Ig Nobel Economics Prize on September 3. The official citation said they had gathered evidence that upper-class people were more likely to “snatch from a children’s supply of candy” and engage in other unethical behavior. Some subsequent headlines sharpened that into the much more literal claim that rich people steal children’s sweets.

That framing goes beyond what the candy experiment established.

There is another important part of the story that most current coverage leaves out: we could not locate a published peer-reviewed direct replication of the candy experiment itself, while several other prominent findings from the same seven-study 2012 paper have subsequently failed direct, preregistered or much larger replication tests.

The evidence does not show that the candy result was false. It does show that the broader conclusion—higher social class generally makes people more unethical—is much less settled than the revived headlines imply.

What actually happened in the candy experiment?

The candy test was Study 4 of a seven-study paper by Piff, Daniel Stancato, Stéphane Côté, Rodolfo Mendoza-Denton and Dacher Keltner, published in the Proceedings of the National Academy of Sciences in 2012.

Study 4 involved 129 University of California, Berkeley undergraduates.

Rather than recruiting objectively wealthy and poor participants, the researchers randomly assigned students to a psychological manipulation intended to make them experience themselves as relatively higher or lower in social class.

Participants were shown a ladder representing socioeconomic standing. One group compared itself with people at the very bottom of the ladder; the other compared itself with people at the very top. The manipulation produced a measurable difference in participants’ subsequent ratings of their own social-class rank.

At the end of the experiment, participants encountered a jar containing individually wrapped candy. According to the paper, the candy was ostensibly intended for children in a nearby laboratory.

Then came a detail that disappears from many retellings:

The participants were told they could take some candy if they wanted.

They were briefly left alone with the jar. After completing other material, participants reported how many candies they had taken. The published outcome was therefore a self-reported candy count, not a record of researchers catching people secretly stealing candy.

The result was statistically significant:

So “almost twice as much candy” is mathematically correct: 1.17 is about 95% higher than 0.60.

But stating the absolute numbers gives readers a much clearer picture of the behavior being measured.

The viral version changes several important facts

The difference between the headline and the experiment is not merely semantic.

None of this means the experiment had no meaning.

Participants believed that taking more candy would leave less for children, which is why the researchers used the behavior as an experimental measure of entitlement or unethical self-interest.

But describing that procedure as scientists discovering that rich people steal from children creates a much stronger image than the experiment warrants.

Were the participants actually rich?

Not in this experiment.

This distinction matters because randomization is simultaneously one of Study 4’s strengths and one of the reasons the viral interpretation is misleading.

By randomly assigning students to the higher- or lower-rank condition, the researchers could investigate whether the induced psychological experience of relative status affected subsequent behavior.

That is stronger causal evidence than merely discovering a correlation between someone’s income and candy-taking.

But what was randomized was not actual wealth.

Researchers did not randomly give people high incomes, inherited fortunes, financial insecurity, elite professional networks or years of experience living in different socioeconomic environments. They changed the comparison participants were asked to make about their relative social position.

The 2012 paper itself described Study 4 as an experiment involving “experimentally primed higher and lower social-class mindsets.”

It is therefore reasonable to interpret the experiment as evidence about a temporary relative-rank manipulation in that laboratory setting.

It is not sufficient by itself to conclude that acquiring real wealth makes people generally unethical.

There is also an opposite error worth avoiding: the broader seven-study paper did not exclusively use artificial rank manipulations. Other studies used different measures of social class, including vehicle status and participants’ own reported socioeconomic standing. The accurate correction is specifically that the famous candy experiment was not a comparison of rich people with poor people.

Did they actually “steal” candy?

Not in the normal meaning of the word.

The paper explicitly says the experimenter told participants that they could take some if they wanted.

That means participants were not secretly taking something they had been forbidden to take.

The researchers nevertheless considered the behavior ethically meaningful because participants had been told the candy was for children. Taking more for oneself supposedly meant leaving less for them.

The distinction is important:

Verified fact: the higher-rank group reported taking more candy.

Researchers’ interpretation: taking more represented greater unethical entitlement or self-interest.

Misleading description: participants were caught stealing candy from children.

Even the original paper’s abstract used the more careful description that higher-class participants were more likely to “take valued goods from others,” rather than saying researchers had observed theft.

Were children actually there?

No child was described as being present at the candy jar.

The participants were told that the candy was intended for children in a nearby laboratory. The paper also calls the candy “ostensibly” intended for those children.

For the experimental manipulation, what mattered was that participants believed taking candy would reduce what remained for children.

But phrases such as “stealing candy from babies” or “stealing children’s sweets” evoke a considerably different event.

How strong was the original candy result?

The result should not be dismissed simply because 129 participants would be considered a modest sample by current standards.

The difference was reported as:

t(124) = 3.18, p < .01, Cohen’s d = 0.57.

