A Third Of It Made Things Worse

You give someone feedback. Their performance gets worse. The natural conclusion is that they took it badly, or were not coachable, or needed to hear it more clearly. There is a large body of evidence suggesting a different conclusion.

Our Grade For This Claim

Applying the scheme from the first article in this series.

The core empirical finding is Grade A. It is a meta-analysis of 607 effect sizes and 23,663 observations published in Psychological Bulletin[1], one of the discipline's most selective review venues, and it is reproduced identically across the publisher's record, the lead author's institutional repository and independent reference works[2][3][4].

Feedback Intervention Theory itself is Grade B. The authors described it as preliminary in their own title, tested it with moderator analyses, and stated that the findings are further moderated by task characteristics that are still poorly understood[1]. It is a theory proposed to explain an anomaly, not an independently established mechanism.

The praise-and-criticism finding is Grade C, described by a reference work as arising from partial tests of the theory[4], and we did not obtain the underlying studies.

The practical translation: the one-third figure is solid, the explanation for it is plausible and provisional, and this article keeps those apart.

A Note On Method

Everything here is verified to August 2026.

We did not obtain the full text of the meta-analysis. We rely on its published abstract, which we obtained from four independent sources that reproduce it identically[1][2][3][5], and on a reference work describing the theory.

Our sources describe the negative proportion in three slightly different ways: the abstract says over one third[1], a reference work says almost one third[4], and a secondary commentary says over 38 percent[3]. We treat the abstract's formulation as authoritative and report the others.

Where we compute anything from the published figures we say so and show the working, and we mark clearly where a calculation is illustrative of scale rather than a prediction.

This article reviews organisational research. It is not human resources, employment law or psychological advice. Performance management carries legal and contractual obligations that vary by jurisdiction and that this article does not address.

The Finding

The result, in the authors' own words.

Kluger and DeNisi published The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory in Psychological Bulletin, 119(2), 254–284, in March 1996[1].

The abstract states: a meta-analysis of 607 effect sizes and 23,663 observations suggests that feedback interventions improved performance on average (d = .41) but that over one third of the feedback interventions decreased performance. And critically, that this finding cannot be explained by sampling error, feedback sign, or existing theories[1].

Three features are worth marking before going further, and these observations are ours.

The scale is large by the standards of this field. Twenty-three thousand observations is not a handful of laboratory studies.

The average effect is genuinely positive. This is not a finding that feedback does not work. On average it works, and d = .41 is a respectable effect.

And the authors ruled out the obvious explanations themselves, in the abstract, which is unusual and worth noticing. They did not leave the reader to wonder whether it was chance or whether the negative cases were simply the harsh ones.

Largely Ignored

The opening sentence of the paper, which is a remarkable thing to find in an abstract.

It reads: since the beginning of the century, feedback interventions produced negative but largely ignored effects on performance[1].

Three observations, ours.

The claim is not that this was discovered in 1996. It is that the negative results had been appearing in the literature for roughly ninety years and had not been assembled into a picture.

Which makes this a different kind of finding from the ones this series has covered so far. It is not a replication failure, a contested meta-analysis or a corrected statistic. It is a pattern that was visible all along in published work and that nobody had aggregated.

And the reason is worth speculating about, carefully. A researcher whose feedback intervention reduced performance would reasonably have treated it as a failed study, a poor manipulation or a sample problem. There was no theory predicting that outcome, so there was nothing for it to be evidence of.

Our own observation: this is close to the opposite of publication bias as usually described. The results were published. What was missing was a framework that made them mean anything.

What The Average Hides

The arithmetic implication, and it is ours.

Hold the two published figures together. The mean effect is +0.41. Over one third of effects are below zero.

Those two facts are not principally about the average. They are about the spread.

A distribution centred on a solid positive effect with a third of its mass on the wrong side of zero is wide. It is the signature of a practice that does substantially different things in different circumstances, not of a practice that reliably does one thing.

Three consequences.

