Contribution
Before it, cognitive psychologists had asserted, mostly from research on perception and memory, that people cannot observe higher mental processes. The paper supplies evidence for higher-order processes (evaluation, choice, inference, problem solving) and answers the question the bare assertion left open: if not from introspection, where do people’s fluent reports about why they acted come from? Its answer is that they come from causal theories about what kind of stimulus produces what kind of response, which an observer who has only a description of the situation holds as well. It then says when such reports will be right and when wrong.
Key insight
A report of why one responded is a judgment of plausibility, not a readout. It feels like introspection and is not. Polanyi’s “we can know more than we can tell” has a converse: people “sometimes tell more than we can know”. The proof is that observers who never met the stimulus predict the same effects that the subjects report, so the subjects need no “fount of privileged knowledge” (Bem’s phrase) to produce them.
Assumptions
- The target is process, not content. The paper grants that people know their own history, current sensations, focus of attention, and much about their emotions, evaluations and plans. What it denies is access to how a stimulus produced a response.
- “Awareness” is redefined. Correct report is not enough. The criterion proposed is a verbal report “which exceeds in accuracy that obtained from observers” given a general description of stimulus and response.
- Reports were taken soon after. Its own studies ask within seconds or minutes, so the claim is that such access as exists is insufficient at that interval, not that access is impossible under coaching or interruption.
- The situations were chosen to mislead. The authors say so: they picked stimuli they expected subjects to be wrong about, and add that their own predictions of which stimuli would matter were no better than the subjects’ reports.
Key results
Three conclusions (stated at the outset). People often cannot report accurately on the effects of stimuli on higher-order responses, and sometimes cannot report the stimulus, the response or that any process occurred. Their reports rest on a priori theories. When reports are right, it is because the theory happened to be right.
From the literature.
- Dissonance and attribution. Of the studies that measured both behaviour and verbal report at about the same time, most found no verbal-report difference at all (17 entries, one of them five experiments), and in most of the rest the behavioural effect was the more reliable. Where the correlation between report and behaviour was reported (two studies) or obtained by writing to the authors (three), it was near zero. Bem and McConnell’s and Goethals and Reckman’s subjects changed their attitudes and recalled their earlier attitudes as their new ones; in the busing study, pro-busing subjects converted by a confederate recalled having been more anti-busing than the anti-busing subjects recalled being.
- Interviews. In Nisbett and Schachter (1966), pill-attribution subjects took four times as much amperage, and 3 of 12 reported the attribution to the pill. Storms and Nisbett’s insomniacs fell asleep 28% faster (arousal pill) or took 42% longer (relaxation pill) and explained the change by exams, roommates and the day of the week. Aronson and Zimbardo, asked by letter, reported the same.
- Subliminal perception (W. R. Wilson 1975): tone sequences played to the unattended ear could not be recognized above chance but were liked more.
- Maier (1931). After the experimenter set a cord swinging, subjects solved the two-cord problem, usually within 45 seconds, and few named the hint. A useless hint given first was credited instead by every subject who received it.
- Latané and Darley (1970). Subjects denied that bystanders affected their helping, and observers predicted no effect.
Its own studies (subjects introductory psychology students unless noted).
- Word pairs. 81 men memorized pairs such as “ocean–moon”; the cues doubled target responses (“Tide”) from 10% to 20% (p < .001), and subjects almost never named the pairs as reasons. “Awareness ratios” ranged from 0 to 244%.
- Position. Passersby in shops chose among four nightgowns (378) or four identical pairs of stockings (52); the right-most was heavily favoured, by almost four to one over the left-most for the stockings. No one named position, and almost all denied it when asked.
- Anchoring (Borgida, Nisbett and Tversky, unpublished; 60 subjects). Subjects reported the same moderate use of the anchor whether it had had a large effect or none.
- Halo. A teacher shown warm or cold changed ratings of his appearance, mannerisms and accent (each p < .001). Subjects denied this; subjects in the cold condition said instead that disliking the attributes had lowered their liking of him, inverting the cause.
- Ineffective factors. Deleting passages from an Updike excerpt (152 subjects) changed rated impact by less than .10 on a 7-point scale, yet 86% said the crib passage had increased it. Noise from a power saw during a film (90 subjects) had no effect on ratings, yet 55% said it lowered at least one; for poor focus most reported correctly that it had no effect, which the authors call their only case of reasonably good accuracy. A “no permanent damage” reassurance (75 men) did not change predicted shock-taking, yet most said it did. In each case subjects not exposed to the factor predicted the same erroneous effect.
- Nisbett and Bellows (unpublished). Five factors in a job applicant’s portfolio were varied factorially for women subjects: attractiveness, academic credentials, a spilled cup of coffee, a past auto accident, and whether they expected to meet her. Mean subject and observer reports of each factor’s influence correlated .89 (liking), .84 (sympathy), .99 (intelligence) and .77 (flexibility). For intelligence, reports correlated .94 with true effects and observer predictions .98. For the other three, reports correlated −.31, .14 and .11 with true effects, and observers did no better or worse.