A d of 0.57 is a moderate standardized difference. This was not a result that barely crossed the conventional p < .05 threshold.

A statistical criticism published shortly afterward by Purdue psychologist Gregory Francis actually estimated Study 4 as the strongest-powered of the seven reported effects under his method. His objection concerned the improbability, in his analysis, that all seven studies would produce statistically successful results.

That criticism became disputed, and it is not sound grounds for declaring the candy finding false.

The more consequential limitations today are different: how narrowly the experiment should be interpreted, whether its behavioral measure captures what headlines imply, whether it generalizes to real socioeconomic differences, and whether independent studies have reproduced it.

Has the candy experiment itself been replicated?

After searching the literature surrounding the 2012 paper and subsequent replication projects, we could not locate a published peer-reviewed direct replication of Study 4’s candy-taking paradigm.

That wording matters.

It would be too strong to say that the experiment has definitively “never been replicated.” An unpublished study, thesis or obscure experiment could exist outside the literature we located.

More importantly, lack of a direct replication is not the same thing as a failed replication.

The evidence supports this conclusion:

The candy result remains a reported finding from the original experiment, but we did not find a published direct independent test showing that the same candy effect reproduces.

That matters because several other highly visible findings from the same paper have subsequently been put through direct replication tests—with considerably less favorable results.

The candy experiment was only one part of a seven-study paper

The 2012 PNAS paper did not base its broad conclusion on candy alone.

The authors reported seven studies spanning several operationalizations of social class and unethical behavior.

The replication record therefore does not justify saying that the entire paper has been “disproved.” Some findings have not received direct tests, and replication studies themselves differ in setting, sample and methodology.

But the record also no longer supports presenting the original seven-study pattern as though seven mutually reinforcing experiments still provide unchallenged evidence for one general rule.

The driving studies failed well-powered preregistered replications

Studies 1 and 2 were among the paper’s most memorable findings.

The original researchers reported that drivers of higher-status vehicles were more likely to cut off other vehicles and fail to yield to pedestrians.

In 2023, Minah Jung, Paul Smeets, Jan Stoop and Joachim Vosgerau published two direct, well-powered and preregistered replications of those field studies.

They found no evidence of the original positive relationship between socioeconomic status and unethical or selfish driving behavior in either replication.

That does not prove the original observations were fabricated or impossible. Traffic patterns, locations, vehicle-status signals and populations can change.

It does establish something narrower and important: the famous vehicle-status relationship did not prove robust when researchers deliberately attempted to reproduce it under stronger modern replication procedures.

The salary-negotiation finding also failed direct replication

Study 5 reported that participants with higher subjective social class were more willing to deceive someone in a hypothetical salary negotiation.

A 2018 replication by Alexa Clerke and colleagues attempted to reproduce the same result using the same basic measures and procedures in two substantially larger samples: 317 participants from Amazon Mechanical Turk and 320 from Prolific.

The correlations between socioeconomic status and truth-telling were essentially zero in both samples: r = .01 and r = .03.

The researchers therefore found no evidence that higher socioeconomic status predicted a greater propensity to lie in that task. They did reproduce some relationships involving positive attitudes toward greed, but not the core class-to-dishonesty association.

That replication was not protected by a conventional time-stamped preregistration before data collection, a limitation its authors disclosed.

Its much larger samples and close reproduction of the original task nevertheless make it directly relevant to the strength of Study 5.

The proposed “greed” mechanism also performed poorly

Study 7 attempted to explain why higher-class participants might behave more unethically by testing attitudes toward greed.

A later project conducted four replications—three direct and one conceptual—and preregistered its methods and analysis on the Open Science Framework. The authors also contacted the original corresponding author before data collection to improve procedural fidelity.

The crucial interaction between socioeconomic status and the greed manipulation did not appear in the replications.

When the researchers meta-analyzed the original study with all four replications, the estimated interaction was approximately b = −0.01, 95% CI [−0.07, 0.05]—effectively centered on zero.

The project also helped uncover a numerical error in the original Study 7 reporting.

What was corrected in 2017?

PNAS published an official correction to the 2012 paper in October 2017.

It changed a reported standard error in Study 7 from 0.18 to 0.10. The journal stated that the other statistics were correct and that the change did not alter the conclusions of Study 7 or the overall article.

The correction did not change:

the Study 4 candy averages, the Study 4 test statistic, or the Study 4 effect size.

So saying “the candy experiment was later corrected” would itself be misleading.

Larger modern studies also challenge the simple “higher class = less ethical” conclusion

The most important question is no longer whether every individual 2012 experiment can be repeated exactly.

It is whether stronger subsequent evidence supports the paper’s broader proposition that people of higher socioeconomic status generally behave more unethically.

Several large modern projects make that proposition difficult to defend as a general rule.

