The marshmallow test, Stanford prison, power poses: what can we still trust in stories about the human mind? Follow 48 ideas from their original promise to the evidence that came later.
Some findings held up. Others became narrower or were contradicted. The difference lies in the question tested and the evidence for each case.
Waiting for a reward was associated with later school achievement. In a more diverse sample, that relationship shrank substantially after accounting for family background and early abilities. The developmental question remains interesting; a test lasting a few minutes calls for care when used to judge a child.
What changed in the marshmallow testThe Stanford prison study became a story about the power of roles. Later analysis of documents and recordings showed how the researchers shaped the situation. Understanding the behavior also requires reading the instructions participants received.
Read the Stanford caseSpread learning across sessions and try answering from memory before rereading. Research on spacing and retrieval practice supports these approaches. The useful interval depends on the material and how long you want to remember it.
The evidence on learningIn 2015, the Open Science Collaboration published replications of 100 studies from three psychology journals. 97 original results had been reported as significant. Of these, 35 crossed the statistical threshold again in the same direction: about 36%.
These 100 studies came from articles published in three journals in 2008. That selection and criterion cannot give a percentage of “true psychology.” The authors also compared effect sizes and the compatibility of results.
A new study tests the same question with different participants. Similarities in procedure, population, and measurement matter when interpreting the result.
How unusual a result at least this extreme would be if the null hypothesis and the test’s assumptions held. Effect size and precision answer different questions.
It weakens confidence in the claim tested. Statistical power, the procedure, and the wider evidence help distinguish a small effect from a context-dependent or absent one.
You have an imaginary study with 30 participants. Exposure X and six outcomes Y were generated independently. Try other analyses on the same observations and watch the p-value. It can fall below 0.05 even though the simulated population contains no relationship.
Start by switching the measured outcome. Compare your selected analysis with the original. Then test your choice on 100 new observations.
The interactive lab needs JavaScript. The idea is simple: if you try many analyses and publish only the convenient result, you hide the attempts that explain how you found it. All case files and sources below remain readable.
Analyses calculated: 1
Exploration can uncover worthwhile questions. The problem arises when the attempts disappear from the report and an analysis chosen after seeing the data is presented as a planned test. Reporting the full path and checking on new data make the difference.
The initial test uses Pearson correlation between X and the first outcome Y for the first 30 participants. Six outcomes and three control variables are generated before the first analysis. Adjustment computes a partial correlation, removing the linear relationship with one control variable. The two-sided p-value uses Student’s t distribution with n − 2 degrees of freedom, or n − 3 after adjustment. The plot shows residuals of both variables when an adjustment is used.
Exclusion tests every available observation, keeps the removal with the lowest p, and repeats once. The counter includes those searches and all three candidate adjustments. Displayed p-values are uncorrected for selection and repeated testing. The final check uses independent observations, the chosen outcome and adjustment, with no exclusions. After that check, the analysis ends; you can restart or generate another study. Restarting reuses all observations, including the checking sample; generate another study for new data.
Simmons, Nelson & Simonsohn, 2011Find a story you know or choose a subject. Each case keeps the claim, subsequent research, and sources together. Its label applies to the specific claim described.
These counts describe this collection only.
48 / 48
No cases match. Try another word or clear the filters.
Resisting temptation reduces performance on a subsequent self-control task by consuming a limited resource.
The radishes-and-cookies experiment made the idea of tiring willpower memorable. Later studies found much smaller effects and different results across protocols; a single reservoir of willpower remains a disputed explanation.
Hagger and colleagues, 2016: 23 laboratories, 2,141 participants, estimated d = 0.04 with a confidence interval including zero. Dang and colleagues, 2021, found a small effect with a different protocol.
Your experience of fatigue deserves to be taken seriously. These findings question one explanation for it and the ability of a short task to reproduce it.
How long a child waits for a larger reward is associated with later academic achievement.
The association exists but shrinks after accounting for family background and early cognitive skills. The test neither determines a child’s future nor shows that waiting causes success.
Watts, Duncan and Quan, 2018, examined outcomes at age 15, not adult earnings. Among children whose mothers had not completed college, statistical adjustments reduced the initial association by about two thirds.
A marshmallow can start a conversation about patience; it cannot deliver a verdict on a child’s character or future.
