Does Wearing a Costume Actually Change How People Behave?

Yes, being unidentifiable changes behaviour — but the change is not "restraint falls off." What the evidence supports is narrower and stranger: anonymity loosens a person's grip on their personal rules and tightens their grip on whatever rule is loudest in the room, which is why the same manipulation that tripled candy theft in one Halloween experiment made other people less aggressive in others.

That condition is the whole story. Costume, uniform, mask, throwaway username: none of them carry a moral direction of their own. They change which audience a person is performing for.

The Halloween experiment everyone half-remembers

On Halloween in the mid-1970s, Ed Diener, Scott Fraser, Arthur Beaman and Roger Kelem set up 27 homes across Seattle and waited. Their paper, "Effects of deindividuation variables on stealing among Halloween trick-or-treaters," was published in the Journal of Personality and Social Psychology in 1976, and it observed 1,352 children who arrived to trick-or-treat between 5:00 p.m. and 9:00 p.m.

At each house a woman greeted the children warmly, told them they could take one sweet each, and then said she had to go back to her work in another room — leaving a bowl of sweets and, in some conditions, a nearby bowl of coins. A hidden observer recorded what happened next.

Three things were varied. Whether the children had been asked their names and where they lived (nonanonymous) or not (anonymous). Whether a child was alone or in a group. And, in a third set of conditions, whether one child in the group was singled out and made responsible for whatever the group did.

The numbers in the paper's Table 1 are the ones worth carrying around:

- Nonanonymous, alone — 7% took extra (40 children)

- Nonanonymous, in a group — 20% took extra (384 children)

- Anonymous, alone — 21% took extra (42 children)

- Anonymous, in a group — 57% took extra (297 children)

Percentage of trick-or-treating children who took more than they were told they could, by condition. Source: Diener, Fraser, Beaman & Kelem, "Effects of deindividuation variables on stealing among Halloween trick-or-treaters," Journal of Personality and Social Psychology, 33(2), 1976 — 1,352 children observed at 27 Seattle homes on Halloween night.*

The altered-responsibility conditions went further. When nobody in the group was anonymous, 10% transgressed (57 children). When only the nominated "leader" was identified and the rest were anonymous, 27% did (139 children). When the whole group was anonymous and one child had been made answerable for them, 80% did (80 children).

That last row is the one that should unsettle the simple reading. It is not anonymity alone. It is anonymity plus a diffused, misdirected sense of who is on the hook.

Why "anonymity makes people worse" was always too neat

The obvious conclusion — mask on, conscience off — is the one that stuck, and it is the one the field spent the following decades failing to confirm.

Philip Zimbardo had proposed the mechanism in 1969: immersion in a group plus anonymity produces a deindividuated state in which self-awareness drops and behaviour stops being governed by internal standards. It is an elegant idea and it explains a lot of folklore about mobs, hoods and uniforms. It also made a prediction that kept coming out backwards.

In Zimbardo's own work, as the Encyclopædia Britannica entry on deindividuation written by social psychologist Karen M. Douglas records, soldiers in one study "gave electric shocks of shorter duration when they were deindividuated" — the opposite of what the theory says should happen. And in 1979, Robert Johnson and Leslie Downing ran the test that made the point unmissable: they dressed female participants in either Ku Klux Klan–style clothing or nurses' uniforms. The ones in nurses' uniforms "were significantly less aggressive."

Same anonymity. Opposite direction. The variable that moved was the costume's meaning.

What the rival explanation actually says

In 1995, Stephen Reicher, Russell Spears and Tom Postmes published "A Social Identity Model of Deindividuation Phenomena" in the European Review of Social Psychology — the model now usually called SIDE. Their argument was that the classic account rests on an individualistic idea of the self, "viewed as a unitary construct referring to that which makes individuals unique," and that this is the wrong unit.

People have more than one self available. There is the personal self, and there is the categorical self — you as a member of a group. SIDE proposes that anonymity does not erase selfhood; it shifts which self is salient. Strip away the individuating cues and what is left to orient by is the group and its norms. The paper's conclusion puts it bluntly: the classic paradigm of anonymity within a group, "far from leading to uncontrolled behaviour, maximizes the opportunity of group members to give full voice to their collective identities."

Three years later Postmes and Spears tested that against the accumulated evidence. Their meta-analysis, "Deindividuation and antinormative behavior," appeared in Psychological Bulletin in 1998 and pooled 60 independent studies. The findings were unkind to the original theory: "little support for (a) the occurrence of deindividuated (antinormative) behaviors or (b) the existence of a deindividuated state." What fitted better was situational: "groups and individuals conform more to situation-specific norms when they are 'deindividuated'."

Read the Diener result through that lens and it stops being about lost restraint. A child anonymous in a group of anonymous children is standing in a small temporary society whose only visible norm is the other children reaching into the bowl. The personal rule ("take one") is not the rule that is loudest.

This is a live disagreement, not a settled one. Deindividuation theory has defenders, the meta-analysis is nearly three decades old, and SIDE has its own critics. What is no longer defensible is the flat claim that anonymity releases a fixed quantity of badness.

The internet ran the experiment at scale, and got both answers

If anonymity were straightforwardly disinhibiting, comment sections should be the cleanest evidence in psychology. They are not.

