The Bias That May Be Correct

You believe your customers want a particular thing. You may be right, and the reason you believe it is that you want that thing. This article is about how much of that reasoning is an error, and the answer turns out to be considerably less than fifty years of textbooks suggest.

The Verdict, Stated First

Five claims, in descending order of confidence.

One. The phenomenon is real and heavily replicated. A 1977 paper reporting four studies, a 1985 meta-analysis of 115 hypothesis tests, and a literature spanning five decades.

Two. A substantial part of it is correct inference, not bias. You are a member of the population you are estimating, so your own view is a genuine observation from it, and a summary of the literature describes the bias as seeming "statistically appropriate."

Three. How much is correct depends entirely on a quantity nobody reports. On our own arithmetic the rationally justified projection is 16.7 percentage points for someone who has observed nobody else and 2.2 points for someone who has observed twenty.

Four. A residual survives when the rational part is removed, on the title of a 1994 paper calling it ineradicable, though we did not obtain that paper.

Five. The field's own ten-year review says the effect is not one thing. It examines four theoretical perspectives and concludes that "no single explanation can account for the range of data."

Our Grades For These Claims

Applying the scheme from the first article in this series.

Grade A for the phenomenon existing, from a 1977 abstract fragment obtained from the publisher and a meta-analysis of 115 tests confirmed in three independent reference lists.

Grade A for the statistical objection being taken seriously, from a phrase in a peer-reviewed paper's own summary describing the bias as statistically appropriate.

Grade C for every effect size in this article. The 0.30 figure and a smaller modern figure both reach us through a practitioner website, and we could not verify either.

Grade A for the ten-year review's conclusion, obtained verbatim.

Grade A for our own arithmetic, which is Laplace's rule and reproducible, on assumptions we state.

Our position: the behaviour is well established, its status as an error is genuinely contested, and nobody reports the number that would settle it.

A Note On Method

Everything here is verified to August 2026.

We obtained a fragment of the 1977 abstract from the publisher[2] and the citation from four independent reference lists[3]. We did not obtain the paper or any of its four studies.

We obtained the 1985 meta-analysis's citation from three independent reference lists[3] and not its abstract, so its effect size reaches us only through a practitioner source.

The statistical objection reaches us as one phrase inside a bibliographic service's summary of a later paper[1]. We did not obtain the 1989 paper.

We obtained the 1987 review's abstract verbatim[4] and not the review.

Several substantive claims about what the objection showed and what a 1994 test found come from a practitioner website[5], flagged at every use, and we could not verify any of them against a paper.

All arithmetic is ours, uses a uniform prior we chose, and one of its results runs against the article's own framing, which we report rather than hide.

This article discusses research on social perception. It is not market research, product or hiring advice.

The 1977 Paper

The source.

Ross, L., Greene, D., and House, P. (1977), The "false consensus effect": An egocentric bias in social perception and attribution processes, Journal of Experimental Social Psychology, 13(3), 279–301[2][3].

The publisher's record gives the abstract's opening: "Evidence from four studies demonstrates that social observers tend to perceive a 'false consensus' with respect to the relative commonness of their own" responses, at which point our reproduction is cut off[2].

A later source states the definition plainly: false consensus is "the tendency to overestimate the extent to which one's opinions are also shared by others."[6]

Four observations, ours.

The authors put "false consensus" in quotation marks in their own title, which is a hedge worth noticing given how confidently the term is now used.

The subtitle calls it an egocentric bias, which asserts the error rather than demonstrating it, and the whole dispute below is about whether that assertion holds.

The paper reports four studies, which for 1977 is a substantial evidential base within a single paper.

And we did not obtain any of them, so everything we say about method comes from later sources.

Four Studies

The best-known of them, reported at the strength our sourcing allows.

A practitioner website describes "the Stanford 'Eat at Joe's' sandwich-board study from that paper" as having become "the textbook example"[5]. This is a practitioner website, not an academic source, flagged here and at every use, and we did not obtain the study.

Three observations, ours.

The design, as generally described, asks people whether they would do something conspicuous and then asks what proportion of their peers would. Both groups estimate their own choice as the more common one.

That structure is what makes the finding measurable. The two groups cannot both be right, so a gap between their estimates is evidence of something even without knowing the true proportion.

And it is also what makes the statistical objection possible, as the sections below set out. A gap between two groups' estimates is exactly what correct Bayesian reasoning would also produce, which is the whole dispute in one sentence.