Two minutes in an expansive posture would change hormones and willingness to take risks.
The larger replication did not confirm the hormonal promise. Participants did report feeling more powerful, a different outcome from hormonal change or interview success.
Ranehill and colleagues, 2015, tested 200 people. They found no evidence of changes in testosterone, cortisol or risk-taking on their task.
You may like a comfortable posture without it delivering the advertised biology. Feeling confident and performing better need separate measurements.
A brief intervention about developing intellectual abilities can improve school achievement.
The US national experiment found modest benefits for initially lower-achieving students. Effects also depended on the school environment; a message about effort cannot replace the conditions needed for learning.
Yeager and colleagues, Nature, 2019: the intervention improved grades among lower-achieving students and increased advanced-mathematics enrolment across the sample.
Encouragement can help. When a student struggles, this study does not justify blaming their attitude.
A universal threshold of 10,000 hours of practice would guarantee expertise.
Research on deliberate practice does not establish a threshold guaranteeing success. The relationship with performance varies by domain, and an association alone does not measure the causal benefit of each hour.
Macnamara, Hambrick and Oswald, 2014, synthesized research across domains. Their 2018 correction changed the overall variance explained from 12% to 14%.
Practise purposefully and track improvement. An hour count alone cannot tell you when you will become an expert.
Perseverance and consistency of long-term interests form a distinct trait called grit that predicts success.
Grit overlaps strongly with conscientiousness. Perseverance relates more clearly to performance than consistency of interests and can add predictive information beyond conscientiousness.
Credé, Tynan and Harms, 2017, synthesized 88 independent samples involving 66,807 people. The findings question the usefulness of a single score combining the two components.
Keeping at a task can matter, but a perseverance questionnaire cannot diagnose why someone succeeds or struggles.
Exerting self-control would lower blood glucose, and consuming sugar would restore the depleted resource.
The central predictions of this model were not supported by the research synthesis. Glucose mouth-rinsing findings also offer no secure replacement explanation: the author identified signs of publication bias.
Dang, Appetite, 2016, separately evaluated glucose depletion, its relationship with later performance and glucose ingestion. None of the three predictions was supported.
These experiments do not provide a basis for recommending sugar as a remedy for low willpower.
Changing how options are presented can influence choices without removing alternatives.
These interventions can work, but results vary with the target and implementation. A small average can matter at scale; it does not promise to solve every social problem.
DellaVigna and Linos, 2022, analysed 126 trials from two US units covering 23 million people. Average take-up increased by 1.4 percentage points, compared with 8.7 in the academic-paper sample used for comparison.
Ask which behaviour changed, by how much and at what cost. This set of trials does not supply a constant for every intervention.
Recalling an immoral act would make cleaning products more attractive.
The link between physical and moral cleanliness is a familiar metaphor. Direct replications of the product-rating task did not confirm the originally reported effect.
Earp, Everett, Madva and Hamlin, 2014, repeated Zhong and Liljenquist’s Study 2 in three experiments using the original materials. They did not detect the effect in these tests.
A compelling metaphor can suggest a hypothesis. Each claim about soap changing moral judgment needs its own test.
Insufficient external justification for advocating a view against one’s beliefs would encourage attitude change.
In the famous experiment, participants received $1 or $20 to describe a dull task as interesting. Later research tests the explanatory mechanism without reducing the whole theory to that single study.
Vaidis and colleagues, 2024, tested an essay protocol in 39 laboratories. Greater freedom of choice did not produce the predicted attitude difference, although writing a counterattitudinal essay differed from the neutral condition.
An attitude change does not automatically identify why it happened. The alternative explanations an experiment rules out matter too.
Merely assigning ordinary people to guard roles would spontaneously turn them into abusers.
Abuse occurred in the simulated Stanford prison. Researchers’ instructions, their involvement in events and differences between guards prevent attributing the behaviour solely to the assigned role.
Le Texier, 2019, analysed archives and participant interviews. He documented researcher interventions and data-collection problems that undermine the account of spontaneous transformation.
The case remains important for research ethics and the exercise of power. It cannot establish what anyone would do in a uniform.
In 1964, 38 neighbours supposedly watched Kitty Genovese’s murder without anyone trying to help.
That account is not supported by the records examined. Witness numbers, what they could see and their interventions were simplified into a story of collective indifference.
