Performance Review Bias: 12-Item Checklist (2026)

Performance reviews are the single most consequential people decision a manager makes each year. They drive promotions, compensation, terminations, and the trajectory of a career. They are also, as four decades of organizational psychology research repeatedly demonstrates, riddled with cognitive bias. Harvard Business Review's longitudinal coverage of the field (Lipman, Wittenberg-Cox), the McKinsey 2025 Women in the Workplace report, Gartner HR research on calibration, and Joan Williams's Bias Interrupters work at UC Hastings all converge on the same uncomfortable conclusion: most ratings tell you more about the rater than about the ratee.

What changed in 2026 is the legal exposure. The EU AI Act high-risk obligations for HR systems apply from 2 August 2026, the EEOC has expanded its scrutiny of algorithmic and rater-driven employment decisions, and class action plaintiffs in both the US and Europe are increasingly pleading bias in performance reviews as the upstream cause of disparate-impact promotion and pay gaps. A biased review process is no longer just unfair — it is documentable, discoverable, and in several jurisdictions actionable.

This guide is the practitioner answer to that reality: the 12 most common performance review biases, defined and counter-measured; a 12-item printable checklist a manager completes before submitting any rating; the calibration meeting design that catches what the checklist misses; and the AI bias detection layer (KS-Agents Torch) that surfaces cross-rater inconsistency, demographic disparity, and language bias in qualitative comments. Read it once, run the checklist every cycle.

For the methodology under the hood, see our 360-degree feedback guide, our best free 360 feedback tools listicle, the AI-powered 360 feedback guide, and the 360 feedback questions examples library. For the manager-coaching complement, see our 1-on-1 meeting template with AI prep and our voluntary turnover analysis root causes guide.

Why Performance Review Bias Matters More in 2026

Bias in performance reviews has been a quiet quality problem since the modern review cycle was invented. In 2026 it becomes a loud legal and reputational problem for three independent reasons that compound on each other.

First, the EU AI Act high-risk classification. Regulation (EU) 2024/1689 places AI systems used for "evaluation, promotion, or termination" of workers under Annex III, point 4. From 2 August 2026 the obligations of Articles 8-15 apply: risk management, training data quality and bias mitigation (Article 10), technical documentation, automatic logging, transparency to affected workers, human oversight, accuracy and robustness. A performance review tool that generates AI summaries, scores, or recommendations carries provider obligations; the employer carries deployer obligations under Article 26 including the duty to inform workers and the human-oversight duty. Bias is no longer a soft outcome metric — it is a documentable conformity failure.

Second, EEOC enforcement. The US Equal Employment Opportunity Commission has expanded its 2024-2025 strategic enforcement plan to include algorithmic decision-making in hiring, promotion, and performance management, on top of its ongoing class-action pipeline against subjective performance evaluation systems (the Walmart-Dukes line of cases left the door open for narrower disparate-impact theories tied to specific managerial discretion practices). Settlements in the $5M-$50M range have become routine for employers whose rating distributions show statistically significant demographic disparities the employer cannot defensibly explain.

Third, workforce fairness expectations. The McKinsey 2025 Women in the Workplace report, now in its eleventh year, continues to document a "broken rung" at the first promotion to manager — for every 100 men promoted, only 81 women are promoted, with the gap wider for women of colour. The pay gap traceable to performance ratings (the "rated lower, paid less, promoted slower" pipeline) is a measurable economic injury. Workforce expectations on procedural fairness — the perception that the review process is itself just, regardless of the individual outcome — is now the strongest single predictor of voluntary turnover among high performers under 40.

The honest reading of the 2026 landscape: a manager who submits a rating without checking it against a structured bias-mitigation protocol is exposing the employer legally, the employee unfairly, and the manager themselves to the consequences when a class action discovery process examines a four-year span of their ratings.

The 12 Most Common Biases in Performance Reviews

Each bias below is defined in operational terms (how it shows up in the rating), grounded in the research that documents it, and paired with the counter-measure that demonstrably reduces it. The list draws on Lipman's HBR work, AIHR bias research, DDI rater training literature, Joan Williams's Bias Interrupters toolkit, and the Lattice State of People Strategy data on calibration impact.