The Meta-Analysis

The quantitative synthesis, whose citation we confirmed and whose contents we did not obtain.

Mullen, B., Atkins, J. L., Champion, D. S., Edwards, C., Hardy, D., Story, J. E., and Vanderklok, M. (1985), The false consensus effect: A meta-analysis of 115 hypothesis tests, Journal of Experimental Social Psychology, 21, 262–283[3].

We obtained the citation from three independent reference lists and not the abstract.

Three observations, ours.

115 hypothesis tests is a substantial base and the number is in the title, so it is not in doubt.

Seven authors on a meta-analysis is unusual and suggests a coding effort rather than a desk exercise.

And we cannot tell you what it found, which for the central quantitative source in this article is a serious gap and the reason the next section exists.

A Figure We Could Not Verify

The effect size, and we are going to be unusually explicit about where it comes from. Ours.

A practitioner website states that the 1985 meta-analysis "reported an average correlation of r ≈ 0.30", and adds that "modern reanalyses applying publication-bias correction (trim-and-fill, p-curve) produce smaller estimates around r ≈ 0.18 to 0.22." It also states that "gaps between groups in their estimates of peer prevalence typically run 10 to 25 points."[5]

Four observations.

The 0.30 figure is attributed to a named meta-analysis and is plausible, and we could not check it because we did not obtain that meta-analysis.

The 0.18 to 0.22 figure is attributed to "modern reanalyses" with no citation whatever, and we could not find the reanalyses it refers to. We report it as an unverified claim rather than a finding.

The 10 to 25 point range is likewise uncited, and we use it below only as an input to our own arithmetic, clearly marked.

And we are labouring this because the sixty-ninth article made the same point about a different literature. A specific number in a confident source is not evidence, and this article's quantitative claims are the weakest part of it.

The Statistical Objection

The argument that reframes the whole finding.

A bibliographic service's summary of a 1994 paper states: "Consensus bias is the overuse of self-related knowledge in estimating the prevalence of attributes in a population. The bias seems statistically appropriate (Dawes, 1989), but according to the" at which point our reproduction is cut off mid-sentence[1].

The reference is Dawes, R. M. (1989), Statistical criteria for a truly false consensus effect, Journal of Experimental Social Psychology, 25[7]. We did not obtain it.

Four observations, ours.

The phrase "seems statistically appropriate" appears in a peer-reviewed paper describing the state of the argument, which is much better sourcing than a commentary would be.

The 1989 title contains the crucial word. "Statistical criteria for a truly false consensus effect" asks how you would tell a real bias from correct inference, which presupposes that the standard measure does not.

And the truncation is unfortunate in a specific way. The sentence continues "but according to the", which is where the 1994 authors state their counter-argument, and we do not have it.

The practitioner source describes the objection: that classical designs "did not statistically separate the rational Bayesian component from the biased component" and that "the size of the residual after removing the rational part is much smaller than the original effect"[5]. Flagged as a practitioner source and unverified.

You Are A Member Of The Population

The objection stated in plain terms, because it is simple once seen. Ours.

Four observations.

When you estimate what proportion of people hold a view, you are estimating a population parameter from a sample. Your own view is one observation drawn from that population.

It is a legitimate observation. There is nothing improper about including yourself in a sample of which you are genuinely a member.

So someone who holds view X and adjusts their estimate upward is updating on evidence, and the only question is whether they update by the right amount.

And that reframes the entire literature. The question is not whether people project but whether they over-project relative to a benchmark, and the classical studies, on the objection, did not compute that benchmark.

Two things follow that are worth stating before the arithmetic. The objection does not deny the data. Nobody disputes that the two groups give different estimates; the dispute is entirely about what the difference means.

And the objection has a shape this series has now seen many times. A finding is established by comparing behaviour against an implicit normative benchmark, and the objection is that the benchmark was wrong. The fifty-third and sixty-fifth articles reported the same structure in two other literatures, and in both cases the finding survived in reduced form.

Sixty-Seven Percent

The benchmark, computed. Our own arithmetic, using a uniform prior we chose.

With a uniform prior over the population proportion, the posterior mean after observing s agreeing cases out of n is (s plus 1) divided by (n plus 2), which is Laplace's rule of succession.

For someone who holds view X and has observed nobody else: n equals 1, s equals 1, and the posterior mean is two thirds.

Four observations.