Reporting d = .41 alone, as a great deal of management writing does, describes the literature accurately and conveys almost nothing useful. The average is the least informative statistic in this dataset.

The operative question is not does feedback work but which population is this instance in.

And that is exactly the structure the third article in this series described in the choice overload literature: a mean that conceals two populations, where the useful research is about the conditions rather than the average. The difference is that here, the authors went looking for the conditions in the same paper.

It Is Not About Positive Or Negative

The exclusion that most changes what a manager should do.

The abstract states the negative effects cannot be explained by feedback sign[1].

Feedback sign means whether the feedback was positive or negative. So the finding is not that criticism hurts performance and praise helps it.

Three consequences, ours.

The intuitive model, in which harm comes from harshness and safety comes from encouragement, is ruled out by the data. Whatever is producing the negative third, it is not simply that some feedback was unkind.

Which means the standard organisational responses, being softening the delivery, training managers in tone, and leading with the positive, are aimed at a variable that the meta-analysis specifically excluded as the explanation.

And it opens an uncomfortable possibility that the next two sections pursue: that positive feedback can also be in the harmful third.

Which Means Praise Is Not Automatically Safe

An implication that follows from the exclusion above. This section is our own reasoning, marked as such.

If feedback sign does not explain the harm, and the theory locates harm in attention moving toward the self, then praise which directs attention to the self should carry the same risk as criticism which does.

Consider the difference between two things a manager might say.

"The reconciliation you did on the Henderson file caught a duplicate posting that would have gone out in the year end." That is about a task and an outcome.

"You're one of the sharpest people we've got." That is about a person.

Both are positive. On the theory described below, only the first is reliably safe.

Two consequences.

An organisation that has responded to concerns about feedback by increasing recognition and praise may have increased the volume of self-directed feedback rather than reducing it.

And the well-intentioned instruction to "catch people doing things right" is only as good as what the praise is attached to. Attached to a specific action, it is task-level. Attached to the person, it is not.

We flag firmly that this is our inference from the theory, not a finding we obtained. The meta-analysis excluded sign as the explanation; it does not follow automatically that self-directed praise is harmful, and we did not locate a study testing that specific proposition.

The Theory They Proposed

Feedback Intervention Theory, as the authors described it.

The abstract states the central assumption: feedback interventions change the locus of attention among three general and hierarchically organised levels of control: task learning, task motivation, and meta-tasks including self-related processes[1].

The three levels, restated plainly and this restatement is ours.

Task learning. Attention on the details of how the work is done. What the correct procedure is, where the error was, what the method should have been.

Task motivation. Attention on effort and goals for this work. How much to invest, whether to keep going, what target to aim at.

Meta-task and self-related processes. Attention on the person. Whether I am good at this, what this says about me, how I compare, what others think.

Two observations, ours.

The levels are hierarchical, and the hierarchy runs from the work upward to the self.

And attention is finite and rivalrous. This is the mechanism's engine: attention spent at one level is not available at another, so feedback does not merely add information, it moves attention from somewhere to somewhere.

The Hierarchy That Predicts Harm

The result of the moderator analyses, and the single most useful sentence in this literature.

The abstract reports: FI effectiveness decreases as attention moves up the hierarchy closer to the self and away from the task[1].

Three consequences, ours.

This gives a usable diagnostic before you speak. The question is not whether your feedback is fair, accurate, kind or well-timed. It is where it will put the other person's attention.

It is a continuum, not a switch. Feedback is not either task-level or self-level. It sits somewhere on a gradient, and the prediction is monotonic: the further up, the worse.

And it explains why the effect could not be predicted from feedback sign. Both praise and criticism can be delivered at any level of the hierarchy, so sign and level are independent, and the theory says level is what matters.

We repeat the authors' own qualification, which they placed in the abstract: these findings are further moderated by task characteristics that are still poorly understood[1]. Thirty years on, a reader should check whether that has been resolved. We did not.