1. Recency bias

Definition. The rater weights events from the last 4-8 weeks far more heavily than events from earlier in the review period. A review meant to evaluate a 12-month performance is in practice often a recall of the last quarter.

How it shows up. A high performer who had a rough Q4 sprint gets rated below their actual annual contribution. A low performer who closed a flashy deal in November gets rated above their actual annual contribution. Recency bias performance review patterns are the most consistent finding in the rater-error literature.

Counter-measure. Continuous documentation. Require managers to maintain a brief monthly performance log (3-5 bullets per direct report) so the annual review draws on twelve data points rather than three. KS-Agents Torch surfaces this log as part of review prep automatically.

2. Halo / horns effect

Definition. The halo effect is the over-generalisation from one positive attribute to all attributes (a strong communicator is also rated high on technical depth, ownership, and judgment, regardless of evidence). The horns effect is the negative mirror: one weak attribute drags ratings on unrelated competencies down.

How it shows up. The halo effect performance review pattern is statistically detectable as suspicious correlation across competency scores from a single rater (a Cronbach's alpha approaching 1.0 across competencies that should be independent). The horns effect appears as a uniform low rating once the manager has decided the employee is "not a fit".

Counter-measure. Force the manager to provide a behavioural example for each competency separately, in the rater UI, before the rating can be submitted. Anchor each competency to its own behavioural rubric.

3. Central tendency

Definition. The rater clusters all ratings around the middle of the scale (3 on a 1-5 scale) to avoid the conflict of giving a low rating or the perceived inflation of giving a high rating.

How it shows up. A manager's rating distribution shows 80% of direct reports at "Meets Expectations", with no genuine differentiation. Central tendency bias destroys the developmental signal of the review and produces compensation outcomes that demotivate top performers.

Counter-measure. Calibration meetings (see section 4 below) are the strongest counter. A second-best counter is to require behavioural-anchored rating scales (BARS) where a "3" requires a specific evidence threshold the manager must meet.

4. Leniency / strictness bias

Definition. Some raters are systematically lenient (most ratings 4-5), some are systematically strict (most ratings 2-3), with no real difference in the underlying employee performance. Idiosyncratic rater bias accounts for an estimated 50-60% of rating variance — more than actual performance.

How it shows up. An employee transferred from a strict manager to a lenient one experiences a one-point rating jump with no behavioural change. Promotion and compensation decisions become a lottery on which manager wrote the review.

Counter-measure. Calibration plus rater-level distribution analysis. Surface each rater's three-year distribution to themselves and to the calibration committee.

5. Similar-to-me bias

Definition. Raters give higher ratings to direct reports who resemble themselves on demographic, educational, communication-style, or background dimensions. Heidrick & Struggles research and DDI longitudinal data both document the effect at 0.3-0.5 standard deviations of rating uplift for in-group reports.

How it shows up. A manager from a top engineering school disproportionately rates direct reports from the same school higher. A manager who values directness rates direct, blunt communicators higher regardless of their actual outcomes.

Counter-measure. Diverse rater panels (the foundational logic of 360 feedback), explicit rater training on similar-to-me, and AI-driven cross-rater inconsistency flags.

6. Gender and race bias

Definition. Systematic rating disparities by gender and race that cannot be explained by performance differences. The most replicated findings: women receive more personality-focused feedback ("she is too aggressive" or "she is not aggressive enough"), men receive more skills-focused feedback. Black and Latino employees receive lower ratings on the same observable performance, with the gap widening at promotion thresholds.

How it shows up. The pay-gap data is direct: McKinsey 2025 reports women earn $0.83 for every $1 paid to white men in comparable roles, with Black women at $0.66 and Latina women at $0.58. A meaningful share of the gap traces upstream to performance ratings: lower ratings, lower merit increases, lower promotion rates compound into the pay-gap statistic. Gender bias performance review and racial bias performance reviews are not separate from compensation discussions — they are the upstream input to compensation discussions.