A perfect Bayesian who knows only their own view should estimate that 66.7 percent of people agree with them. Not 50 percent.

That is 16.7 percentage points of apparent false consensus which is not false at all. It is the correct answer to the question asked.

The size of that number is what makes the objection serious rather than technical. A third of the distance from an even split to certainty is rationally justified from a single observation.

And the uniform prior is our choice. A person with a genuine prior belief about the population would update differently, and no source we obtained specifies what prior the classical studies assumed, because on the objection they assumed none.

One more property of that number is worth noticing, because it is what makes the objection bite. Two thirds is not a small correction that could be waved away as noise. It is a third of the distance from an even split to certainty, produced by a single observation and no bias whatsoever.

And it is also why the classical paradigm looks the way it does. Comparing two groups who hold opposite views doubles the effect, so a rational 16.7 point excess on each side appears as a 33.3 point gap between them, which is larger than most reported effects. The design amplifies exactly the quantity the objection says is legitimate.

How Fast It Washes Out

The part that makes this usable. Our own arithmetic, assuming the truth is an even split and that any others you have observed are evenly divided.

Having observed 0 others, your rational estimate is 66.7 percent, an excess of 16.7 points.

Having observed 2: 60.0 percent, an excess of 10.0 points.

4: 57.1 percent, excess 7.1. 10: 53.8 percent, excess 3.8. 20: 52.2 percent, excess 2.2. 50: 50.9 percent, excess 0.9. 100: 50.5 percent, excess 0.5.

Three observations.

The justified excess decays roughly as one over the number of observations, which is the standard shape and means it disappears quickly.

By twenty observations the rational projection is worth about two percentage points, which is far below any reported effect.

So the objection's force depends entirely on how much other evidence the person had, and that single quantity determines whether the observed effect is mostly rational or mostly bias.

What Remains After Subtraction

Putting the two together. Our own arithmetic, using a reported range we could not verify.

The Bayesian gap between two people holding opposite views is twice the excess above. Subtracting it from an observed gap of 10, 17 or 25 points:

Having observed 0 others, the Bayesian gap is 33.3 points, and the residuals are negative 23.3, negative 16.3 and negative 8.3 points.

Having observed 2: Bayesian gap 20.0, residuals negative 10.0, negative 3.0, positive 5.0.

Having observed 5: gap 12.5, residuals negative 2.5, 4.5, 12.5.

Having observed 10: gap 7.7, residuals 2.3, 9.3, 17.3.

Having observed 20: gap 4.3, residuals 5.7, 12.7, 20.7.

Reading That Table Honestly

Including the row that goes against this article's own framing. Ours.

Four observations.

The top row produces negative residuals, meaning that if the people in these studies had genuinely observed nobody else, the reported effects would be smaller than correct Bayesian reasoning requires. On that reading people under-project rather than over-project.

We do not believe that reading and we report it because the arithmetic produces it. Participants in these studies are adults with years of social experience, so the realistic rows are the lower ones, where the Bayesian gap is small and most of an observed effect is genuine bias.

But the top row is not useless, because it establishes the shape of the argument. There exists an evidence level at which the observed effect is rationally justified, and the whole dispute is about whether real people are above or below it.

And the honest summary of our own table is this. Somewhere between five and ten prior observations, the Bayesian gap falls below the reported effects, and above that level the residual is real and grows. Everything turns on a number no source we obtained reports.

The Truly False Consensus

What survives the subtraction, from a title we obtained and a paper we did not.

Reference lists identify Krueger, J., and Clement, R. W. (1994), The truly false consensus effect: An ineradicable and egocentric bias in social perception, Journal of Personality and Social Psychology, 67(4), 596–610[3].

A bibliographic service records that a paper on the topic sought "to detect truly false consensus effects (TFCEs)" and that "the false consensus effect involves adequate inductive reasoning and egocentric biases."[1]

We did not obtain the paper and report the title and those phrases.

Four observations, ours.

The phrase "adequate inductive reasoning and egocentric biases" is the settlement, and it is a both-and rather than an either-or. The effect contains a correct component and an incorrect one.

The word "ineradicable" in the title is a strong claim and it is the authors' own. Whatever remains after removing the rational part apparently does not go away.

The word "truly" is doing the technical work, and it exists because of the 1989 objection. The field had to invent a term for the part that is actually an error, which tells you how seriously the objection was taken.

