Correspondence bias, and the situation a chat leaves out
The named caution for reading behaviour without the situation that produced it, and what a chat export structurally cannot hold.
Correspondence bias is the tendency to infer a person’s lasting dispositions from behaviour that the situation they were in already explains. Daniel Gilbert and Patrick Malone set it out under that name in Psychological Bulletin in 1995. A chat export is close to the condition they describe: it keeps the behaviour and almost none of the situation.
You read what somebody wrote and you form a view about the kind of person who would write it. The experiments behind this page found that the view survives even when the reader has been told, plainly, that the situation left the writer very little room.
Why that matters for a conversation you can scroll is mechanical rather than moral. The messages are preserved. The hour that produced them is not.
The definition names behaviour the situation already explains
Gilbert and Malone open their 1995 paper with a definition: correspondence bias is “the tendency to draw inferences about a person’s unique and enduring dispositions from behaviors that can be entirely explained by the situations in which they occur”. The same sentence is the abstract on the paper’s PubMed record.
Two phrases carry the weight. “Enduring dispositions” means a lasting feature of a person rather than a passing state. “Entirely explained by the situations” means the behaviour did not need the disposition at all in order to happen.
The authors are blunt about their own title. They write that they “consider the correspondence bias to be something of a misnomer inasmuch as several different psychological mechanisms can give rise to the same general effect”. One name, four causes, as the next section sets out.
Lee Ross had named a neighbouring idea in 1977. In Gilbert and Malone’s account of that work, Ross used the term fundamental attribution error for the tendency to underestimate the power of situations, and his larger point was that the tendency belonged to the people taking part in the experiments and not only to the researchers studying them.
An essay someone was assigned to write, read as a belief
The experiment everyone cites came first. In Gilbert and Malone’s account, Ned Jones and Keith Davis published the first systematic model of dispositional inference in 1965, and two years later Jones and Harris set out to test one of its less subtle predictions: that observers infer dispositions from an unconstrained actor’s behaviour and make no such inference when the actor is entirely constrained.
Participants read essays that supported or opposed Cuba’s president, Fidel Castro. They were told either that the writer had been free to choose which side to argue, or that his debate coach had instructed him to defend a particular view. Where the writer had chosen, observers inferred strong attitudes, as predicted. Where the writer had been ordered, observers made similar inferences, “albeit much weaker”.
Jones and Harris called the result observer bias, and Gilbert and Malone quote their verdict on the handy explanation available at the time. Behaviour “tends to engulf the total field”, Jones and Harris wrote, and that “describes the results without really explaining them”.
The primary 1967 paper is behind a publisher that refuses automated requests, so the experiment is reported here from Gilbert and Malone’s description of it, which is a secondary account by two authors writing about a study they did not run.
Notice what was measured: ratings of a stranger’s attitude, made from one short essay, by participants who had been told how the essay came to exist. Nothing was measured about anybody reading a person they know.
Four separate causes wearing one name
The paper’s abstract lists four mechanisms that produce distinct forms of the bias: lack of awareness, unrealistic expectations, inflated categorizations, and incomplete corrections.
Later the authors restate the same four as reasons an observer draws an unwarranted inference about a person’s traits. The observer lacks awareness of the situation as it objectively is or as the actor construed it; or holds inappropriate expectations for how a person would behave in such a situation; or has let an awareness of the situation distort how the behaviour itself was perceived; or, in their words, they “lack either the motivation or the capacity to correct the trait inferences they may have spontaneously and effortlessly made”.
This is the part summaries drop, and it is the part that decides what you can do about it. The four causes are not four names for one slip. Each says that a different thing is missing, and only the first one is answered by better information.
Situations are often invisible, and often somewhere else
The first mechanism has the clearest consequence for anything written down. “To avoid the correspondence bias,” the authors write, “an observer must realize that a situation is playing a causal role in an actor’s behavior. But one can implicate situational forces as causes only when one is aware that such forces exist in the first place.”
They then say why that awareness is hard to come by outside a laboratory: “many situational forces are temporally or spatially removed from the behavioral episodes they constrain”. Social norms and a parent’s threat, in their example, “physically exist only in the brains of the people whose behaviors they are constraining, and nothing in the behavioral episode itself may bring these forces to the observer’s attention”.
Their worked case is a study by Ross, Amabile and Steinmetz, also from 1977, in which participants were assigned at random to be contestant or quizmaster in a mock game show. Quizmasters wrote questions from their own store of arcane knowledge, contestants mostly failed to answer them, and observers of the show concluded that the quizmasters were genuinely brighter. The observers, Gilbert and Malone write, could not see the “invisible jail” the contestants had been put in.
A chat export is behaviour with the day stripped off
Put those two halves together. The bias is at its strongest where the situation is invisible, or removed in time or space from the behaviour it produced. A chat export is a record in which the behaviour is complete and the situation is almost entirely absent.
The file holds turns, timestamps and words. It does not hold what either person had walked out of that afternoon: the row at work, the bad news, the three hours of sleep, the separate conversation that one of these messages is quietly answering.
None of that was removed from the file. It was never in it. On Gilbert and Malone’s first mechanism, a reader of the thread cannot begin the work a fair inference would need, because the forces that shaped the messages were somewhere else at the time.
