Performance Review Bias: 6 Types and How to Fix Them

Two people on the same team, same year, similar output.

One of them shipped steadily from January, solved a hard problem in April that nobody outside the team noticed, and spent most of the autumn quietly keeping a difficult account from leaving. The other had a slower year and then presented a well received piece of work at all hands in early November.

Guess which one gets the stronger rating.

Performance review bias is the systematic distortion of an employee's assessment by something other than their actual performance, and it happens most often because the manager is working from partial memory instead of a record. Nobody in that story did anything wrong. The manager wasn't careless and didn't play favorites. They wrote an honest review from the information available to them in November, and the information available to them in November was mostly about November. That's what performance review bias looks like in practice, and it's why the usual remedy fails.

• The six biases that show up most
• Why training, rubrics and calibration only go so far
• What reduces it, and what doesn't
• A checklist you can use this cycle without buying anything

The 6 Types of Bias in Performance Appraisal

Recency bias.
Weighting the last few weeks far more heavily than the full period. The most common and the most damaging, because it systematically rewards visibility timing rather than contribution. Anyone who becomes prominent shortly before review season benefits; anyone whose best work happened in March pays for it.

The halo effect.
One strong attribute coloring everything else. Somebody presents well, so their judgment, reliability and technical depth all get rated generously. It's particularly common where a manager has limited direct visibility and fills the gaps with a general impression.

The horns effect.
The same mechanism in reverse. One visible failure, one difficult meeting, one missed deadline, and every other dimension gets pulled down with it. Horns tend to outlast halo, because negative impressions are stickier.

Leniency and severity bias.
Whole populations rated high or low depending on who assessed them. Two managers with identical teams will produce different distributions, and the difference reflects their comfort with conflict more than their people's performance.

Central tendency bias.
Everyone lands in the middle. Safe, defensible, and it destroys the entire signal. A rating scale where 80% of the population sits on the same point isn't measuring anything.

Similarity bias.
Rating people more generously when they resemble you in background, working style or communication. This is the one with the clearest legal and fairness exposure, and the hardest to see in yourself.

Most organizations recognize all six types of performance review bias. Most also assume the answer is to tell managers about them.

Why Bias Training Doesn't Fix It
Awareness helps at the margins. It doesn't restore forgotten information, and that's the problem. Take recency bias. A manager who knows about it and is determined to correct it still can't recall what happened in March. They can try, and what they'll produce is a reconstruction shaped by whatever fragments survived. Knowing your memory is unreliable doesn't make it reliable.

The same limit applies to the standard remedies. Rating rubrics standardise the output of biased inputs, which makes the result more consistent without making it more accurate. Calibration sessions have real value, and they surface disagreement between managers who are all working from equally thin recall, so what emerges is a negotiated position rather than a better evidenced one. More frequent reviews help, and they multiply the documentation burden that caused the thinness in the first place.

None of these is useless. They're all downstream of the real issue, which is that the manager opened a blank form in November and was asked to account for a year. The outcome shows up in how employees experience the result. Gallup's research on performance reviews finds only 29% of employees  strongly agree the reviews they receive are fair, and just 26% strongly  agree they are accurate. 

What Bias Costs You Actual
Being specific matters here, because "reviews are a bit unfair" doesn't move a budget conversation. Steady contributors leave. The person who delivers consistently and quietly gets rated below the person who became visible in the final quarter, notices the pattern over two or three cycles, and concludes the organization rewards something other than the work. That's regretted attrition caused by a measurement failure. Promotion decisions drift toward visibility. Over several years this reshapes who holds senior roles, and not in the direction anyone intended.

And the exposure is increasing. Where ratings feed pay, an employer may be asked to show the objective criteria behind a progression decision. A review assembled from memory is hard to justify to the person it describes, and harder still to anyone reviewing the decision afterwards.

What Reduces It, and What Doesn't
The interventions that work share one property. They change what the manager has in front of them rather than what the manager has been told, which is why awareness campaigns underperform and process changes don't.

Shorten the recall window.
Anything that captures a judgment closer to the event beats a better attempt to remember it later. Notes after a project closes, feedback given in the week it applies, a check in that records a decision rather than a status.

Use multiple sources where the manager's view is partial.
For matrixed roles or anyone whose influence runs sideways, a single assessor is working from a fraction of the picture. More perspectives dilute individual bias, provided the raters are chosen properly and the feedback is developmental rather than evaluative.

Calibrate, but calibrate against evidence.
The session is only as good as what managers bring to it. Nine people comparing impressions produces a negotiation. Nine people comparing specific examples produces a correction.

Separate the pay conversation in time.
Candour improves measurably when nothing is being decided in the room, and candour is what surfaces the examples that counteract a general impression.

Make the assessment about what happened, not about the person.
A rubric asking what somebody did and what changed as a result is harder to fill in from a vague impression than one asking how collaborative they are.

The Structural Fix
Every one of those interventions is an attempt to get more evidence in front of the manager before they write. Bias in performance reviews isn't primarily a judgment problem. It's an input problem. A manager facing a blank form and twelve months of work to characterize will produce a biased review no matter how well trained they are, because the only material available is memory, and memory is selective in exactly the ways described above.

Change what's on the screen when the cycle opens and the six biases lose most of their grip. A manager who can see the whole period, the projects and the shipped work, the feedback from April as readily as the feedback from October, and how somebody's skills shifted, all in one place rather than across four systems, isn't reconstructing anything. Recency bias has far less to work with when March is as visible as November. The halo and horns effects weaken when there are specific examples across multiple dimensions rather than one strong impression. Similarity bias is harder to sustain against a record of what people did.

Historically that record required somebody to maintain it, which is why it never existed. What's changed is that much of it can now gather itself, drawn from the systems where the work is already being done, so the picture exists without anyone having assembled it. Feedback given in April sits in the same record as the goal it related to and the skill it demonstrated, which is what makes March as visible as November when the cycle opens. That doesn't remove judgment from the review, and it shouldn't. Weighing what an achievement cost, and what it took to get there, is human work. It means that judgment gets applied to a full year of detail instead of to a handful of recent impressions.

A Checklist for This Cycle
None of this requires a purchase. Six things you can do before the next cycle opens.

Write examples before ratings.
Ask managers for three specific examples per person before they open the rating scale, so the rating follows the evidence rather than the other way round.

Require evidence in calibration.
An example for every rating at the top and bottom of the scale, produced in the room.

Check the distribution by the manager.
Flag anyone whose spread differs sharply from their peers, then ask why.

Test whether visibility is winning.
Look at whether your highest rated people are the most visible rather than the most productive.

Ask employees for their own three.
Particularly the work nobody saw, which is the material most likely to be missing.

Diary a note when projects close.
Fifteen minutes, immediately, while it's fresh. The cheapest intervention here and the one that shortens the recall window most. If you want to see what a review looks like when March is as visible as November, bring one manager and one past cycle to a PossibleWorks demo.

Then ask the question that sits underneath all six biases: at the moment a review cycle opens here, how much of the year is visible to the person writing it? If the honest answer is the last six weeks, you don't have a bias problem that training will solve. You have a record problem wearing a bias costume.