And the effect being ineradicable is compatible with it being small. A residual that resists correction and is worth a few percentage points is a different proposition from the headline effect, and we obtained no magnitude for it.

What Base Rates Did To It

The one experimental test of the objection we found described, and we can report it only at low grade.

A practitioner website states: "Krueger and Clement (1994) tested this directly. They gave participants explicit base-rate information about peers' choices on a target item, then asked for estimates. Some of the FCE survived, the 'truly false consensus', and some of it disappeared. The residual is real but smaller than 1977 suggested."[5]

It adds a practical implication: "correction strategies that simply provide base-rate information work better than the original FCE literature implied. Show product managers actual user-research data, and a substantial part of their projection bias goes away. The remaining bias is real but tractable."[5]

This is a practitioner website, flagged at every use, and we could not verify any of it against the paper.

Four observations, ours.

If the description is accurate, the design is exactly right. Supplying the base rate removes the informational justification for projecting, so whatever projection remains is the error.

The reported result, that some survived and some disappeared, is what both sides of the dispute would predict, which is why it settles it.

The practical implication follows directly and is the most useful sentence in this article. Giving someone real data about their audience removes the part of the bias that was rational, because it was rational only in the absence of that data.

And we would not repeat the practitioner's framing without its caveat. We could not verify that this is what the 1994 paper did, and a reader relying on it should read the paper.

No Single Explanation

The field's own ten-year stocktake, obtained verbatim.

Marks, G., and Miller, N. (1987), Ten years of research on the false-consensus effect: An empirical and theoretical review, Psychological Bulletin[4].

Its abstract: "Ten years of research on the false-consensus effect (Ross, Greene, & House, 1977) and related biases in social perception (e.g., assumed similarity and overestimation of consensus) are examined in the light of four general theoretical perspectives: (a) selective exposure and cognitive availability, (b) salience and focus of attention, (c) logical information processing, and (d) motivational processes. The findings indicate that these biases are influenced by a host of variables and that no single explanation can account for the range of data. Instead, each theoretical perspective appears to have its own domain of application, albeit with some degree of overlap into other domains."[4]

Four observations, ours.

Perspective (c), logical information processing, is the rational account that the 1989 objection developed formally, and it is listed here two years earlier as one of four live explanations.

"No single explanation can account for the range of data" is the eighth time this series has recorded a literature saying that about itself, and by now we treat it as the expected finding rather than a surprise.

The phrase "each theoretical perspective appears to have its own domain of application" is more useful than it sounds. It says the mechanisms are real and situational, so which one operates depends on the setting.

And the review names selective exposure as a perspective, which is worth flagging for a commercial reader. If you mostly talk to people like you, your sample really is skewed, and projecting from it is a sampling problem rather than a psychological one.

A Third Explanation Entirely

Neither bias nor Bayesian updating, and it may be the most commercially relevant of the three.

A bibliographic service records that research "examined whether this 'false consensus effect' is partly due to people's failure to recognize that their choices are not solely a function of the 'objective' response alternatives, but of their subjective construal of those alternatives."[1]

We did not obtain the study and report the description.

Four observations, ours.

The claim is that two people answering the same question are not answering the same question. They have understood it differently, and each assumes their reading is the obvious one.

On that account the error is not about people at all. It is about the question, and the projection follows correctly from a mistaken belief that everyone read it the same way.

This is directly the survey problem a business faces. A customer answering "would you pay more for faster delivery" is answering a question they constructed, and their answer tells you about their construction as much as their preference.

And it points at a different remedy from the other two accounts. Not more observations and not less confidence, but asking what the respondent thought the question meant, which is a change to instrument design rather than to sample size.

The Minority Asymmetry

A consequence of the arithmetic that we have not seen drawn out, and it has an uncomfortable edge. Our own calculation, same assumptions as before.

With a single self-observation, the Bayesian estimate of your own side is 66.7 percent regardless of the truth. So the error depends entirely on which side you are on.

At a true split of 50/50: both sides overestimate by 16.7 points.

At 70/30: the majority holder understates by 3.3 points; the minority holder overstates by 36.7.

At 90/10: the majority holder understates by 23.3 points; the minority holder overstates by 56.7.

Four observations.

The crossover sits at a majority share of 66.7 percent. Below it both sides overestimate themselves; above it the majority begins to understate its own prevalence.

And the minority error is larger at every split, growing as the minority shrinks. The more unusual your view, the more badly this rule misleads you, from exactly the same reasoning that serves a typical person well.

