- Description
- Princeton Theological Seminary
- Source
- Scan by NYPL · original page
- Licence
- PD Public domain
The fundamental attribution error
When someone else fails, we see their nature; when we fail, we see the circumstances.
In brief
- What: we attribute other people's behavior to their traits (“they're careless”), underestimating the weight of the situation (workload, instructions, tools, deadlines).
- Why: the person is visible, the situation is not; and for ourselves it is the opposite, because we know our own constraints.
- Detect: “they're not invested,” “she's sloppy” with no fact or context, an incident closed with a person's name.
- Avoid: look at the situation first, describe behaviors rather than character, ask “what would I do in their place?”
Definition#
Fictitious illustration
Fictitious illustration
Fictitious illustrationThe fundamental attribution error is the tendency to explain other people's behavior by internal causes (personality, motivation, competence) rather than by situational causes (instructions, constraints, pressure, tools, chance), even when the situation is enough to explain everything.
The term comes from Lee Ross (1977), who considered it “fundamental” because it underlies many other errors of social judgment. The phenomenon had been demonstrated ten years earlier by Edward Jones and Victor Harris (1967): participants read an essay for or against Fidel Castro's regime and had to guess the author's opinion. Even when they were told that the author had been ordered to defend that position, they still credited the author, in part, with the opinion expressed.
It is also called correspondence bias: we assume the act “corresponds” to the person. It has a counterpart regarding ourselves, the actor-observer asymmetry: we explain our own actions by the situation and other people's by their nature (see the self-serving bias).
The mechanism#
- The person fills the field of view. When we watch someone, it is them we see acting; the situation (deadlines, missing information, contradictory instructions) stays in the background, often invisible. Our own situation, we live from the inside.
- A trait-based explanation is economical. “They're careless” takes two words and predicts future behavior; reconstructing the context takes effort (see the halo effect, which plays on the same economy).
- Insufficient correction. We start with the explanation based on the person, then adjust to account for the situation, but this adjustment takes attention: under pressure or overload, it does not happen.
- The just world. Believing that people get what they deserve is reassuring; blaming failure on the person avoids asking whether the same situation could happen to us.
In the experiment by Ross, Amabile and Steinmetz (1977), students were randomly assigned to be “questioners” or “contestants” in a general-knowledge quiz game. The questioner wrote their own questions, which gave them an obvious advantage, and the contestant had to answer. Observers, like the contestants, nonetheless judged the questioner more knowledgeable. Everyone had seen that the role was assigned at random, and nobody took it into account.
Where it shows up#
In project management#
- Incident review: after a failed production release, the report concludes with “an error by the on-call engineer.” Nobody asks why the procedure allowed this error, or why they were alone at night.
- Code review: “they write dirty code” is read as a trait, whereas the code was written on the eve of a delivery, with no style guide and changing specifications.
- Delays: a supplier or a neighboring team “doesn't keep its commitments,” when it is facing contradictory priorities that you cannot see. When it is your team that is late, you immediately cite dependencies and scope changes.
- Bugs and acceptance testing: “the tester wasn't rigorous enough,” without looking at the test environment, the available data or the planned coverage.
In management#
- Feedback: “you lack rigor” (a trait, which one can only endure) rather than “Tuesday's report didn't list the decisions” (a behavior that can be changed).
- Annual review: an employee whose team changed scope three times is judged “inconsistent.” The context, poorly recorded, has vanished by the time ratings are given.
- Conflict management: each person explains the other's behavior by their temperament (“they act in bad faith”) and their own by the situation (“I had no choice”). Both stories are symmetrical and irreconcilable.
- Remote work: a colleague who answers slowly is judged “not engaged”; we ignore that they have three back-to-back meetings, a different time zone or a faulty connection.
- Hiring: a candidate who is tense in an interview is judged “naturally stressed,” when the interview itself is a stressful situation.
Famous cases#
The Good Samaritan at Princeton: haste rather than virtue 1973#
Bias at play: fundamental attribution error (helping behavior is read as a character trait).
What happened. John Darley and Daniel Batson recruited students from the Princeton Theological Seminary. Each had to go from one building to another to give a talk, some on the parable of the Good Samaritan, others on an unrelated topic. On the way, they were told they were early, on time or late. In an alley, a man slumped over, coughing and groaning, lay in their path (an accomplice of the experimenters). The hurry condition weighed heavily on behavior: 63% of students who were not in a hurry helped, versus 10% of those in a great hurry. The topic of the talk (having thought about the parable) had no significant effect.
Who stops to help?
Loading the chart…
View the data
| Not in a hurry | Intermediate hurry | In a great hurry | |
|---|---|---|---|
| Not in a hurry | 63 | - | - |
| Intermediate hurry | - | 45 | - |
| In a great hurry | - | - | 10 |
What this illustrates
An observer who saw a hurried student walk past the man on the ground would say “they're heartless.” That is the bias's reading: the same student, a few minutes earlier, would have stopped. An incident review that concludes “they didn't pay attention” often describes the same thing: a person under pressure, in a situation that allowed the error. The study has a small sample and does not establish that character does not matter; it shows that the situation matters far more than we think.
The questioners in the general-knowledge quiz 1977#
Bias at play: fundamental attribution error (the role is mistaken for competence).
What happened. The setup described above is a miniature model of what happens in an organization: someone occupies a role that gives them a visible advantage (running the meeting, presenting their results, asking the questions) and is credited with superior skills. In the study, observers and contestants judged the questioner more knowledgeable than the contestant, although the advantage came from the role alone.
What this illustrates
A manager who chairs, an architect who leads the review or an expert who answers in committee seems more competent than those who have to answer. Do not confuse position with competence: when evaluating, ask what the role allowed them to show, and what it prevented the other person from showing.