Why The Self Level Is The Worst

The mechanism, as a reference work describes it.

Feedback that directs attention to the self is the most problematic, as it diverts cognitive resources away from task performance and produces affective reactions that can interfere with task performance[4].

Two distinct harms in that sentence, and separating them is ours.

A resource cost. Thinking about yourself consumes attention that the work needed. This operates even if the feedback was welcome.

An affective cost. Emotional reactions arise which further interfere. This is the one people expect, and on this account it is only half the problem.

Two consequences.

The resource cost explains why positive self-directed feedback could also impair performance. Being told you are talented is pleasant and still moves attention to the self.

And it explains why delivery skill has limits. A manager can control tone, timing and framing. They cannot make self-focused attention free.

The Technique Everyone Is Taught

A specific finding about the most widely taught feedback method in management.

A reference work summarising tests of the theory reports: partial tests of FIT found that feedback interventions containing both praise and criticism produce lower positive and some negative effects on performance when they direct attention to the self[4].

Feedback containing both praise and criticism describes the standard construction taught in most management training: an opening positive, the criticism, a closing positive.

Three observations, ours, and the qualification in the third matters.

The reported finding is conditional. The clause when they direct attention to the self is doing essential work, and the finding is not that combining praise and criticism is harmful in itself.

But there is a reason to think the combination tends toward the self, and this is our own reasoning. Praise and criticism about the same person in the same conversation invites comparison of the two, and comparison of positive and negative statements about oneself is a self-level operation. The structure itself points attention upward.

And we did not obtain the underlying studies. The finding reaches us through a reference work describing partial tests, which is the authors' own characterisation of their evidentiary status.

Reading That Finding Carefully

A caution against the overclaim this article could easily make. This section is our own.

It would be easy, and wrong, to write the headline the feedback sandwich is scientifically disproven.

Three reasons we are not writing it.

The finding as reported is about interventions containing both praise and criticism that direct attention to the self, which is a narrower category than the sandwich technique and does not map onto it exactly.

It comes from partial tests of a theory the authors themselves labelled preliminary, reported to us through a reference work rather than from the studies.

And lower positive and some negative effects is a carefully hedged phrase that we should not convert into a stronger claim than it makes.

What we will say is narrower and defensible. The most widely taught feedback technique in management has not been shown to be safe, and the one piece of relevant evidence we located points the wrong way. That is a reason to prefer a method with a clearer rationale, which the sections below set out, rather than a reason to announce a debunking.

The Failure At The Other End

A complication that stops this from being a simple rule.

The same reference work notes that feedback directing attention to the details of the task can have positive effects, but there is also a potential problem of feedback recipients focusing too much on details[4].

Two observations, ours.

So task-level is not automatically safe either. The hierarchy predicts that lower is better; it does not promise that the bottom is costless.

The plausible mechanism, and this is our inference, is that attention driven entirely into procedural detail can crowd out the judgment needed to do the work well, which is a familiar phenomenon in skilled performance.

We report this because it is the honest state of the evidence and because a rule that says always be more specific would be a simplification the source does not support.

What This Means At Ordinary Volumes

Arithmetic, computed by us and explicitly illustrative.

Before the numbers, the caveat that governs them. The one-third figure is the share of interventions in the research literature that had negative effects. It is not a validated base rate for any particular manager, team or workplace. The arithmetic below shows the scale of the question, not a prediction about your firm.

On that basis, taking weekly one-to-ones over a 48-week year:

One manager, one person, weekly: 48 feedback events, of which roughly 16 would fall in the harmful third.

A team of five: about 240 events a year, roughly 80 in the harmful third.

A team of twelve: about 576 events, roughly 192.

Two observations, ours.

Even discounted heavily for the fact that this is a literature-derived proportion rather than a workplace one, the volume is the point. Feedback is not an occasional intervention. It is continuous, and anything with a substantial failure rate applied continuously accumulates.

And unlike almost every other intervention this series has covered, you are already doing it. There is no decision to adopt feedback. The only decision available is what kind.