The commercial reading is direct. A founder with an unusual preference is in the minority position on this table, and unusual preferences are common among people who start things.

And note that none of this requires anyone to be biased. Every figure above is what correct inference produces, which means the minority holder is being misled by a sound rule rather than by an error, and cannot fix it by thinking harder.

What Actually Survives

Our reading, stated directly.

Five statements.

People project their own views onto others, robustly and across five decades. That much is not in dispute.

A substantial part of that projection is correct inference. A summary in a peer-reviewed paper describes the bias as seeming statistically appropriate, and on our own arithmetic a single self-observation justifies a 16.7 point excess.

The justified part shrinks fast with evidence. By twenty prior observations it is worth about two points on our own figures.

A residual survives and is described as ineradicable, on a 1994 title, and we obtained no magnitude for it.

And every effect size in this article is unverified, reaching us through a practitioner website rather than a paper.

The Question Nobody Reports

The gap that would settle this, and it is a single number. Ours.

Four observations.

On our own arithmetic the whole dispute reduces to one quantity: how many other people's views the estimator had observed before estimating.

Below about five, the reported effects are smaller than a Bayesian would produce. Above about ten, most of the effect is residual. The crossover sits in a narrow band and nobody measures where a given study's participants were.

That is not a criticism of the classical studies, which were asking a different question. It is a criticism of how the finding is used, since a business reader is being told about a bias whose size depends on evidence they can actually change.

And it produces the most useful question in this article, which is not about psychology at all. Before treating your own projection as an error, ask how many customers you have actually asked. The answer determines whether you are biased or merely under-sampled.

Three Accounts, Three Remedies

Setting the explanations side by side, because they prescribe different actions and a business reader has to pick one. Ours.

Four observations.

The bias account says people over-weight themselves egocentrically. Its remedy is awareness and effort, which this series has repeatedly found to be the weakest kind.

The inference account says people are updating correctly on thin evidence. Its remedy is more evidence, and on our own arithmetic ten observations does most of the work.

The construal account says people misread how differently others understood the question. Its remedy is instrument design: ask what the respondent thought you meant.

And the 1987 review's conclusion is the reason to hold all three. Each perspective appears to have its own domain of application, so the right question is not which account is true but which one is operating in the decision in front of you.

Your Market Research

The first application. Ours, untested, and not market research advice.

Four points.

The commercial version of this finding is "our customers want X," where the evidence is that you want X. The literature says that reasoning is partly valid and the validity collapses as soon as you have real data.

On our own arithmetic the justified projection from a single observation is 16.7 percentage points, and from twenty observations about 2.2. Twenty conversations is not a survey and it removes almost all of the rational case for projecting.

Which reframes what small-sample research is for. Its value is not that twenty people are representative. It is that twenty people destroy the excuse for reasoning from one.

And the review's selective exposure perspective supplies the warning. If your twenty are all people you already know, you have not sampled the population; you have sampled your own neighbourhood of it.

Your Product Decisions

The second application, and the one with the most money attached. Ours.

Four points.

Every feature decision made without customer data is a projection by construction, because the only preference in the room is the builder's.

The practitioner source's version of the 1994 finding is directly on point: show people actual research data and a substantial part of the projection goes away, though we could not verify it.

The residual is the part to plan around. If the bias is ineradicable, then data reduces it and does not remove it, which argues for a decision rule rather than a discussion.

And the rule we would use follows from the sixty-seventh article. Write down what proportion of customers you expect to want the thing, before you look at the data, and compare. A recorded prior turns a bias into a measurable error.

Hiring And Culture

The third application, and the one where the mechanism compounds. Ours, and nothing here bears on the lawfulness of any employment practice.

Four points.

A founder estimating what motivates staff is projecting from a sample of one, and that one is unusually unrepresentative of employees by definition.

The compounding problem is selective exposure. People who share your outlook are more likely to have stayed, so your observations are drawn from a filtered population and each additional one confirms rather than corrects.

That is the sixty-sixth article's missing cell in a different guise. You observe the people who agreed enough to stay and never the distribution you were estimating.

And the remedy is the same shape as everywhere else in this series. Ask, record, and count, rather than reasoning from a sample you did not choose.

One specific version of that is worth stating because it is cheap and almost nobody does it. Ask departing staff what they think the median employee would say, not what they themselves think. The first question is about the population and the second is another observation of one.