Counter-measure. Demographic disparity analysis at calibration time (rating distribution by gender, race, age, with statistical significance flags), language bias detection in qualitative comments (Joan Williams's Bias Interrupters identifies 22 patterns of gendered review language), and structured promotion pipelines that require independent justification when a candidate is rated lower than peer-group median.

7. Idiosyncratic rater effects

Definition. Beyond systematic biases, individual raters carry idiosyncratic patterns — they over-weight specific behaviours, anchor on specific events, or interpret competency definitions differently. The Scullen, Mount, and Goff (2000) study found that more than half of variance in performance ratings is rater-specific rather than ratee-specific.

How it shows up. The same employee gets meaningfully different ratings from two equally-informed managers. The idiosyncratic rater effect is invisible to any single review but visible across raters.

Counter-measure. Multi-rater (360) input where raters cancel each other out statistically; rater training on common rubric interpretation; AI cross-rater inconsistency flags at calibration.

8. Contrast bias

Definition. The rater evaluates an employee relative to the previous employee they reviewed rather than against the rubric. A solid B-grade performer reviewed immediately after a star looks weaker; the same employee reviewed after a struggling colleague looks stronger.

How it shows up. Ratings cluster differently depending on the order in which the manager wrote them. Contrast bias performance review is invisible from a single review but visible from rating-time-series analysis.

Counter-measure. Force the manager to rate against the rubric criteria explicitly, with behavioural anchors. Randomise the order of direct reports in the review UI rather than alphabetical.

9. First impression / primacy bias

Definition. The rater's initial impression of an employee in the first months of tenure (or in a role) becomes the lens through which subsequent performance is interpreted. A weak first three months becomes a permanent ceiling; a strong first three months becomes a permanent halo.

How it shows up. An employee who recovered from a difficult onboarding remains rated "average" for years despite trajectory data showing top-quartile growth. Primacy bias performance review patterns are particularly damaging at the boundary of promotion eligibility.

Counter-measure. Trajectory-based assessment alongside snapshot ratings. Require the manager to evaluate growth-rate explicitly as a separate dimension.

10. Stereotype confirmation

Definition. The rater unconsciously interprets ambiguous behaviour through the lens of demographic stereotypes (the "assertive" male employee is "abrasive" when female; the "ambitious" young employee is "pushy" when older).

How it shows up. Qualitative comments use systematically different vocabulary across demographic groups. Stereotype bias performance review language is the easiest bias to surface with NLP analysis: the Joan Williams Bias Interrupters dictionary catches the most common patterns automatically.

Counter-measure. Language bias detection on qualitative comments at submission time (Torch surfaces flagged patterns for manager review before submission). Bias interrupters training for managers.

11. Distance bias (remote vs in-office workers)

Definition. Raters give higher ratings to employees they see physically more often. The proximity effect is well-documented in pre-2020 research; post-2020 hybrid work made it the bias most actively studied.

How it shows up. Remote employees rate 0.2-0.4 points lower on identical performance compared to office-based peers, with the gap widest for visibility-dependent competencies ("strategic thinking", "leadership presence"). Distance bias remote workers patterns are now the leading cause of remote-work attrition. Some research finds remote workers rated more leniently for output (because output is the only signal), but rated more harshly for "soft" competencies (because soft signals are filtered).

Counter-measure. Output-anchored competencies where remote work actually shows full signal; calibration committees that explicitly review remote-vs-in-office rating distributions; rater training on the specific distance bias patterns.

12. Length-of-tenure halo

Definition. Long-tenured employees receive a positive rating premium that does not reflect their current performance — they "earned" it through past contribution and the rater is uncomfortable adjusting the rating downward.

How it shows up. An employee with 12 years at the company who is no longer high-performing remains rated "Meets" or "Exceeds" out of inertia and discomfort. The rating distortion most damages high-potential newer employees who deserve the promotion slot the long-tenured employee implicitly holds.

Counter-measure. Tenure-blind calibration where each rater is asked to defend any tenure-rating correlation that exceeds expected performance variance.