Where The Damage Gets Attributed

The structural reason this persists. This section is our own analysis.

A manager observes the aggregate. Across a team and a year, performance improves, because on average feedback improves performance.

What the manager does not observe is the counterfactual for any individual case.

Three consequences.

When performance falls after feedback, the available explanations are all about the employee: they were defensive, they lacked resilience, they were not coachable, they disengaged.

The explanation that the feedback caused it requires a comparison that does not exist, since you cannot observe the same person in the same week without the conversation.

So the practice is self-validating at the aggregate level and unfalsifiable at the individual level, which is close to the ideal condition for a harmful practice to persist.

This is the fourth time this series has encountered that shape. Unreported errors, deferred purchases, unobserved counterfactual settlements, and now this. The behavioural failures that survive are the ones that leave no artefact and have a ready alternative explanation.

The Annual Review Problem

Applying the hierarchy to the most formalised feedback event in most organisations. This section is our own reasoning.

Score a conventional annual performance review against the three levels.

It is about the person over a period, not about a task. The unit of analysis is the employee.

It frequently carries a rating or ranking, which is a comparative statement about the person relative to others.

It is usually linked to pay, which attaches a financial consequence to a self-relevant judgment.

And it is distant in time from the work, so the specific tasks it concerns are no longer live and cannot be acted on.

On the theory as described, that is close to a maximally self-directed feedback intervention.

Two things we will and will not say.

We will say that the design has the features the theory identifies as harmful, and that anyone running one should know that.

We will not say the evidence shows annual reviews reduce performance. We did not locate a study testing that, and the theory's authors called it preliminary. Reviews also serve documentation, legal and compensation purposes that have nothing to do with performance improvement, and those may independently justify them.

What Task-Level Feedback Sounds Like

The practical translation, constructed by us from the theory rather than taken from a source.

Four properties that keep attention low on the hierarchy.

It names an artefact, not a person. The file, the quote, the call, the reconciliation. Something that exists outside the individual.

It is close in time to the work, so the task is still live and the information can be used.

It describes what happened and what the alternative was, which is task-learning content.

It contains no comparative or dispositional statement. No ranking, no reference to other people, no claim about what kind of worker the person is.

Two tests you can apply to a sentence before saying it, and these are ours.

Could this be said about the work if a different person had done it? If yes, it is probably task-level. If the sentence only makes sense about this individual, it is not.

Does it invite the answer "what does that say about me?" If a reasonable person would leave the conversation thinking about themselves rather than about the work, the theory predicts trouble regardless of how positive the content was.

The Owner Problem

Why this is harder in a small firm, in our assessment.

Three structural features work against task-level feedback in an owner-managed business.

The relationship is personal. In a firm of eight, the owner knows the employee's circumstances, and feedback naturally takes a personal register because everything else does.

There is nowhere to put the comparative statement. In a large organisation, ranking happens in a calibration process the employee does not witness. In a small one, comparison happens in the room.

And the owner's approval is the whole reward system. Where there is no formal progression, promotion track or salary band, the owner's view of you is your standing, which raises the self-relevance of everything they say.

Two consequences.

The same words carry more self-directed weight in a small firm than they would in a large one, which if the theory holds means the hierarchy problem is worse where the relationship is closest.

And the most valuable single change is probably temporal rather than verbal: moving feedback closer to the work, where the task is still live, does more to keep attention low on the hierarchy than any amount of rephrasing at a scheduled review.

Not An Argument For Saying Nothing

The misreading to head off, stated plainly.

Three things follow from the evidence, and a fourth does not.

Feedback improves performance on average, at d = .41[1]. That is a real benefit and this article is not disputing it.

A substantial minority of interventions make performance worse, and the proportion is large enough to matter at ordinary volumes.

And the best available account of why locates the problem in where attention goes, which is something a manager can influence.