And the same move works on the customer side. Asking someone what they believe most customers want elicits their estimate of the distribution, which you can compare against your own, and disagreement between two estimates of the same population is more informative than agreement between two preferences.

The Remedy That Works

Because this article has an unusually clean answer. Ours.

Four observations.

Most biases in this series resist correction because the person cannot see the error from inside. This one is different in a specific way.

A large part of it is a response to missing information, and missing information is the one thing a business can straightforwardly fix.

On our own arithmetic the fix is cheap and steep. Going from zero to ten observations cuts the justified projection from 16.7 points to 3.8, which is most of the benefit for very little effort.

And that is why we would rank this among the most tractable findings in the series. The residual may be ineradicable, but the majority of the effect is a data problem, and data problems have solutions.

Two cautions on that optimism, both of which follow from earlier sections. The construal account is not fixed by more data at all, since a hundred people misreading a question the same way you did produces a confident wrong answer rather than a corrected one.

And the minority asymmetry means the fix is least effective for the people who most need it. Someone holding an unusual view is furthest from the truth on our own table and is also least likely to encounter disconfirming observations in an ordinary week, so the ten observations that fix a typical person will not be drawn from a representative sample for them.

When Projection Is The Right Move

The section this article needs, because it would otherwise recommend never trusting your own judgment. Ours.

Four observations.

Projection is correct inference in the absence of better data, and that is not a grudging concession. It is the objection's whole point.

So a founder with no research who reasons from their own preference is doing something defensible, provided they hold it at the strength the evidence supports, which on our own arithmetic is about two thirds rather than certainty.

The error is not projecting. It is projecting with more confidence than one observation warrants, and continuing to project after data arrives.

And there is a case where you are the best available sample. If you are genuinely typical of your market, your preference carries real information, and the honest test is whether you can name three ways you differ from your customers. If you cannot, that is the finding.

How This Pairs With The Last Article

Two findings that look similar and are not. Ours.

Four observations.

The curse of knowledge says knowing something makes you misjudge what others know. False consensus says holding a view makes you misjudge how many share it.

The difference is that one is never defensible and the other often is. Your knowing a fact is not evidence that others know it, because you learned it for a reason. Your holding a preference is genuine evidence about the distribution of preferences, because you are a draw from it.

That asymmetry has a practical consequence. Correct hard against the curse of knowledge and correct proportionately against false consensus, and the second correction should shrink as your data grows.

And the sixty-ninth article flagged this connection as an untested inference of ours. Having now read the false consensus literature, we would state it more carefully: the two mechanisms are different, and only one of them has a rational component.

That correction is worth marking as such. The previous article speculated that the curse of knowledge should operate symmetrically on less-informed parties, on the grounds that the mechanism looked general. Reading this literature, we think that speculation conflated two things.

The symmetric mechanism is social projection, which is this article's subject and which is partly rational. The asymmetric one is the curse of knowledge, which concerns knowledge rather than preference and which has no rational component, because acquiring a fact is not a random draw from a population of fact-holders.

We would rather correct a speculation from one article ago than let it stand, and we note that the correction makes the earlier claim narrower rather than wrong. A less-informed party does project, and what they project is their preferences and not their ignorance.

What To Do

Count your observations before calling it a bias. On our own arithmetic the rationally justified projection is 16.7 percentage points from a sample of one and 2.2 from a sample of twenty.

Get to about ten observations. That cuts the justified projection from 16.7 points to 3.8, which is most of the available benefit for a very small effort.

Check whether your observations were selected. The field's own review names selective exposure as one of four mechanisms, and people who share your outlook are more likely to be in your sample.

Record your estimate before you look at the data. A written prior turns an unmeasurable bias into a measurable error, which is the same instrument this series has recommended in nine articles.

Expect a residual and plan for it. A 1994 title calls the surviving component ineradicable, so data reduces the effect rather than removing it.

Do not treat your own preference as worthless. Projection is correct inference in the absence of better data, and the error is holding it more firmly than one observation warrants.

Ask how you differ from your customers, and name three ways. If you cannot, your sample of one is less representative than you think and that is the finding.

Treat the published effect sizes here with suspicion. Every one of them reaches this article through a practitioner website rather than a paper, and one is uncited even there.