The 12-Item Checklist: What Every Manager Completes Before Submitting a Rating

The single most effective tool for reducing bias at the moment of decision is a structured pre-submission checklist that forces the manager to confront each bias against the specific rating they are about to enter. The checklist below — twelve yes-no items — takes 5-7 minutes per direct report and has been shown in multiple internal studies (and in Bias Interrupters field deployments) to reduce demographic rating disparities by 30-50%.

Print this. Run it before every rating submission.

1. Recency check. Have I drawn evidence from at least three different time windows across the review period (Q1, Q2-Q3, Q4), not only the last 6-8 weeks? Yes / No

2. Halo / horns check. Have I evidenced each competency rating with a behavioural example specific to that competency, rather than rating multiple competencies on a single overall impression? Yes / No

3. Central tendency check. Looking at my full team's ratings on this competency, do I have actual differentiation, or am I clustering around the middle? Am I willing to defend the differentiation in calibration? Yes / No

4. Leniency / strictness check. How does my three-year rating distribution compare to my peer managers? Have I reviewed my own distribution before submitting? Yes / No

5. Similar-to-me check. Are my highest-rated direct reports more demographically, educationally, or stylistically similar to me than my lowest-rated? If so, can I defend the rating delta with behavioural evidence? Yes / No

6. Demographic disparity check. Looking at my ratings by gender, race, age, and disability status, do I see disparities I cannot explain by performance evidence? Yes / No

7. Idiosyncratic input check. Have I incorporated input from at least three different sources (peers, skip-level, cross-functional collaborators), not relied solely on my own observation? Yes / No

8. Contrast check. Am I rating against the rubric criteria, or relative to the other team members I have just reviewed? Yes / No

9. First impression check. Has my rating evolved with the employee's trajectory, or am I anchored on an impression formed in the first 3-6 months? Yes / No

10. Language bias check. Re-read my qualitative comments — does the language I would use for a man (or in-group member) match the language I am using here? Does any comment focus on personality where it would focus on skill for another employee? Yes / No

11. Distance check. If this employee is remote or hybrid, have I weighted output evidence rather than visibility evidence? Have I avoided downgrading "presence" competencies for absence-driven reasons? Yes / No

12. Tenure check. Does my rating reflect this employee's current performance against the rubric, or am I rewarding (or penalising) tenure? Yes / No

A "no" on any item is not a failure — it is a flag to revisit that specific rating before submission. KS-Agents Torch can surface this checklist as a required pre-submission step in the review workflow, with auto-flagging for items 1, 4, 6, 10, and 11 based on data the system already has.

How Calibration Meetings Reduce Bias

The Lattice State of People Strategy 2024 found that organisations running structured calibration sessions report 23% lower rating-distribution variance across managers and 31% lower demographic disparity in promotion outcomes. Gartner HR research is consistent: calibration is the single highest-ROI bias intervention in the performance review cycle.

A calibration meeting is a structured cross-manager session, typically held after individual ratings are drafted and before they are finalised, where managers review each other's ratings against shared rubric examples, surface inconsistencies, and adjust where the evidence does not support the rating delta. The goal is not consensus on each individual but consistency on the rubric — what does "Exceeds Expectations on strategic thinking" actually look like, and is every manager applying the same standard?

The practical design: 8-12 managers per session, 3-4 hours per session for a team of 80-120 employees, an HR business partner facilitating, the rubric and behavioural anchors visible to all. The session reviews rating distributions first (statistical view), then individual borderline cases (clinical view), with explicit attention to demographic disparity flags surfaced by HR or by the AI bias detection layer.

Calibration is most effective when the AI layer surfaces the right data: rater-level distribution outliers, cross-rater inconsistency on shared peer-rated employees (in 360 contexts), demographic group disparity at the unit level, language bias flags from qualitative comments. We will publish the dedicated calibration meeting guide separately; for now the principle is that calibration without surfaced data becomes a discussion of opinions; calibration with surfaced data becomes a discussion of evidence.