Manning, Levine and Collins, 2007, re-examined the 38-witness account. Their analysis is a historical correction, not an experimental replication of the bystander effect.
An inaccurate historical story needs correcting separately from the theory it inspired. Research on helping needs its own assessment.
In an experiment backed by scientific authority, many participants continue administering what they believe are electric shocks despite the other person’s protests.
The pressure to continue is documented, and refusal also occurs. The experiments do not measure the proportion of people willing to kill; the shocks were simulated and the situation depended on the protocol.
Burger, 2009, conducted a partial replication with 70 adults, stopping at the 150-volt threshold. He observed substantial willingness to continue without testing behaviour through 450 volts.
The result invites scrutiny of institutional pressures. It cannot predict an individual’s conduct or remove individual responsibility.
Competition between groups can foster hostility, while pursuing shared goals can reduce conflict.
The camp included researcher-organized competitions followed by problems the boys had to solve together. Conflict did not arise in an untouched setting, and cooperation was built gradually.
Sherif and colleagues’ 1961 report of the 1954 study describes activities including restoring the water supply. It is the original study report, not an independent replication.
Interdependence offers a useful idea for cooperation. One camp study does not establish a universal remedy for historical conflicts.
A unanimous majority giving wrong answers can influence someone’s response even on a simple visual task.
Asch’s task involved comparing line lengths. Conformity appears in subsequent research, but its strength depends on the situation and cultural context. A public answer alone does not establish changed perception.
Bond and Smith, 1996, synthesized studies using Asch’s task and found variation across periods and societies. The analysis supports the phenomenon without establishing a universal conformity rate.
When a group appears to agree, it is worth asking how answers were collected and whether people could respond independently.
Time pressure can reduce the likelihood of stopping to help someone.
Theology students encountered a slumped person on the way to give a talk. Those in a hurry helped less often. The talk’s topic, including the Good Samaritan parable, did not produce the expected difference in this sample.
Darley and Batson, 1973, compared hurry conditions and speech topics. The cited source is the original experiment; it does not substantiate thousands of replication participants.
Circumstances can prevent a good intention from becoming action. The study does not show that moral or religious values have no influence.
Anonymity and immersion in a group would weaken self-control and encourage norm-breaking.
Anonymity has no fixed moral direction. Behaviour also depends on the group’s relevant norms, which can encourage helping or aggression.
Postmes and Spears, 1998, analysed 60 studies. Results better supported conformity to situation-specific norms than a general loss of norms.
To understand an anonymous group, examine which behaviours it encourages.
In observed public conflicts, more bystanders are associated with a greater chance that at least someone intervenes.
The chance that a particular person helps and the chance that a victim receives help from someone are different questions. A group can offer a higher chance of intervention even when each member hesitates.
Philpot and colleagues examined 219 recordings from three countries. At least one bystander intervened in 90.9% of the filmed conflicts. This observation alone does not establish causation or represent every emergency.
The figure describes these conflicts, not a personal guarantee. It shows why the perspective used to measure helping matters.
Grouping people by apparently trivial criteria can favour allocating rewards to members of their own group.
In minimal-group studies, participants distributed rewards between people identified by group membership. Favouritism could arise without prior conflict; this does not establish that every classification produces hatred.
Tajfel, Billig, Bundy and Flament, 1971, reported favouritism in reward allocation. The source documents behaviour in these tasks, rather than explaining real-world discrimination on its own.
A group label can influence a decision. The size and consequences of that influence need studying in the situation we care about.
People would divide into logical, left-brain-dominant types and creative, right-brain-dominant types.
Some functions and connections are lateralized. This local specialization does not justify dividing personalities by a dominant hemisphere across the whole brain.
Nielsen and colleagues, 2013, analysed resting-state images from 1,011 people. They observed local lateralization without evidence for the global left-brained or right-brained types proposed by the myth.
You can have both logical and creative abilities. A hemisphere personality quiz does not measure how your brain regions work together.
Students would learn better if teaching matched a fixed sensory style, such as visual or auditory.
Presentation preferences exist, but do not establish a benefit from matching. The hypothesis requires tests showing that different methods benefit the supposed learner categories differently.
Pashler, McDaniel, Rohrer and Bjork, 2008, found no adequate basis for using style assessments in general educational practice. This concerns style matching, not every form of teaching adaptation.