The Limits Of This Analysis

Several caveats matter, and the quantitative sourcing here is the weakest in this series after the previous article. This article discusses research on social perception and is not market research, product or hiring advice; nothing here bears on the lawfulness of any employment practice, and the applications are our own reasoning and untested. Everything is verified to August 2026. We did not obtain the 1977 paper, only a truncated fragment of its abstract from the publisher, and we report nothing about its four studies except through later sources. We did not obtain the 1985 meta-analysis, only its citation from three reference lists, so its effect size reaches this article solely through a practitioner website. That website's further claim that modern reanalyses produce smaller estimates around 0.18 to 0.22 is uncited even there, and we could not find the reanalyses referred to; we report it as an unverified claim. We did not obtain the 1989 paper that makes the statistical objection, and have one phrase describing it from inside another paper's summary, which is truncated mid-sentence at exactly the point where the 1994 authors state their counter-argument. We did not obtain the 1994 paper, and its account of what base-rate information did comes from the same practitioner website, unverified. All arithmetic is ours. It uses a uniform prior we chose, an assumption that the truth is an even split, and an assumption that any others observed are evenly divided; different assumptions move every figure. The residual table takes as input a 10 to 25 point range of observed gaps which is uncited in our source. And the top row of that table produces negative residuals, implying that under the stated assumptions the reported effects would be smaller than rational; we do not believe that reading, we explain why in the body, and we report it because our own arithmetic produced it.

Related Reading

Behavioural Finance The Song In Your Head The neighbouring failure, which has no rational component at all. No. 71 The Question This Series Has Been Avoiding People shown how a bias could have affected them insisted their judgment was objective anyway. After seventy articles on bias, this is the finding... Behavioural Finance Nine Out Of Ten Flags Are False The same Bayesian machinery, applied where people update too little rather than too much.

Frequently Asked Questions

What is the false consensus effect?

The tendency to overestimate the extent to which one's opinions are shared by others. It was named in a 1977 paper reporting four studies and has been replicated across five decades. The authors put the phrase in quotation marks in their own title, which turned out to be prudent.

Why might it not be a bias?

Because you are a member of the population you are estimating, so your own view is a legitimate observation from it. A summary in a peer-reviewed paper describes the bias as seeming statistically appropriate. On our own arithmetic, someone who holds a view and has observed nobody else should rationally estimate 66.7 percent agreement.

So how much of it is real bias?

It depends on a number nobody reports: how many other people's views the estimator had observed. On our own figures the justified projection falls from 16.7 percentage points at zero observations to 3.8 at ten and 2.2 at twenty. Above about ten observations, most of an observed effect is residual.

Does anything survive?

Yes. A 1994 paper's title calls the surviving component ineradicable, and a bibliographic summary describes the effect as involving both adequate inductive reasoning and egocentric biases. We obtained no magnitude for the residual, which is a serious gap.

How reliable are the numbers in this article?

The effect sizes are the weakest part of it. Every one reaches this article through a practitioner website rather than a paper, and one claim about modern reanalyses is uncited even there. Our own arithmetic is sound but rests on a uniform prior we chose and assumptions we state.

What is the practical fix?

Get real observations, and about ten of them captures most of the benefit. This is one of the more tractable findings in the series precisely because a large part of it is a response to missing information, and missing information is fixable. Record your estimate before you look at the data.

Should I ever trust my own preference?

Yes, at about the strength one observation supports, which is roughly two thirds rather than certainty. The error is not projecting; it is projecting too confidently and continuing after data arrives. A useful test is whether you can name three ways you differ from your customers.