AI Bias Detection in Performance Reviews: The KS-Agents Torch Approach

AI in performance reviews is a double-edged tool. Used badly, it amplifies the bias already in the historical data — a model trained on past promotion decisions reproduces past promotion bias. Used well, it surfaces patterns no human manager could catch unaided.

KS-Agents Torch is built around the second use case explicitly. The bias detection layer operates on three signals.

Cross-rater inconsistency flags. When a 360 review collects ratings from peers, manager, skip-level, and cross-functional collaborators, statistical inconsistency across raters on the same competency for the same employee is a bias signal. Torch surfaces ratings where one rater is >1.5 standard deviations from the rater-cohort mean, asks that rater to evidence the outlier, and presents the flag to the calibration committee. AI bias detection performance review at the cross-rater level catches idiosyncratic rater effects, similar-to-me bias, and personal conflicts that the manager-only view misses.

Demographic group disparity surfacing. At calibration time, Torch surfaces rating distributions by gender, race, age, disability status, and remote-vs-in-office status, with statistical significance flags where the disparity exceeds expected variance. The committee reviews flagged groups against the underlying performance evidence; HR documents the resolution. This is the single most effective tool for turning a Tier 3 EEOC-actionable rating distribution into a defensible one — not by adjusting ratings to mask disparity, but by examining whether the disparity is performance-explained or bias-explained.

Language bias detection in written feedback. Torch's NLP layer scans qualitative comments for the patterns documented in Joan Williams's Bias Interrupters work: personality-focused language for women where skill-focused would be used for men, "abrasive / aggressive / emotional" pattern flags, age-coded language ("digital native", "old-school"), tenure-coded language. The flags are surfaced to the manager at submission time as suggestions ("consider whether this comment focuses on style where the other comments focus on outcome") rather than blocks — preserving manager judgment while surfacing the bias for conscious review.

Torch is conformity-assessed for the EU AI Act 2 August 2026 high-risk obligations: technical documentation file, training data quality and bias-mitigation documentation (Article 10), human oversight design (Article 14), automatic logging (Article 12), transparency to affected workers (Article 13). The bias detection layer is itself transparent — a manager flagged for a language-bias pattern sees the specific Bias Interrupters reference, not a black-box "high-bias score".

EU AI Act Implications for High-Risk HR AI (August 2026 Enforcement)

From 2 August 2026, any AI system that influences "evaluation, promotion, or termination" decisions in employment is a high-risk system under Annex III, point 4 of Regulation (EU) 2024/1689. The performance review tool that uses AI for scoring, summarising, recommending, or predicting falls squarely inside the scope; the bias detection layer described above is itself in scope.

The seven concrete documentation deliverables every European HR buyer should require from a performance review tool with AI features:

1. Conformity assessment (Article 43). The provider's documented conformity declaration, typically internal control under Annex VI for HR systems.

2. Risk management system (Article 9). Iterative risk management process documentation across the system lifecycle, with bias as a named risk dimension.

3. Training data documentation (Article 10). Quality criteria for training, validation, and testing datasets, with explicit bias identification and mitigation work — particularly relevant for HR systems where Article 14 GDPR special categories may be implicated.

4. Technical documentation file (Article 11 + Annex IV). System purpose, design, training process, validation, monitoring — the document a competent authority would request in an investigation.

5. Automatic logging (Article 12). Review-cycle event logs, AI inference logs, override logs, sufficient for traceability.

6. Transparency to users (Article 13) and to affected workers (Article 26 deployer obligation). Affected workers must be informed when an AI system processes their performance data; Article 86 right-to-explanation applies in employment contexts.

7. Human oversight design (Article 14). The AI summary, score, or bias flag must be reviewable, override-able, and contestable by the affected worker. For performance reviews this is the single most procurement-critical detail — the difference between a defensible deployment and an Article 22 GDPR violation on automated decisions.

Penalties under Article 99 are material: up to €35M or 7% of global turnover for prohibited practices, up to €15M or 3% for high-risk system non-compliance. These stack on top of GDPR penalties for the same processing.