Katja Rost, Lea Stahel and Bruno Frey analysed 532,197 comments on 1,612 online petitions on the German platform OpenPetition.de between May 2010 and July 2013, publishing in PLOS ONE in 2016. Their stated aim was "challenging the popular assumption that online anonymity is one of the principle factors that promotes aggression." Across that dataset, 20.62% of comments contained at least one aggressive expression — and the result was that "in the context of online firestorms, non-anonymous individuals are more aggressive compared to anonymous individuals." When the loud local norm is public moral outrage, having your name attached is not a brake. It is a reason.

The second piece comes from a natural experiment. Alfred Moore, Rolf Fredheim, Dominik Wyss and Simon Beste examined 45 million comments on the Huffington Post from January 2013 to May 2015, published in Political Studies in 2021. Over that window the site changed its identity rules twice: from what the authors call non-durable pseudonyms, to authenticated accounts in December 2013, to Facebook-based commenting in June 2014 — "approximating a 'real-name' environment." Cognitive complexity, their proxy for one aspect of deliberative quality, rose significantly in the middle phase and then fell after the move to real names.

Not anonymity, then, and not identification either. The thing that worked was a durable pseudonym — a persistent identity inside the forum, disconnected from the rest of a person's life. Something to build a reputation in. Something to lose.

What to do with this on a practical level

None of the above tells you who you are. It tells you which features of a situation tend to move behaviour, and those you can actually inspect. Before you join something — a costume party, a group chat, a pile-on, a new anonymous forum — run this:

1. Find the loudest norm, not the official one. Not what the rules say. What are the people already here visibly doing? In the Diener setup, the stated rule was "take one." The visible behaviour was everyone reaching in. Visible behaviour wins.

2. Check whether anyone is actually accountable. The 80% condition combined anonymity with a nominated scapegoat. "Someone else is responsible" is the most permissive arrangement in the whole study. Diffused responsibility with a named figurehead is not the same as accountability.

3. Ask what the costume means. A KKK robe and a nurse's uniform are both anonymising. They pointed behaviour in opposite directions. This applies to team colours, work uniforms, group handles and account aesthetics: the outfit is an instruction about which norms apply.

4. Prefer a durable name to a disposable one. If you want a space you can post in honestly without linking it to your job, the Huffington Post pattern suggests a persistent pseudonym is the better structure — not a throwaway, and not your legal name.

5. Notice when being identified is the accelerant. If a group's shared norm is punishing someone, your name attached is an asset, not a restraint. That is the Rost finding. Being visible does not make an action fair.

6. Give yourself one individuating cue. Saying your own name, seeing your own face, addressing someone directly by theirs — each reintroduces the personal self that anonymity was crowding out. It is not willpower. It is changing which self is in the room.

Take the quiz

We built a short quiz to go with this piece: six questions about situations you actually find yourself in — group chats, uniforms, comment threads, costume parties. It takes about two minutes, stores nothing, and asks for no email. It will not tell you what kind of person you are. It will tell you which of the conditions above your usual environments happen to contain.

Frequently asked questions

Did the Halloween study really find that masks make children steal?

Not quite. Diener and colleagues manipulated whether children were asked their names and addresses, not whether they wore masks — every child was in costume. Among the 1,352 children observed at 27 Seattle homes, 7% of identified children alone took extra, against 57% of anonymous children in groups. Anonymity and group presence both mattered, and they compounded. But the study measured one behaviour on one night in one city in the 1970s, and it describes rates across conditions, not a rule about any individual child.

So is deindividuation theory wrong?

It is contested, which is not the same thing. Postmes and Spears pooled 60 independent studies in 1998 and found little support either for the antinormative behaviour the theory predicts or for the deindividuated mental state it proposes. Situation-specific norms explained the results better. Plenty of researchers still find the framework useful for describing crowd behaviour. Mind Decoded's position is that the strong version — anonymity reliably releases antisocial impulses — does not survive contact with the evidence.

Why would being anonymous ever make someone behave better?

Because what anonymity amplifies is the group norm, and group norms are frequently prosocial. In Johnson and Downing's 1979 study, anonymous participants in nurses' uniforms were significantly less aggressive than those in Klan-style robes. Zimbardo himself recorded a study in which deindividuated soldiers delivered shorter shocks. If the salient identity implies care, restraint or professionalism, the anonymity pushes that way.

Does a work uniform change how someone behaves?

The research does not test workplace uniforms directly, so this is an extension rather than a finding. What the costume studies suggest is that a uniform does two jobs at once: it reduces individuating cues, and it supplies a ready-made set of norms to fall back on. Which way behaviour moves would depend on what the uniform means in that setting — to the wearer and to everyone watching.

Are anonymous accounts the reason comment sections are hostile?

The largest datasets say it is more complicated. Rost, Stahel and Frey found that in online firestorms on a German petition platform, non-anonymous commenters were more aggressive than anonymous ones. Moore and colleagues found comment quality on the Huffington Post peaked under durable pseudonyms and declined after the shift to real-name Facebook commenting. Design, norms and reputation appear to do more work than identifiability alone.

Is there anything here that predicts what a particular person will do?

No, and it is worth being explicit. Every figure in this article is a rate across a condition — a proportion of a sample, measured once. In the anonymous group condition of the Halloween study, 57% took extra, which means 43% did not. These are descriptions of how situations shift distributions, not forecasts about individuals, and nothing here diagnoses anybody.

Connect

Questions or thoughts? Reach out anytime.

Email

subscribe to our newsletter

hello@minddecoded.com

© 2026. All rights reserved.

Terms of Service Privacy Policy Disclaimer