References

1. Bibliographic service record for the 1977 paper, confirming the citation as Lee D. Ross, David Greene and Pamela House, Journal of Experimental Social Psychology, 1977, volume 13, pages 279–301, and carrying summaries of citing works, including one recording that consensus bias is the overuse of self-related knowledge in estimating the prevalence of attributes in a population, that the bias seems statistically appropriate citing Dawes (1989), and that according to the, at which point the reproduction is cut off mid-sentence; another recording that the false consensus effect involves adequate inductive reasoning and egocentric biases, and that the work sought to detect truly false consensus effects; and another recording that research examined whether the effect is partly due to people's failure to recognize that their choices are a function of their subjective construal of the response alternatives rather than of the objective alternatives themselves. Note: a bibliographic service record. Our only source for the statistical objection, which reaches us as one phrase inside a summary of a later paper and is truncated exactly where that paper's counter-argument begins. semanticscholar.org
2. Publisher record for Ross, L., Greene, D., & House, P. (1977), The "false consensus effect": An egocentric bias in social perception and attribution processes, Journal of Experimental Social Psychology, reproducing the opening of the abstract, that evidence from four studies demonstrates that social observers tend to perceive a false consensus with respect to the relative commonness of their own responses, at which point the reproduction is cut off. Note: the publisher's record. We obtained a truncated abstract fragment only and not the paper or any of its four studies. sciencedirect.com
3. Academic preprint reference list confirming Ross, L., Greene, D. and House, P., The "false consensus effect": An egocentric bias in social perception and attribution processes, Journal of Experimental Social Psychology, 13, 279–301 (1977); Mullen, B. and colleagues, The false consensus effect: A meta-analysis of 115 hypothesis tests, Journal of Experimental Social Psychology, 21, 262–283 (1985); Krueger, J. and Clement, R. W., The truly false consensus effect: An ineradicable and egocentric bias in social perception, Journal of Personality and Social Psychology, 67, 596–610 (1994); and Krueger, J., From social projection to social behaviour, European Review of Social Psychology, 18, 1–35 (2007). The 1977 citation including issue number is separately confirmed by two further independent preprint reference lists. Note: reference lists; citations only. We obtained none of the works named, including the meta-analysis whose effect size is the central quantitative claim in this article. arxiv.org
4. Repository record reproducing the abstract of Marks, G., and Miller, N., Ten years of research on the false-consensus effect: An empirical and theoretical review, Psychological Bulletin: on ten years of research on the false-consensus effect and related biases in social perception, including assumed similarity and overestimation of consensus, being examined in the light of four general theoretical perspectives, namely selective exposure and cognitive availability, salience and focus of attention, logical information processing, and motivational processes; on the findings indicating that these biases are influenced by a host of variables and that no single explanation can account for the range of data; and on each theoretical perspective appearing to have its own domain of application, albeit with some degree of overlap into other domains. Note: a repository record. We obtained the abstract verbatim and not the review. researchgate.net
5. Practitioner website guide to the false consensus effect, recording that Lee Ross, David Greene and Pamela House published the foundational paper in 1977; that the Stanford sandwich-board study from that paper became the textbook example; that the 1985 meta-analysis of 115 studies reported an average correlation of approximately 0.30; that modern reanalyses applying publication-bias correction produce smaller estimates around 0.18 to 0.22; that gaps between groups in their estimates of peer prevalence typically run 10 to 25 percentage points; that Dawes showed classical designs did not statistically separate the rational Bayesian component from the biased component and that the residual after removing the rational part is much smaller than the original effect; that Krueger and Clement (1994) tested this by giving participants explicit base-rate information about peers' choices before asking for estimates, with some of the effect surviving and some disappearing; and that correction strategies providing base-rate information therefore work better than the original literature implied. Note: a practitioner website, not an academic source, flagged at every use. Every effect size in this article reaches us through this source alone; its claim about modern reanalyses is uncited even here, and we could not verify any of its statements against a paper. yukaichou.com
6. Repository page carrying academic citing text on the false consensus effect, recording that false consensus is a form of social projection whereby individuals overestimate the degree to which others share their characteristics or beliefs; that it is a well-documented cognitive bias expressed as the tendency to overestimate the extent to which one's opinions are also shared by others; and that the bias was initially brought to scholarly attention by Ross, Greene and House in 1977 and has since been corroborated through numerous studies across psychology, organizational behavior and planning theory. Note: a repository page reproducing text from citing papers. Our source for the plain-language definition. researchgate.net
7. Reference list carried on an academic paper about predicting others' knowledge, confirming Dawes, R. M. (1989), Statistical criteria for a truly false consensus effect, Journal of Experimental Social Psychology, 25; alongside Camerer, C. F., Loewenstein, G., and Weber, M. (1989), Journal of Political Economy, 97, 1232–1254. Note: a reference list; citation only. We did not obtain the 1989 paper, which is the source of the statistical objection this article turns on. ouci.dntb.gov.ua

This article discusses research on social perception and is not market research, product or hiring advice; nothing here bears on the lawfulness of any employment practice. The 1977 paper, the 1985 meta-analysis, the 1989 objection and the 1994 test were none of them obtained. Every effect size reported here reaches this article through a practitioner website rather than a paper, and one claim is uncited even there. All arithmetic is the authors' own and uses a uniform prior and assumptions the body states; different assumptions move every figure, and one row of the resulting table argues against the article's framing and is reported anyway.