Manager Training Mini-Script: How to Coach Managers Out of Bias in 15 Minutes

Long bias training programmes have a poor evidence base — meta-analyses show short-lived attitude shifts and minimal behaviour change. What works is short, structured, repeated practice tied to the specific decision moment. The 15-minute manager script below is designed to be delivered before each review cycle, ideally in the review-prep 1:1 between the manager and their HR business partner.

Minute 1-3: Surface the manager's own three-year rating distribution. Show the manager their own ratings split by gender, race, tenure, remote/in-office, alongside peer-manager comparison. Ask: "what do you see, and is there anything here you want to investigate before this cycle?" Most managers respond honestly when given the data privately rather than in a calibration room.

Minute 4-7: Walk through the 12-item checklist with two real direct reports. Pick one the manager rates highly and one they rate lower. Walk through items 5 (similar-to-me), 6 (demographic disparity), 10 (language bias), and 11 (distance) for each. The exercise alone surfaces 60-70% of the bias the manager will avoid this cycle.

Minute 8-11: Practice rewriting one biased phrase. Take one qualitative comment from a previous review (anonymised) that contains personality-focused language, and rewrite it as outcome-focused. The Joan Williams Bias Interrupters cheat sheet helps. This minute alone, repeated each cycle, has the strongest measurable effect on language bias in subsequent reviews.

Minute 12-15: Commit to the calibration mindset. Tell the manager: "in calibration you will defend each rating against the rubric, not against the other ratings. Your job is to evidence, not to advocate." Set the expectation that "I'm not sure" is an acceptable position in calibration; "I have data" is the better one.

Run this script every cycle. Bias mitigation is a habit, not a training module.

Frequently Asked Questions

What are the most common biases in performance reviews?

The twelve most consistently documented biases in performance reviews are recency bias, halo / horns effect, central tendency, leniency / strictness bias, similar-to-me bias, gender and race bias, idiosyncratic rater effects, contrast bias, first impression / primacy bias, stereotype confirmation, distance bias (remote vs in-office), and length-of-tenure halo. The Scullen, Mount, and Goff (2000) finding that more than half of rating variance is rater-specific (not ratee-specific) frames why these biases dominate the signal in unstructured review processes.

How do you avoid bias in a performance review?

Three layered counter-measures, in order of effectiveness: structured pre-submission checklists (the 12-item checklist in this guide is the proven format), calibration meetings with cross-manager review of rating distributions and demographic disparities, and AI bias detection that surfaces cross-rater inconsistency, demographic group disparity, and language bias in qualitative comments. None of these alone is sufficient; all three together reduce demographic rating disparities by 30-50% in field deployments.

What is recency bias in performance reviews?

Recency bias is the rater's tendency to weight events from the last 4-8 weeks far more heavily than events from earlier in the review period. A 12-month review becomes in practice a recall of the last quarter. The counter-measure is continuous documentation: a brief monthly performance log per direct report, surfaced automatically at review prep so the rating draws on twelve data points rather than three. Recency bias is the single most prevalent rater error in the literature.

Is the halo effect a performance review bias?

Yes. The halo effect is one of the most-documented performance review biases — the over-generalisation from one positive attribute to all attributes. A strong communicator gets rated high on technical depth, ownership, and judgment regardless of evidence. The mirror is the horns effect: one weak attribute drags ratings on unrelated competencies down. Statistically, halo effect appears as suspiciously high correlation across competency scores from a single rater. The counter-measure is requiring a competency-specific behavioural example before the rating can be submitted.

Can AI reduce performance review bias?

Yes, when designed for that purpose explicitly. AI used for cross-rater inconsistency detection, demographic group disparity surfacing at calibration, and language bias detection in qualitative comments has demonstrably reduced rating disparities in deployed systems. AI used to score or rank employees directly tends to amplify historical bias rather than reduce it. The procurement-critical detail is whether the AI layer is a bias-detection tool (transparent, override-able by managers) or a bias-amplification tool (opaque, ratings-driving). KS-Agents Torch is designed for the first use case.