What does not follow is that managers should say less. Withholding feedback has its own costs, including the ones the second article in this series described: a team that hears nothing learns nothing, and silence is frequently read as a judgment anyway.

The defensible conclusion is about composition rather than volume. The same amount of feedback, aimed lower on the hierarchy, is the change the theory recommends.

What To Do

Stop treating tone as the variable. The meta-analysis specifically excluded feedback sign as the explanation for the negative third, which means softening delivery addresses something the data ruled out.

Ask where the attention will go, before you speak. The theory's prediction is monotonic: the closer to the self, the worse the expected effect.

Name an artefact, not a person. The file, the quote, the call. If the sentence only makes sense about this individual, it is high on the hierarchy.

Move feedback closer to the work in time. In a small firm this is probably the highest-value single change, because it keeps the task live and the attention on it.

Watch self-directed praise as well as criticism. On the theory, being told you are talented moves attention to the self just as being told you are careless does. This is our inference and not a finding we obtained.

Be cautious with praise and criticism in the same conversation. The one piece of evidence we located on that combination points the wrong way, though it is a partial test reported through a reference work.

Do not expect specificity alone to solve it. The same source notes recipients can focus too much on detail, so the bottom of the hierarchy is not costless.

Recognise your annual review for what it is. Person-focused, comparative, pay-linked and distant from the work is the profile the theory identifies as most risky, whatever other purposes the review legitimately serves.

Stop attributing every post-feedback decline to the employee. That explanation is always available and the alternative is unobservable, which is precisely why it needs deliberate scepticism.

The Limits Of This Analysis

Several caveats matter. This article reviews organisational research and is not human resources, employment law or psychological advice; performance management carries obligations that vary by jurisdiction and are not addressed here. Everything is verified to August 2026. We did not obtain the full text of the meta-analysis and rely on its published abstract, obtained from four independent sources reproducing it identically, together with a reference work describing the theory. Our sources describe the negative proportion three different ways, being over one third, almost one third, and over 38 percent; we treat the abstract as authoritative and report the others. The authors described their theory as preliminary in their own title and stated that the findings are further moderated by task characteristics that are still poorly understood; we did not establish whether that has since been resolved, and a reader should check the thirty years of subsequent literature we did not review. The praise-and-criticism finding comes from a reference work describing partial tests, and we did not obtain the underlying studies. We did not locate any study testing whether annual performance reviews reduce performance and make no such claim. The inference that self-directed praise carries the same risk as self-directed criticism is our own reasoning from the theory and is not a finding we obtained. Our arithmetic on feedback volumes is illustrative of scale only; the one-third figure is a property of the research literature, not a validated base rate for any workplace, and we do not predict that a third of any manager's conversations are harmful. The two sentence tests, the four properties of task-level feedback, the small-firm analysis and the attribution-trap observation are our own construction from the theory rather than findings from the literature.

Related Reading

No. 8 Debate The Idea, Not The Person A meta-analysis found task conflict negatively related to performance, against what textbooks predicted. A larger one found no relationship either way.... Behavioural Finance Correcting The Corrections The other place where the number everyone quotes concealed more than it revealed. Behavioural Finance Thirty-Six Percent The grading scheme, and why a solid average can conceal two populations.

Frequently Asked Questions

Does feedback improve performance or not?

On average it does, at d = .41 across 607 effect sizes. But over a third of the interventions in that meta-analysis decreased performance. Both facts are from the same abstract, and the second is the one that carries the practical information.

Is the problem that people take criticism badly?

The authors ruled that out explicitly. The abstract states the negative effects cannot be explained by feedback sign, meaning it is not simply that negative feedback hurts and positive feedback helps. That excludes the intuitive model most feedback training is built on.

So what determines whether feedback helps?

On the theory the authors proposed, where it sends the recipient's attention. Effectiveness decreases as attention moves up a hierarchy from task learning through task motivation toward the self. Self-directed attention both consumes cognitive resources and produces emotional reactions that interfere with the work.

Is the feedback sandwich bad?

