Guide
Performance review bias: 5 types and how to design them out
When ratings reflect the reviewer’s personality rather than actual output, promotions and bonuses go to people who are liked, not people who perform. Here are the five biases that do the most damage, and the process changes that remove them.
Where this fits: shared scorecards and calibration, steps 2 and 4 of the 5-step performance management framework, are where most bias is designed out. Next step: score your own review process with the free bias checker.
The 5 biases
Recency bias
What it looks like: Managers judge employees on their last couple of weeks of work instead of output over the whole cycle. Without real-time data, reviews default to grading recent events.
How to fix it: Pre-fill scorecards with real-time and historic data (goals, tickets, commits, feedback) and summarise a whole cycle of 1:1 notes before the review starts.
Central tendency
What it looks like: Defaulting everyone to 3/5 to play safe and avoid tough conversations. You can no longer tell top performers from the rest.
How to fix it: Use observable Yes/No behaviours and a 5-grade scale with clear distinctions, so ratings capture genuine performance variance.
Halo and horn effect
What it looks like: A single high or low area skews the entire review score.
How to fix it: Score goals, skills and values separately, with recommended grades auto-calculated from data across all three.
Leniency
What it looks like: Some managers habitually over-rate their teams, inflating grades.
How to fix it: Calibration flags such as “Manager Leniency” (a historical habit of over-rating) and “Grade not supported by deliverables” (e.g. an “Exceeding” recommendation when task completion was only “Developing”).
Apples-to-oranges standards
What it looks like: Uneven grading bars across managers for identical output: a 4/5 means meeting expectations to one manager and saving a failing project to another.
How to fix it: A shared skills library and universal anchors, so a high performer in one team is held to the same standard as one in another.
Three structural fixes

In HR Chief, Andrei describes multi-leader assessment: multiple leaders provide assessments of an employee – rather than a single individual – to ensure a diverse perspective. (Source: HR Chief interview)
1. Replace rating scales with observable Yes/No behaviours
Break skills and values into specific, observable milestones assessed on a strict Yes/No basis. An employee either demonstrates the behaviour or doesn’t, which leaves far less room for argument than a 1–5 scale. See employee performance rating scales explained.
2. Calibration flags and mandatory justification
Let data surface the outliers before the meeting: lenient managers, grade inflation, and grades not supported by deliverables. Any manual change to a system-recommended grade should need a documented rationale with evidence. See how to run a performance calibration meeting.
3. The dual-manager system (optional)
So one biased supervisor can’t bottleneck a career, the final grade can average inputs from two leaders: a line manager (focused on delivery) and a functional manager (a discipline expert focused on quality and craft). This helps neutralise individual favouritism.
In practice
Ebury moved away from disconnected grading spreadsheets that let managerial bias creep in, and cut departmental calibration time from 4 hours to 1 hour. Huspy replaced gut-feel promotions with a grading system and measurable KPIs.
Check your own process with the free performance review bias checker (you can also paste in a handful of ratings to spot central tendency or leniency), take the 2-minute performance health check, or start from the performance review template.
Seven steps to reduce manager rating bias
The three structural fixes above set the system. The seven steps below are how a People team puts them into practice over one or two cycles, with a way to tell whether each step worked.
Why manager rating bias deserves attention
The evidence that ratings reflect raters is strong. In a study published in the Journal of Applied Psychology, Scullen, Mount and Goff analysed ratings of 4,492 managers by multiple raters (doi:10.1037/0021-9010.85.6.956). As summarised by Buckingham and Goodall in Harvard Business Review, 62% of the variance in those ratings could be accounted for by raters’ individual peculiarities of perception.
Bias also shows up in the words, not just the numbers. Correll and Simard’s HBR article “Research: Vague Feedback Is Holding Women Back” describes how vague, non-specific feedback can hold some groups back more than others. And Kahneman and colleagues’ work on noise shows that people applying the same guidelines can reach very different judgements on similar cases. Each of these points to the same conclusion: a review process needs structure to be fair.
Step 1: Write down what “good” means for each level
Bias thrives in ambiguity. If “meets expectations” is not written down, each manager supplies their own definition, and the definitions differ. Publish a talent bar by seniority: the minimum standard for deliverables, skills and culture at each level. Then turn it into scorecards with observable statements (“Delivers team-level KPIs without close supervision”) rather than adjectives (“strong”, “proactive”).
How to tell it worked: two managers rating the same anonymised example reach the same grade, or can point to the exact statement they disagree on.
Step 2: Rate each area separately
A single overall rating invites the halo and horns effect, where one standout trait colours everything. Rate deliverables, skills and culture separately, each against its own criteria, and calculate the overall grade from the parts. The performance review template shows the structure.
How to tell it worked: people start receiving mixed profiles (strong deliverables, developing skills) instead of uniform scores across every area.
Step 3: Require evidence for every rating
Ask for a one-line evidence note for each scorecard area: a deliverable, a behaviour, a piece of feedback. Specific evidence counters both recency (the last few weeks crowding out the rest) and vagueness. It also makes calibration faster, because the discussion starts from facts.