Warning signs#
- An explanation starts with “they're…” or “she's…” (careless, lazy, disorganized) and cites no dated fact.
- The incident report stops at a name and the action list contains nothing about the process or the tool.
- We excuse our own delay by the context, and another person's by their character, in the same sentence or the same week.
- The feedback cannot be followed by an action: you don't “fix” a character trait.
- The same type of error occurs with different people, but each is treated as an individual case.
- You discover the context after the fact (“I didn't know they had these constraints”) and you change your mind about the person.
Habits to avoid it#
- Look at the situation first. Before explaining a behavior by the person, list what the situation could have produced: workload, instructions, available information, tools, deadlines, incentives.
- Put yourself in their place. “What would I have done with the same information, the same workload, the same deadlines?” If the answer is “the same thing,” the cause is in the situation.
- Test interchangeability. Could another competent person, in the same situation, have made the same error? If so, fix the system.
- Describe behaviors, not characters. “In Thursday's review, three tests were not rerun” rather than “you're sloppy.” The situation, behavior, impact model helps.
- Ask for their version before concluding. An open question (“what happened on your side?”) brings the context you don't have.
- Blameless incident reviews. Set as a rule that the root cause is never “a person”; drill down with five whys until you reach a cause the organization can change.
- Record the context as you go. At the time, note the constraints of a period (scope change, absence, incident). The year-end evaluation will use them.
- Apply the same lens to yourself. The symmetrical question: “if it were me, what explanation would I give?”
With AI: detect, fix, mitigate#
AI is useful for spotting wording that judges the person and for rephrasing in terms of facts. It does not know your team's context: you must supply the circumstances, and a language model often reproduces the stereotypes in its data (Echterhoff et al., 2024). It replaces neither the conversation with the person nor your knowledge of the field.
Detect#
- Spot, in an incident review, an evaluation or meeting minutes, character judgments with no supporting fact.
- Flag root causes that stop at a person.
- Compare the explanations given for the same type of event depending on whether “we” or “they” caused it.
Fix#
- Rephrase feedback as observable behavior, impact and request.
- Propose a list of situational causes to check before concluding.
- Prepare open questions to gather the person's version.
Mitigate#
- Have an anonymized evaluation reread to spot character adjectives.
- Have the factual timeline of an incident written up before any analysis of responsibility.
- Use an incident review template that requires a “context factors” section.
Prompts to build into your routines#
1. From character judgment to observable fact. To use before feedback or an evaluation, on an anonymized text.
Here is a draft of feedback (first name removed): [text].
1. Identify every expression that judges a character trait rather than a behavior (adjectives, “always,” “never,” “they're…”).
2. For each, ask me for the dated fact that justifies it, or point out that it has none.
3. Rewrite the feedback in this format: specific situation, observed behavior, observed impact, request.
4. List five situational factors I should check before sending it (workload, instructions, tools, deadlines, available information).
Assume nothing about the person: if information is missing, ask me.2. Timeline and context factors of an incident. For an incident review, before naming a root cause.
Here are the elements of an incident: [factual description, no names].
1. Build the timeline (who decided what, with what information, at what moment).
2. For each decision, indicate what a competent and careful person could have done in the same situation with the same information.
3. List the system factors that made the error possible or likely (procedure, tool, missing alert, workload, absent review).
4. Propose corrective actions aimed at the system, not at the person.
If one of your conclusions rests on an assumption, flag it as such.How to build them in#
- Before feedback: run the draft through prompt 1, then reread it aloud, asking whether the person can take an action from it.
- After an incident: prompt 2 serves as a framework for the discussion, not as a verdict; participants complete and correct the timeline.
- In evaluation reviews: have the character adjectives in a batch of evaluations pulled out so you can discuss the facts that support them.
Precautions#
- An evaluation text is personal data: anonymize it, and use only a tool authorized by your organization.
- AI can itself produce character-based explanations or stereotypes: reread its rephrasings.
- It does not know the context: a plausible but invented “situational cause” is not a fact. Verify with the person and with the records (tickets, history).
- Do not let AI conclude on individual responsibility.
Limits#
- Character matters: traits do predict part of behavior. The bias is not in invoking them, but in underestimating the situation. An employee who makes the same error in very different situations does pose a real problem of competence or engagement.
- The size of the effect varies across cultures: it seems more pronounced in individualistic Western societies than in several East Asian societies, where the situation is given more weight (work by Morris and Peng, 1994, and Choi and Nisbett, 1998).
- Some famous experiments (including the Princeton one) rest on small samples; their logic is sound, but they should be read as illustrations, not precise measurements.
- Suspecting the situation at all costs is a symmetrical error: it can excuse you from evaluating performance or from addressing a genuinely problematic behavior.
Sources and verification#
- Ross (1977), “The intuitive psychologist and his shortcomings: Distortions in the attribution process,” Advances in Experimental Social Psychology, vol. 10 — the term “fundamental attribution error.”
- Jones & Harris (1967), “The attribution of attitudes,” Journal of Experimental Social Psychology — the Castro essay experiment.
- Ross, Amabile & Steinmetz (1977), “Social roles, social control, and biases in social-perception processes,” Journal of Personality and Social Psychology — the quiz game.
- Darley & Batson (1973), “From Jerusalem to Jericho: A study of situational and dispositional variables in helping behavior,” Journal of Personality and Social Psychology — the Princeton study (percentages cited from the article).
- Morris & Peng (1994), “Culture and cause: American and Chinese attributions for social and physical events,” Journal of Personality and Social Psychology — cultural variations.
- Echterhoff et al. (2024), “Cognitive Bias in Decision-Making with LLMs” — cognitive biases of language models.