We are not making that claim. What we found is that partial tests of the theory reported interventions containing both praise and criticism producing lower positive and some negative effects when they direct attention to the self. That is narrower and more hedged than a debunking, and it reaches us through a reference work rather than the studies.

Should I just give less feedback?

No. The average effect is positive, and withholding feedback has its own costs, including that a team which hears nothing learns nothing. The defensible conclusion is about composition rather than volume: the same amount, aimed lower on the hierarchy.

What is the single highest-value change?

Probably moving feedback closer in time to the work, especially in a small firm. It keeps the task live and the attention on it, which does more to hold attention low on the hierarchy than rephrasing the same points at a scheduled review months later.

References

1. Kluger, A. N., & DeNisi, A. (1996). The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory. Psychological Bulletin, 119(2), 254–284. DOI 10.1037/0033-2909.119.2.254, on feedback interventions having produced negative but largely ignored effects on performance since the beginning of the century; on a meta-analysis of 607 effect sizes and 23,663 observations suggesting that feedback interventions improved performance on average at d = .41 but that over one third of them decreased performance; on this finding not being explicable by sampling error, feedback sign, or existing theories; on the central assumption of the proposed theory being that feedback interventions change the locus of attention among three general and hierarchically organised levels of control, being task learning, task motivation, and meta-tasks including self-related processes; on effectiveness decreasing as attention moves up the hierarchy closer to the self and away from the task; and on these findings being further moderated by task characteristics that are still poorly understood. Note: we did not obtain the full text and rely on the published abstract as reproduced by the publisher's record. ovid.com
2. Hebrew University of Jerusalem research repository, institutional record for Kluger and DeNisi (1996), reproducing the same abstract and confirming the journal, volume, issue, date and page range. Note: the lead author's own institution; used to corroborate the abstract text independently of the publisher. cris.huji.ac.il
3. Academia.edu hosted copy and commentary on Kluger and DeNisi (1996), reproducing the abstract and adding that research shows feedback interventions can either improve, have no effect, or debilitate performance, with over 38 percent of effects being negative, and that task-focused interventions enhance performance while meta-task directed interventions can reduce effectiveness. Note: a document-hosting platform with added commentary, not peer-reviewed; its 38 percent figure is more specific than the abstract's "over one third" and we do not adopt it. academia.edu
4. SAGE Encyclopedia of Educational Research, Measurement, and Evaluation, entry on Feedback Intervention Theory, on the theory having been first proposed by Kluger and DeNisi in 1996 to explain why feedback is not always effective; on it growing out of a meta-analysis of over 600 effect sizes; on the results indicating that in almost one third of cases feedback had a negative effect on subsequent performance; on feedback directing attention to the details of the task having positive effects but with a potential problem of recipients focusing too much on details; on feedback directing attention to the self being the most problematic, as it diverts cognitive resources away from task performance and produces affective reactions that interfere with it; and on partial tests of the theory finding that feedback interventions containing both praise and criticism produce lower positive and some negative effects on performance when they direct attention to the self. Note: a reference work summarising the theory; we did not obtain the partial tests it describes. methods.sagepub.com
5. ProQuest bibliographic record for Kluger and DeNisi (1996), Psychological Bulletin, 119(2), 254–284, reproducing the abstract and confirming the DOI. Note: a fourth independent reproduction of the abstract, used to confirm that the figures quoted throughout this article are the published ones. proquest.com

This article reviews organisational research and is not human resources, employment law or psychological advice. The meta-analysis was not obtained in full; its abstract was obtained from four independent sources reproducing it identically. Sources describe the negative proportion three different ways and the discrepancy is reported rather than resolved. The theory was described as preliminary by its own authors and its moderators were described by them as poorly understood; approximately thirty years of subsequent literature was not reviewed. No claim is made that annual performance reviews reduce performance, and no study testing that was located. Arithmetic on feedback volumes is illustrative of scale and is not a base rate for any workplace.