When reports will be wrong: when the report is removed in time from the process; for “mechanics of judgment” (order, position, contrast, anchoring); for context; for nonevents; for nonverbal behaviour; and when a small cause has a large effect. When right: when the influential stimuli are available and plausible and few plausible noninfluential ones are, as in verbal conditioning, and in judgment tasks where clinicians and stockbrokers apply rules they were taught.
Why we think we introspect: we confuse access to content with access to process; we report intermediate results (Maier’s professor’s “monkeys swinging from trees”) as if they were the process; feedback that would disconfirm our theories of ourselves is rare; and it is frightening to think we know our minds no better than an informed outsider.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | People’s reports of how stimuli affected their responses are often wrong: they miss effective stimuli and credit ineffective ones | strong for the situations studied | the reviewed literature and eight small studies, though the situations were chosen to produce error |
| C2 | Subjects’ reports match observers’ predictions from a description alone | strong | Nisbett and Bellows (r .77 to .99), the Updike, reassurance and word-pair studies, Latané and Darley |
| C3 | Reports come from a priori causal theories, not from access to the process | moderate | inferred from C2; an explanation consistent with the data, not tested against a rival |
| C4 | Reports are accurate where the culture specifies the stimulus–response rule | moderate | the intelligence judgment in Nisbett and Bellows; the Slovic and Lichtenstein review, interpreted |
| C5 | There may be no introspective access to higher-order process at all | weak | the paper itself shows only that such access as exists does not produce reports better than observers’ |
| C6 | Accuracy should fall with time, for order and context effects, for nonevents and nonverbal cues, and for small causes of large effects | weak | speculation offered as testable |
Concepts
- Higher-order cognitive process: the mediating process between a stimulus and an evaluation, judgment, choice, inference or behaviour; contrasted with perception and memory, and with content.
- A priori causal theory: a belief, held before the event, about what kind of stimulus produces what kind of response; from explicit cultural rules, implicit theories (“psychological implication”, Abelson; “causal schemata”, Kelley), observed covariation, or connotative similarity.
- Representativeness of a cause: judging a stimulus influential because it resembles the kind of thing that produces the response, by analogy with Tversky and Kahneman’s heuristic.
- Awareness (proposed criterion): report more accurate than an observer’s prediction from a general description.
- Causal theory controls: comparing a subject’s reported weights not with actual weights directly but with what the subject, or a peer, would have predicted beforehand.
Connections
It builds on Bem’s self-perception theory (1967, 1972), which had argued that people infer their own attitudes as observers do; on Kelley’s and Abelson’s work on causal schemata; and on Tversky and Kahneman’s representativeness and availability heuristics, which it applies to causal reports. It reads the dissonance and attribution literature against itself, through what those papers did not publish. Johansson, Hall, Sikström and Olsson (LIT-1417, NOTE-1398) carry the question from group-level comparisons to single reports shown false of the person who gave them.
Bearing on the record
- CASE-114, the imagined owner. The case’s note of 2026-10-23 quotes A5 that this paper shows “reported authorship and actual causal history can diverge”. It does, in a specific form: reports of why one responded diverge from what caused the response, and match what observers would guess. The paper is about the influence of stimuli on responses, not about attributions of agency or control; it never discusses who did something. So it supports the case’s premise, that a self-report of why one acted can be a construction from a theory, and is silent on the case’s thesis, that the content of a self-model changes error correction. Its causal-theory account suggests one prediction for the case: a system whose self-model is centralized should explain its errors as an observer holding the same model would, whatever its actual division of labour.
- NOTE-1398 (choice blindness). NOTE-1398 says the gap between what people do and the reasons they give was shown before only in group-level patterns, citing this paper. That is right for the evidence, since every comparison here is of condition means. The paper’s interview data are individual (the insomniacs’ explanations, Maier’s professor), but none of them can be shown false of the person who gave it.
No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Selected situations. The authors chose stimuli they expected subjects to misjudge. The paper shows that reports fail often, not how often in daily life.
- Unpublished and small studies. The studies of its own are reported briefly, several with no test statistics. Nisbett and Bellows is an unpublished manuscript, and the paper gives its correlations with no number of subjects.
- Necessary versus sufficient causes. The authors grant that an effective stimulus may have been necessary but not the main cause, so a subject who named another cause may not have been wrong; and that demand or embarrassment may explain some denials (position, noise).
- The causal-theory account is not tested against a rival. That subjects and observers agree is consistent with both using the same theory; it does not exclude subjects having some access that happens to agree with the theory.
- Errors and slips. The awareness ratio as defined (reporters over an estimate of the influenced) can exceed 100%, as its reported 244% does; the text does not say so. The text cites Kornhauser and Lazarsfeld as 1935 and the reference list as 1955 (the 1955 being a reprint in The Language of Social Research). The reference list gives Storms and Nisbett (1970) as JPSP volume 2; the article is in volume 16, issue 2, so the list gives the issue for the volume.
Open questions
- Does any procedure yield reports of process that beat an observer’s prediction? The paper’s criterion makes this the test and runs it only on judgments where both fail or both succeed.
- Do reports become more stereotyped with delay, as it predicts?