How does the EU AI Act regulate performance review bias?

The EU AI Act classifies AI systems used for "evaluation, promotion, or termination" of workers as high-risk under Annex III, point 4. From 2 August 2026, providers must document training data bias mitigation (Article 10), implement risk management with bias as a named dimension (Article 9), provide human oversight that allows override and contestation (Article 14), inform affected workers (Articles 13 and 26), and maintain automatic logs (Article 12). Bias is no longer just an HR concern — it is a documentable conformity obligation, with penalties up to €15M or 3% of global turnover for non-compliance.

What's a checklist to spot bias before submitting a review?

The 12-item checklist in this guide covers recency, halo/horns, central tendency, leniency/strictness, similar-to-me, demographic disparity, idiosyncratic input, contrast, first impression, language bias, distance bias, and tenure halo. Each item is a yes-no question the manager answers before submission; a "no" is a flag to revisit that specific rating. Field studies show 30-50% reduction in demographic rating disparities when the checklist is required. KS-Agents Torch can integrate the checklist as a required pre-submission step with auto-flagging for the data-driven items.

How does calibration reduce performance review bias?

Calibration meetings — structured cross-manager review of ratings before finalisation — reduce rating-distribution variance by 23% and demographic disparity in promotion outcomes by 31% per Lattice State of People Strategy 2024 data. The mechanism is consistency on the rubric (what does "Exceeds Expectations" actually look like) plus cross-rater accountability (managers must defend their distributions). Calibration is most effective when paired with AI surfacing of rater outliers, demographic disparities, and cross-rater inconsistency.

Are remote workers rated more leniently or harshly?

Both, depending on the competency. Remote workers tend to be rated more leniently on output-based competencies (because output is the only signal management sees) and more harshly on visibility-dependent competencies like "leadership presence", "strategic thinking", or "executive presence" (because the soft signals that drive these ratings are filtered by physical distance). The net effect across studies is a 0.2-0.4 point downgrade for remote workers on a 5-point scale, with the gap widest at promotion thresholds. Distance bias is now the leading documented cause of remote-worker attrition.

What evidence shows gender bias in performance reviews?

Multiple replicated findings. McKinsey 2025 Women in the Workplace documents the "broken rung" at first promotion to manager: 81 women promoted for every 100 men, with the gap wider for women of colour. Joan Williams's Bias Interrupters research at UC Hastings catalogues 22 patterns of gendered review language (personality-focused for women, skill-focused for men; "abrasive / aggressive / emotional" pattern). Stanford's Correll & Simard work on tech industry reviews finds women receive 76% more critical feedback than men on the same observable performance, with the criticism more often personality-targeted. Pay-gap data follows: women earn $0.83 for $1 paid to white men, with Black women at $0.66 and Latina women at $0.58 — a meaningful share traceable to upstream rating disparities.

Start Free: Run a Bias-Audited Review Cycle Before You Buy

The fastest way to find out whether your review process has the bias patterns described above is to run one cycle with the structured checklist, calibration design, and AI bias detection layer described in this guide.

KS-Agents 360° Starter is free forever for teams up to 10 employees, no credit card, no time limit. The Starter tier ships the 12-item checklist as a pre-submission step, the calibration committee workflow, and Torch's bias detection layer (cross-rater inconsistency flags, demographic disparity surfacing, language bias detection on qualitative comments). EU residency on Hetzner Falkenstein, EU AI Act conformity assessed for the 2 August 2026 deadline, ISO/IEC 42001-aligned AI management posture, native EN/IT/DE/FR/ES, Mitbestimmungsgesetz documentation pack on request.

→ Start free with KS-Agents 360° or see how Torch surfaces bias signals at calibration time.

For deeper reading, see our 360-degree feedback guide for the multi-rater methodology that addresses idiosyncratic rater effects, our best free 360 feedback tools listicle, the AI-powered 360 feedback guide, the 360 feedback questions examples library, our 1-on-1 meeting template for the coaching follow-up, and our voluntary turnover analysis guide for the downstream cost of unaddressed bias.