Choose representations suited to the material and the learner’s actual needs. A style label should not restrict the methods available to them.
Listening to a Mozart sonata can temporarily improve performance on some spatial tasks.
The original finding concerned students and spatial tasks, not children’s intellectual development. Similar advantages occur after other music, and publication bias complicates estimation.
Pietschnig, Voracek and Formann, 2010, found little evidence for an advantage specific to Mozart. Comparing music with silence asks a different question from comparing it with another piece.
Music can remain a pleasure without a promise of higher IQ. An immediate test result does not demonstrate lasting change in intelligence.
Changing facial expression can influence felt emotion, including when a pen held between the teeth mimics a smile.
Voluntary expressions and mimicking a smile received support in recent research. Evidence for the pen-in-mouth procedure is less conclusive. A fragile procedure does not invalidate every version of the hypothesis.
Many Smiles, 2022, tested 3,878 participants across 19 countries with a preregistered plan. Findings differed across methods of producing a smile.
This laboratory effect does not justify pressuring someone to smile when distressed and does not establish a treatment for depression.
We use only 10% of the brain and can unlock the remainder through training.
Brain activity continues at rest. Regions contribute differently across tasks; a scan highlighting a few areas often shows a contrast between conditions, leaving ongoing activity out of the picture.
Raichle and colleagues, 2001, measured baseline brain activity and task-related changes. The study identifies no unused 90% reserve.
A brain scan is not a gauge of the percentage of intelligence you are using.
Instinct, emotion and reason occupy three brains layered by evolution.
The model separates functions too rigidly and misrepresents evolution. Reptiles and mammals share ancestral structures; change involved cell diversification and circuit reorganization.
Tosches and colleagues, 2018, compared gene expression in reptilian brain cells, identifying evolutionary continuities and differentiated cell types.
The lizard-brain metaphor is memorable, but it does not anatomically explain an impulsive purchase.
Unscrambling sentences containing words associated with old age makes participants walk more slowly.
The slower-walking effect was not convincingly reproduced with automated timing. Experimenter expectations affected results in a subsequent test. This does not refute every form of priming.
Doyen and colleagues, 2012, used sensors to time walking and separately manipulated experimenter expectations.
A detail such as who holds the stopwatch can change the interpretation of a striking experiment.
Cells responding both to performing and observing an action could explain human empathy.
Cells with these responses exist. Finding them does not establish that they alone explain empathy, language or understanding intentions. Their role requires testing for each function.
Mukamel and colleagues, 2010, recorded 1,177 neurons in 21 patients observing and performing gestures. Some cells responded in both situations.
A cell responding when you see a gesture is not, by itself, an explanation of compassion.
The moment you hear shocking news is preserved with photographic precision.
Memories can remain vivid and convincing even as reported details change. Consistency across retellings is itself different from accuracy against the event.
Talarico and Rubin, 2003, followed 54 students after September 11. Consistency declined for both the news and an everyday event; confidence held up better for the former.
A person’s sincerity and certainty deserve respect without guaranteeing every remembered detail.
Hypnosis and suggestive questioning can reliably recover trauma completely blocked in the unconscious.
Suggestion can contribute to false memories. A memory recalled after a long interval can nevertheless have independent corroboration; its timing neither proves nor disproves it.
Loftus and Pickrell, 1995, studied a suggested memory of getting lost in a shopping centre. Geraerts and colleagues, 2007, found corroboration for some spontaneously recovered abuse memories.
Listening without leading questions protects both the person and the quality of the evidence.
Addressing minor disorder prevents serious crime.
The intervention matters. Programs solving local problems with communities show more favorable results than aggressive enforcement against minor offenses. These evaluations do not by themselves test the full theoretical mechanism.
The Braga, Welsh and Schnell review, 2019, finds crime reductions associated with community interventions, without a significant reduction for aggressive order maintenance.
Repairing a place and penalizing its users are different interventions with different evidence.
Fear of confirming a negative stereotype can lower test performance.
There are findings supporting the hypothesis, but effect size and generalizability are disputed. An analysis about girls and mathematics cannot settle the effect for every group, examination or form of discrimination.
Flore and Wicherts, 2015, analyzed 47 effect sizes among girls under 18. The average was small, with publication-bias indicators casting doubt on its estimate.