The evidence has to exist before review time. Encourage managers to keep short running notes from 1:1s, or to log feedback as it happens, so the review is reading, not remembering.
How to tell it worked: calibration discussions reference specific work, and fewer ratings change because “I had forgotten about that project”.
Step 4: Add a second perspective
A single rater is the highest-risk setup. Ask one other manager or a senior peer who worked closely with the person for brief input on each review. McKinsey’s 2024 article “In the spotlight: Performance management that puts people first” reports that only two in five companies in its research use both upward and downward evaluation in individual reviews, and describes 360° input as a way to build a more comprehensive fact base.
How to tell it worked: the second input changes or confirms ratings with stated reasons, rather than being ignored.
Step 5: Brief raters with examples, before every cycle
A 20-minute briefing before the cycle opens does more than a long annual training. Walk through the biases most likely to show up (recency, halo and horns, leniency, central tendency, similarity) with one concrete example each, and remind managers of the evidence rule. The performance review bias guide has examples you can reuse.
How to tell it worked: managers start challenging their own first draft before submitting it.
Step 6: Check distributions by manager
Before calibration, look at each manager’s ratings side by side. A manager whose team is almost all at the top, or almost all in the middle, is not necessarily wrong, but needs to evidence it. Paste each manager’s ratings (numbers only) into the ratings spotter to flag possible leniency, severity or central tendency.
Use the distribution as a prompt, not a quota. McKinsey describes companies shifting from forced distribution to distribution guidance, and the playbook’s guideline weights (15–25% A-players, 65–85% above bar, 0–10% underperformers) are a starting point for questions rather than a rule. For the line between that prompt and a forced curve, see stack ranking vs calibration.
How to tell it worked: the spread of ratings across managers narrows over two or three cycles without anyone being told to hit a number.
Step 7: Calibrate with evidence on the table
Calibration is where the steps above come together. Managers present evidence for their top and bottom ratings against the bar; peers ask questions; the facilitator logs every change with a reason. Three checks matter most: generous graders, missing or thin evidence, and outliers or sudden jumps. The calibration meeting guide and the calibration meeting template show how to run it.
How to tell it worked: you can explain every final rating with evidence, and employees in different teams with similar output get similar grades.
Worked example: two managers, one standard
An illustrative product company has two engineering managers with similar teams. Manager A rates five of six engineers at the top grade; Manager B rates five of six in the middle. Before the company introduced a written bar, both distributions went straight into pay decisions, and Manager A’s team received noticeably larger rises for similar output.
After steps 1 to 3, both managers rate against the same mid-level scorecard, separately for deliverables, skills and culture, with an evidence note per area. The ratings spotter flags possible leniency for Manager A and possible central tendency for Manager B. In calibration, Manager A evidences two of the top ratings convincingly and moves three down a grade; Manager B, prompted to find their strongest performers, moves one engineer up with clear evidence of work above the bar. The final distributions still differ, but now the difference reflects the teams rather than the managers, and every change is logged with a reason both managers can explain to their reports.
What order to do this in
- High: Steps 1 and 7: a written bar and structured calibration. Without both, other fixes have little to hold on to.
- Medium: Whichever three fixes the checker shows; usually evidence (step 3), a second rater (step 4) or distribution checks (step 6).
- Low: Keep step 6 running each cycle and re-run the checker after any process change.
Common mistakes when reducing bias
- Relying on training alone. Awareness helps, but ratings change most when the process changes.
- Forcing a curve. A quota swaps one unfairness for another, especially in small teams. See stack ranking vs calibration.
- Removing ratings entirely. Decisions on pay and promotion still get made; without ratings they are made less visibly.
- Fixing bias in one cycle. Expect two or three cycles before distributions settle.
- Ignoring the words. Read a sample of written reviews for vague language, not just the scores.
To see how much time your current calibration takes, use the review cost calculator. For the full method behind these steps, read the performance management framework.
Sources: Scullen, Mount & Goff (2000), Journal of Applied Psychology; Buckingham & Goodall, HBR (2015); Correll & Simard, HBR (2016); Kahneman et al., HBR (2016); McKinsey, In the spotlight (2024).
FAQ
What is recency bias in performance reviews?
Judging an employee on their most recent weeks of work instead of their output over the whole cycle, usually because the manager is relying on memory rather than data.
What is central tendency bias?
Defaulting everyone to the middle of the scale, such as 3 out of 5, to play safe and avoid tough conversations.
What is the halo or horn effect?
When one strong or weak area skews the entire review score.
How do you reduce bias in performance reviews?
Base reviews on full-cycle data, replace subjective rating scales with observable Yes/No behaviours, use calibration that flags lenient managers and unsupported grades, require written justification for manual changes, and consider a second, functional manager’s input.
Related: stack ranking vs calibration, the 9-box grid, including a free template and when to replace it and measuring subjective roles fairly.
Prefer to talk? On the 15-minute review you walk Ash through your current review cycle. You leave with 3 specific fixes for it, whether or not you buy anything, plus an optional price for your headcount if you want one.
Customer results from Revolut People’s published customer stories (each story page links its source); capabilities as described by Revolut People. Rewritten in our own words.
Follow us on Google