More than 11,000 participants in two German experiments

A 2022 paper in The Economic Journal used two large survey experiments involving more than 11,000 participants from heterogeneous German samples.

Researchers examined both participants’ socioeconomic characteristics and experimentally primed social status, then measured behavior in an incentivized ethics task.

Their conclusion was direct: the data rejected the proposition that higher socioeconomic status predicted less ethical behavior, both correlationally and under their experimental manipulation.

This was not a replication of the candy jar.

It therefore cannot tell us whether the original 1.17-versus-0.60 candy difference would reproduce.

But it is highly relevant to the much broader claim that the candy experiment is often used to illustrate.

A 33,536-person preregistered replication project

An even larger reassessment appeared in Nature Human Behaviour in 2025.

Anatolia Batruch, Nicolas Sommet and Frédérique Autin preregistered replications of 35 hypotheses derived from 22 influential social-class studies and tested them in 33,536 participants across the United States, France, Switzerland and India.

The results were not a wholesale collapse of social-class psychology: approximately half of the effects successfully replicated. The researchers found substantial support for several theories involving social constraints, uncertainty and status.

But the behavioral predictions relevant to the idea that higher-class people are especially self-oriented and unethical performed poorly.

The researchers found no evidence that higher-class participants were more likely to use deception in the hypothetical negotiation task derived from the earlier literature. In most samples, higher social class instead corresponded with less unethical responding on that measure. The paper likewise found no support for two related behavioral hypotheses involving self-benefiting unethical behavior and public-versus-private generosity.

That is a substantially stronger reason for caution than simply pointing out that an old experiment had a modest sample.

The study received an author correction in February 2026 concerning an incorrectly coded education contrast in the Indian sample. The correction affected particular estimates and figures but did not concern the Piff-derived negotiation result discussed here.

Does newer research prove richer people are actually more ethical?

No.

Replacing “rich people are less ethical” with “rich people are more ethical” would commit the same basic error.

A particularly large 2025 Psychological Bulletin meta-analysis synthesized 1,106 effect sizes from 471 independent studies involving 2,340,806 participants across 60 societies.

It found a small overall association between higher social class and greater prosociality, r = .065.

That result is important because it is difficult to reconcile with a simple stereotype in which increasing social class reliably makes people less concerned with others.

But prosociality and unethical behavior are not identical concepts.

Helping someone, donating resources or cooperating more does not automatically mean a person will never cheat or behave selfishly in another context.

The defensible conclusion from the wider literature is therefore not that either class is inherently more moral.

It is that social class and social behavior have a much more heterogeneous and context-dependent relationship than the revived candy headline suggests.

Was the original paper accused of publication bias?

Yes, but that allegation should not be confused with an established finding of misconduct.

Shortly after the original paper appeared, Gregory Francis published a statistical critique in PNAS arguing that seven statistically successful studies in succession appeared implausibly consistent given their estimated statistical power.

Francis calculated that, using his approach, the probability of all seven reported findings rejecting their null hypotheses was about 0.02. He argued that some form of publication or reporting bias could explain the pattern.

Piff and his coauthors responded directly.

They said they had conducted seven studies and reported all seven, explicitly denied selective reporting, and argued that Francis’s analysis was inappropriate because it relied on post-hoc observed power, too few heterogeneous studies and assumptions about common effect sizes.

There is no basis in the evidence reviewed here to accuse the original researchers of fabricating results or deliberately suppressing failed experiments.

Fortunately, the dispute no longer needs to carry most of the evidentiary weight.

The best test of whether a result is robust is actual new evidence. And subsequent direct replications and large preregistered studies provide much more informative evidence about the broad claim than a retrospective argument over the probability of the original seven significant findings.

One small unresolved issue in Study 4

The candy experiment reports 129 participants, and the successful rank manipulation is reported with t(127), consistent with 129 observations in a conventional two-group comparison.

The candy result, however, is reported as t(124), apparently corresponding to 126 observations. The subsequent ethical-scenarios measure uses t(125).

The published article does not explain the apparent difference in usable sample size.

Missing responses, procedural exclusions or other ordinary data-loss explanations are possible. Without the original participant-level data, there is no responsible basis for deciding which explanation applies.

It is a methodological loose end, not evidence of misconduct.

We also found no preregistration or publicly archived raw Study 4 participant dataset. That limits retrospective auditing, although neither practice was standard in psychology to the degree it is today when the experiment was conducted more than a decade ago.

Why did the 2026 headlines become stronger than the experiment?

The drift is visible even within the language used to describe the same result.

The official 2026 Ig Nobel citation says the researchers amassed evidence that upper-class people were more likely to “snatch from a children’s supply of candy.”

The Guardian then headlined its coverage: “Rich people more likely to steal children’s sweets.”