Uncertainty about a particular intervention does not make the experience of discrimination imaginary.
Differences in category-association speed could predict an individual’s discriminatory behavior.
The IAT measures relative associations under test conditions. Its relationships with behavior vary and do not justify a confident individual verdict about character or future actions.
Oswald and colleagues, 2013, find poor prediction. Kurdi and colleagues, 2019, find significant but heterogeneous associations in a broader synthesis. Conclusions depend on criteria and methods.
A score can open a conversation about associations; it should not become a moral label.
Income inequality generally harms mental health beyond personal income.
The recent synthesis does not support a universal negative relationship. Associations vary with context, including income and inflation; the result does not refute the health effects of poverty.
Sommet and colleagues, published online in 2025, combine 168 studies. The average mental-health association is near zero after publication-bias correction, but remains adverse in low-income samples.
Inequality between people and a person’s lack of money require distinct questions and measurements.
Financial concerns reduce attention available for other tasks.
Initial research found lower performance under certain conditions. Later results are uneven; the popular 13-IQ-point figure is not a fixed loss of intelligence caused by poverty.
Mani and colleagues, 2013, studied financial scenarios and farmers before and after harvest. Carvalho, Meier and Wang, 2016, found no cognitive differences before versus after payday.
Financial difficulty is not a diagnosis of someone’s intellectual capacity.
Messages hidden in a film automatically increase drink and popcorn sales.
Vicary’s story provides no credible evidence for the advertised sales increases. Laboratory research on subliminal stimuli is a separate question, with condition-dependent effects.
Karremans, Stroebe and Claus, 2006, reported an effect of a subliminal brand name on drink choice among thirsty participants.
A limited influence on a laboratory choice does not demonstrate covert control over everyday shopping.
A large assortment can discourage buying, as the comparison of 24 versus 6 jams suggested.
More options do not automatically cause paralysis. Studies find different results; a near-zero average can combine situations where variety helps and where it hinders.
Scheibehenne, Greifeneder and Todd, 2010, synthesized 50 experiments with 5,036 participants. The average effect was near zero, with substantial variation between studies.
For a difficult choice, clear criteria may matter more than arbitrarily shortening the list.
Contact between members of different groups can reduce prejudice.
Evidence supports an average reduction in prejudice, with variation across interventions. The favorable conditions proposed by the theory are not requirements proven necessary in every case.
Paluck, Green and Green, 2019, found 27 randomized studies with delayed outcomes. Effects on racial or ethnic prejudice were weaker, with gaps in evidence about adults.
It is worth testing which kind of encounter helps, whom it helps and for how long.
Poor performers overestimate themselves because they also lack the skills to recognize their mistakes.
The classic self-assessment pattern can be amplified by measurement error and regression to the mean. The contribution of a specific metacognitive mechanism remains disputed; popular graphs about experts and fools exceed the evidence.
Gignac and Zajenkowski, 2020, tested measured and self-assessed intelligence in 929 people without finding the nonlinear pattern predicted by the classic explanation.
Compare your estimate with a checkable result. Labeling an opponent does not test your own judgment.
Five dimensions describe much of the variation in personality.
The model is useful and well studied, particularly in literate populations. Its dimensions are continuous, and their structure and measurement do not automatically transfer to every culture.
Gurven and colleagues, 2013, did not reproduce the standard structure in responses from Bolivia’s Tsimane population. That limitation matters when presenting an instrument as universal.
A profile describes tendencies; it neither determines your future nor confines you to a personality type.
Spacing study sessions and trying to retrieve material from memory can improve long-term retention.
These are two distinct practices with supporting evidence. Benefits depend on material, timing and assessment; rereading can feel easier and help immediately without offering the same later advantage.
Roediger and Karpicke, 2006, found a testing advantage on delayed assessments. Cepeda and colleagues, 2006, synthesized evidence on distributed study.
Close the material, try to explain what you remember, check your answer and return after an interval.
Under certain conditions, short-term memory holds approximately four information units.
Capacity is limited, but the number depends on the task, grouping and prior knowledge. Four is a useful estimate under controlled conditions, not a universal ceiling on thought.
Cowan, 2001, reexamines research and proposes about four units when rehearsal and long-term-memory support are restricted.
