The Online Disinhibition Effect: Anonymity Is Not the Reason

Almost everyone believes some version of this: people are worse online because they are anonymous, and if you took the masks off, they would behave.

It is one of the most-cited ideas about internet behaviour. It is also, on the experimental evidence, wrong - and the way it became conventional wisdom is a good case study in how a taxonomy turns into a fact.

The founding paper contains no data

Suler (2004) named the online disinhibition effect and proposed six mechanisms: dissociative anonymity, invisibility, asynchronicity, solipsistic introjection, dissociative imagination, and minimization of authority.

It is a conceptual essay. No sample, no measurements, no statistics. It has been cited thousands of times, and its framework organises much of the field, but Suler tested nothing. Everything downstream is an attempt to check what he proposed.

That is not a criticism of Suler, who was proposing rather than concluding. It is a warning about what happened next.

The experiments that manipulated anonymity found little

Lapidot-Lefler and Barak (2012) ran the direct test: 142 participants in 71 dyads, randomly assigned across a 2×2×2 design manipulating anonymity, visibility and eye contact, discussing a dilemma over instant messaging.

Lack of eye contact was the factor that mattered. It was significant on threats (F = 5.11, p < .05, η² = .04) and on self-reported flaming (F = 5.92, p < .01, η² = .04). Anonymity was non-significant on three of the four measures.

Two honest caveats. The remaining anonymity result - F = 2.87 on threats, reported as p < .05 - does not correspond to a two-tailed p below .05 at that degrees of freedom, so we treat anonymity as null rather than weak-but-real. And eight cells drawn from 71 dyads is roughly nine dyads per cell, with effect sizes around η² = .04. This study cannot bear much weight in either direction.

Rösner and Krämer (2016) manipulated anonymity (anonymous guest posting versus Facebook login) crossed with the aggressiveness of the surrounding comments, in 62 German football fans.

• Anonymity: H = 1.85, p = .179 - not significant.
• Group norm: H = 6.24, p = .012 - significant.

Aggressive expressions averaged 1.5 (SD 2.45) when the surrounding norm was aggressive against 0.36 (SD 0.56) when it was not. Again the sample is small and underpowered to detect a modest anonymity effect - but the norm effect showed up in the same underpowered study.

The best study in the field did not test anonymity at all

Cheng, Bernstein, Danescu-Niculescu-Mizil and Leskovec (2017) ran the most convincing experiment here - 667 participants on Mechanical Turk, randomised across a 2×2 of mood (an easy or a difficult quiz beforehand) and context (benign or troll-like seed comments).

Condition	Trolling rate
Positive mood, positive context	35%
Positive mood, negative context	49%
Negative mood, positive context	47%
Negative mood, negative context	68%

Negative mood raised the odds by 89%; negative context by 68%. Together they roughly doubled the baseline.

They then analysed 16.5 million posts from 865,248 users on a major news site. 26% of flagged posts came from users with no prior history of being flagged. And in a predictive model reaching AUC = 0.78, the discussion context alone reached 0.74 - mood and context outperformed a user’s own history of trolling.

The paper’s title states the finding: anyone can become a troll. Note carefully what it does and does not show. It is strong evidence against dispositional explanations - bad behaviour online is not mostly the work of a fixed population of bad people. It is not evidence about features of the medium. Mood and social context are situational, but an offline crowd has both.

Real names do not fix it

If anonymity were the mechanism, removing it should help. The largest study of this is a natural test.

Rost, Stahel and Frey (2016) analysed 532,197 comments across 1,612 online petitions on a German platform over three years.

Only 29.2% of commenters chose anonymity - most used their real names. 20.62% of comments contained at least one aggressive expression. And the non-anonymous commenters were the more aggressive ones.

Their interpretation is that this aggression is norm enforcement - people punishing perceived violations - and that using a real name increases the credibility of the sanction. Whether or not that account is right, the direction of the correlation is the opposite of the prediction.

South Korea ran the closest thing to a policy experiment, requiring real-name verification on large sites in 2007. Cho (2013) analysed 784,107 postings from 74,949 participants across the implementation and found a short-term drop in participation that normalised over time, with identified users showing more discreet behaviour. The frequently quoted figures for how much abusive content actually fell we could not verify against a primary source, so we are not printing them.

The classical theory was dismantled in 1998

The disinhibition framework inherits from deindividuation theory - the idea that anonymity in a crowd dissolves individual restraint.

Postmes and Spears (1998) meta-analysed it: 60 independent studies, 70 effect sizes, 4,714 participants.

Pooled effect: r = .09. Statistically reliable, small, and highly inconsistent (Q = 218.25, p < .001, effects ranging from -.49 to .56).

Anonymity to one’s own group: r = .03, non-significant. Anonymity to an out-group: r = .16.

Their conclusion is worth quoting in shape if not in full: the results showed little support for the occurrence of antinormative behaviour or for the existence of a deindividuated state at all. Deindividuated participants behaved more in line with the local norm, not less (β = -.40, p < .001).

The essential limitation: this literature is almost entirely pre-internet offline laboratory work - five studies before 1970, only six in the 1990s. It demolishes the theory the disinhibition framework inherits. It is not direct evidence about online behaviour, and it should not be used as though it were.

That reversal - anonymity increasing conformity to whatever norm is salient - is the core of the SIDE model, which proposes that anonymity obscures individual features and thereby raises the salience of shared group identity. On that account, anonymous spaces are antisocial only when the local norm is antisocial. Which is what Rösner and Krämer found.

The result that cuts the other way

Honesty requires reporting this one prominently.

Ma and Zhang (2026) meta-analysed the online media characteristics associated with cyberbullying: 97 studies, 248,316 participants, spanning 2004 to 2025. Of six characteristics examined, perceived online anonymity showed the strongest association with perpetration.

We could not verify the pooled correlation and are not printing it.

Why this does not overturn the argument, but does constrain it: the measure is perceived anonymity, self-reported, correlated with self-reported cyberbullying. That is a textbook common-method problem, and it is entirely non-causal - people who behave badly may well report feeling more anonymous afterwards.

Every randomised manipulation of actual anonymity in this literature found null or marginal effects. The large correlational literature finds anonymity is the strongest predictor. That contrast is the most interesting fact in the field, and it is the sort of divergence that usually means the correlational measure is picking up something other than what it claims.

What this does not establish

That anonymity never matters. The experiments are small - 62 and 142 participants across many cells. “No evidence anonymity matters” is defensible; “evidence anonymity does not matter” is not.

That the medium is irrelevant. Lack of eye contact was the one factor that reached significance, and eye contact is a feature the medium removes. That is a medium effect, just not the one everybody names.

That real-name policies are useless. The evidence is thin and mixed, and the Korean numbers we could not verify.

What follows

Stop reaching for anonymity as the explanation. It is the least supported of Suler’s six factors, and the one everybody quotes.

Look at norms and context instead. They won in the underpowered experiment, they won in the well-powered one, they won in the 16-million-post observational analysis, and they are what the 1998 meta-analysis found anonymity amplifying rather than dissolving.

Which means moderation is not mainly about identity. A space where the visible norm is hostile will produce hostility from people with real names attached, and the largest study of real-name commenting found exactly that. What changes behaviour is what the last ten comments look like.

And be careful about yourself. The mood finding is the least comfortable result here: a bad quiz beforehand raised trolling odds by 89% in ordinary people who had never been flagged for anything. The medium did not make them worse. The day did.

Online behaviour is one of several areas where the named cause and the measured cause come apart - see the persuasive design guide for the wider set, the filter bubble for another idea tested and found narrower than claimed, and left on read for what a small interface signal does between two people.