Those formulations progressively collapse several different propositions:

higher experimental social-rank condition → upper class → rich people

and

taking permitted candy supposedly reserved for children → snatching → stealing children’s sweets.

Curiously, some coverage from when the paper first appeared in 2012 explained the experiment more precisely than the 2026 revival.

A National Science Foundation report republished by ScienceDaily explicitly told readers that Study 4 was the experiment in which researchers manipulated participants into temporarily feeling higher or lower in social-class rank. It also stated that participants had been informed they could take candy.

Fourteen years later, the popular description became more literal at roughly the same time the subsequent scientific record was becoming less supportive of the paper’s broadest interpretation.

So what does the Ig Nobel candy study actually tell us?

The most defensible interpretation is narrower than both the viral headline and a sweeping debunk.

Verified: In one 2012 experiment, Berkeley students randomly placed in a higher relative-rank condition reported taking an average of 1.17 candies, compared with 0.60 among students in a lower-rank condition.

Verified: Participants were told they could take candy, and the candy was described as intended for children in another laboratory.

Verified: The experiment manipulated perceived relative social rank rather than comparing groups of objectively rich and poor people.

Verified: Several other prominent findings from the same paper later failed direct or preregistered replications.

Uncertain: We could not locate a published direct replication of the candy paradigm itself, meaning the specific candy result should be treated as not independently confirmed rather than disproven.

Reasonable inference: The accumulated evidence no longer supports treating “higher social class makes people more unethical” as a simple, broadly established scientific rule.

That leaves a less sensational but more accurate conclusion:

The famous candy experiment found what its authors reported. What it did not do was catch rich people stealing candy from children—and 14 years of subsequent research have made its sweeping moral interpretation harder, not easier, to defend.

References and Further Reading

• Piff et al. — “Higher Social Class Predicts Increased Unethical Behavior” (PNAS, 2012) — The original seven-study paper and primary source for the candy experiment, its methods, reported results and authors’ interpretation.

• PNAS — Correction for Piff et al. (2017) — Official correction changing a Study 7 standard error; it does not alter Study 4’s candy result.

• Francis — “Evidence That Publication Bias Contaminated Studies Relating Social Class and Unethical Behavior” (PNAS, 2012) — Contemporary statistical critique of the seven-for-seven pattern of significant results.

• Piff et al. — Reply to Francis (PNAS, 2012) — The original researchers’ response denying selective reporting and challenging Francis’s statistical method.

• Balakrishnan et al. — Four-Study Replication of SES, Greed and Unethical Behavior (Scientific Data, 2017) — Preregistered replication project testing Study 7; the key SES-by-greed interaction did not survive meta-analysis.

• Clerke et al. — “Association Between Social Class, Greed, and Unethical Behaviour: A Replication Study” (Collabra, 2018) — Two larger direct replications of Study 5 that did not reproduce the association between higher SES and willingness to lie.

• Gsottbauer et al. — “Social Class and (Un)Ethical Behaviour: Causal and Correlational Evidence” (The Economic Journal, 2022) — Two large German experiments with more than 11,000 participants finding no evidence that higher social status predicts less ethical behavior.

• Jung et al. — “Social Status and Unethical Behavior: Two Replications of the Field Studies in Piff et al. (2012)” (Journal of Experimental Psychology: General, 2023) — Two direct, well-powered, preregistered replications that did not reproduce the original traffic-study relationship.

• Batruch, Sommet and Autin — “Advancing the Psychology of Social Class With Large-Scale Replications in Four Countries” (Nature Human Behaviour, 2025) — Registered Report testing 35 hypotheses in 33,536 participants; roughly half of all effects replicated, but behavioral hypotheses portraying higher class as more unethical/self-oriented did not.

• Nature Human Behaviour — 2026 Author Correction to Batruch et al. — Correction concerning an education contrast in the Indian sample; it does not concern the Piff-derived negotiation finding discussed above.

• Wu et al. — “Social Class and Prosociality: A Meta-Analytic Review” (Psychological Bulletin, 2025) — Meta-analysis of 471 independent studies and more than 2.34 million participants finding a small positive association between higher social class and prosociality.

• Improbable Research — Official 2026 Ig Nobel Prize Winners — Primary source for the 2026 Economics Prize citation and confirmation that Paul Piff attended the ceremony.

• National Science Foundation/ScienceDaily — 2012 Coverage of the Original Study — Useful contemporary account that explicitly explains the experimental rank manipulation and participants’ permission to take candy.

• The Guardian — 2026 Ig Nobel Coverage — An example of the renewed headline framing describing the result as rich people being more likely to “steal children’s sweets.”

Editorial currency note: This article reflects the published literature located through September 2026. A future direct replication of the candy experiment, release of the original Study 4 participant-level data, or additional high-powered research could change how strongly the result should be interpreted.