Write down task steps and group information meaningfully. An external list reduces the burden of holding everything in mind.
Naming ink color becomes harder when the printed word names a different color.
In fluent readers, word information can interfere with the required response. The difference depends on the task version; the test alone neither localizes a brain problem nor makes a diagnosis.
Stroop, 1935, compared reading and color naming with and without conflicting verbal information. Incompatible words slowed color naming.
Hesitating over RED printed in blue illustrates ordinary competition between responses.
Cognitive behavioral therapy can reduce depression symptoms compared with control conditions.
Randomized trials support efficacy. Benefits vary between people and comparisons; superiority over other psychotherapies is unclear. Treatment includes behavioral practices as well as work on thoughts.
Cuijpers and colleagues, 2023, synthesize 409 trials with 52,702 patients. They find benefits against conditions including waitlist and usual care.
If a therapy does not help enough, the result deserves discussion and treatment adjustment; it is not a personal failure.
Negative information can attract more attention than positive information.
The asymmetry appears in studied contexts without being identical across stimuli, people or memory. A click effect establishes neither an evolutionary explanation nor how every algorithm works.
Robertson and colleagues, 2023, analyzed randomized Upworthy headline tests. For an average-length headline, an additional negative word increased click-through rate by 2.3% in relative terms.
A headline attracting you through fear is not necessarily more relevant or better supported.
Genetic differences contribute to variation in many human traits.
Twin studies support genetic contributions, with estimates depending on population, environment and model assumptions. Heritability does not say how much of a person is genetic or mean that a trait cannot change.
Polderman and colleagues, 2015, synthesize 2,748 publications on human traits, many nonpsychological. Twin pairs in the total partly overlap across measurements.
An estimate about differences within a population cannot determine a child’s future.
A starting value can pull subsequent numerical estimates toward it.
The effect occurs in many estimation tasks. In the classic demonstration, a wheel rigged to show 10 or 65 preceded estimates of African countries’ UN share. Influence is not inevitable for every person.
Klein and colleagues, 2014, found strong anchoring evidence in Many Labs 1, a project with 36 samples. Its protocol did not exactly repeat the wheel demonstration.
Before a negotiation, prepare an independent estimate and the criteria behind it.
When discovering a rule, people can favor examples matching their first hypothesis and miss alternatives.
The task shows a difficulty in hypothesis testing without proving that people always reject contrary evidence. Results also depend on how the problem is constructed.
Wason, 1960, used the 2–4–6 sequence and a hidden rule, not the four-card task. Only 6 of 29 participants found the rule without an earlier incorrect conclusion.
Look for an example that would distinguish your explanation from a plausible alternative.
Start with what participants did and what was measured. Then ask whether other researchers found something similar, how large the effect was, and how closely the situation resembles your life. A finding can stay interesting as its promise becomes more modest.
This is an editorial selection of 48 familiar experiments, theories, and practices. It is not a representative sample of psychology or a systematic review of the entire literature. We retained the original edition’s subjects and checked their claims against original studies, replications, meta-analyses, and archival research. The year in each case is a historical reference, not the date of its latest evidence.
Labels summarize our reading of the cited sources at the revision date. New research may change them. Sample sizes appear beside identified studies; a meta-analysis and a direct replication are different kinds of evidence. The therapy case describes research evidence; treatment decisions require assessment of a person’s clinical situation with a professional.
We replaced categorical verdicts, corrected references and attributions, and removed participant totals without an identifiable source. The lab now uses independent observations preserved across analyses and a separate check on fresh data. All 48 subjects remain available.
20Individual effort in teamworkReduced effort in groups is documented, but depends on task meaning, evaluation of contributions and expectations of colleagues.Supported
Maximilien Ringelmann / Bibb Latané (1979) · 1913
The starting claim
People can exert less effort when their contribution is pooled into a collective result.
What the research shows
Reduced effort in groups is documented, but depends on task meaning, evaluation of contributions and expectations of colleagues. Coordination losses are a separate explanation for poor group performance.
Evidence that changes the reading
Karau and Williams, 1993, synthesized 78 studies and identified many factors that change the effect. The article appeared in the Journal of Personality and Social Psychology.
What you can take away
A weak collective result does not prove that team members are lazy. Check whether each understands their contribution and whether the work can be coordinated.
Read the sources