That is why this is the most useful named caution for anybody reading a thread back, including anybody reading a written analysis of one. The limits of such a reading are set out in what a chat analysis cannot tell you, and the narrower problem of a single message landing wrong is covered in why text messages get misunderstood. A reader with little to go on also fills gaps deliberately, which is the subject of uncertainty reduction theory. Correspondence bias sits above all three, because it is about the step from a record of behaviour to a claim about a person.
The window you pick decides what counts as context
A written analysis of a conversation inherits that gap and narrows the input further, and the shape of the narrowing is worth stating exactly rather than in general.
Portraitor’s privacy policy says that before any part of your conversation is transmitted, your browser removes names, email addresses, telephone numbers, postal addresses and credentials and replaces them with anonymous placeholders, entirely on your own device, and that the list matching each placeholder back to a real detail never leaves that device. The policy is explicit about the effect: “This removes who is speaking. It does not remove what was discussed.”
The same page says that chat exports often contain photographs, videos and voice recordings, that the service never opens them and never transmits them, and that only the written messages are read.
The upload flow on Portraitor’s AI chat analyzer adds two more boundaries, and both are chosen by you. It asks which person in the conversation you want analyzed, and says that their messages are read while everyone else stays context. It then asks which time period, offering presets or a range you select yourself.
Read those four facts together as a description of what reaches the reading, and it is narrow by construction: the written turns of one person you picked, inside a window you picked, with the identifiers replaced on your own machine and the attachments never opened. Whatever situation sat outside that window was not dropped by the analysis. It was never in the file. What picking one person does is covered in analyzing your own messages, and what the masking step removes is covered separately.
This is a claim about architecture rather than about quality, and it cuts both ways. The same design that keeps identities on your device also keeps the surrounding day off it.
What the experiments measured, and what they left open
This literature measured ratings. Participants in a room judged a stranger from one essay, or watched a staged quiz, and their judgements were compared with what the situation should have led them to expect. That is enough to establish that observers draw dispositional inferences from constrained behaviour, and enough to show the inference is hard to suppress. It is not a measurement of how accurate any particular reading of any particular person is.
Gilbert and Malone say as much themselves. Asking at the end of the paper how capably people navigate their social worlds, and how accurate their understandings of those around them are, they answer: “We do not know. Nobody does.”
They also set two caveats against their own argument. An observer who underestimates the power of a situation does make a logically incorrect inference, but “those incorrect inferences need not be dispositional”. And the error runs the other way too: they cite Lepper, Greene and Nisbett on children overestimating how much a reward drove their own interest in a toy, and Strickland on participants overestimating how far a watchful supervisor was responsible for an employee’s honesty.
So the honest use of this is a small one. It is a reason to hold any reading of a thread loosely, and a reason to see that the missing context in a file is structural rather than an oversight by whoever read it. It is not a verdict about you, about anyone you were talking to, or about what either of you is like.
A different failure is worth keeping apart from this one. A description can feel exact while being about nobody in particular, which is the subject of why personality descriptions feel accurate. Correspondence bias is the opposite direction of error: a reading that is specific, and specifically unfair to a situation it never saw.
Frequently asked questions
Frequently asked questions
What is correspondence bias in simple terms?
It is inferring what somebody is like from behaviour that their situation already explains. Gilbert and Malone’s 1995 definition is the tendency to draw inferences about a person’s unique and enduring dispositions from behaviours that can be entirely explained by the situations in which they occur. Put plainly: giving a person the credit, or the blame, for something the circumstances were doing.
Is correspondence bias the same as the fundamental attribution error?
They overlap, and the two names were coined for different emphases. In Gilbert and Malone’s account, Lee Ross used fundamental attribution error in 1977 for the tendency to underestimate the power of situations. Their own 1995 paper uses correspondence bias for the inference of dispositions from situationally induced behaviour, and calls even that name “something of a misnomer”, because four different mechanisms produce the same effect.
Who ran the first experiment on it?
Jones and Harris, in 1967, in what became known as the attitude attribution paradigm. In Gilbert and Malone’s account of it, participants read an essay for or against Fidel Castro and were told either that the writer chose his side or that a debate coach had assigned it; attitudes were inferred in both conditions, much more weakly in the assigned one. The 1967 paper itself could not be opened from a public page for this article, so the study is reported here from that secondary description.
Does knowing about correspondence bias stop it?
Not on its own. The fourth mechanism in Gilbert and Malone’s list is incomplete corrections, and they describe observers who “lack either the motivation or the capacity to correct the trait inferences they may have spontaneously and effortlessly made”. The inference arrives first. Correcting it is separate work, and that work can fail for want of either effort or attention.
Can a chat analysis put the missing situation back?
No. Nothing in a conversation file records what either person had just been through, so no reading of that file can restore it. What a reading can do is describe patterns in the written turns, which is a smaller and different claim. Where that line falls is set out in what a chat analysis cannot tell you.
The messages survived and the day around them did not, and a reading of the first is not a reading of the second.
Your conversations, read the way you would read someone else’s.
Portraitor takes a chat export you already have, anonymises it in your browser, and returns a structured reading of how you come across. Identifying details are removed in your browser before anything is sent.
See how it works →- Gilbert, D. T., & Malone, P. S. (1995). The correspondence bias. Psychological Bulletin, 117(1), 21 to 38. PDF on Daniel Gilbert’s page at Harvard.
- PubMed record for Gilbert and Malone (1995), The correspondence bias, with the abstract.
- Portraitor, Privacy Policy.
- Portraitor, home page, the upload flow.