Does using AI make performance reviews fairer?
Not by itself. A company needs to examine what counts as performance, whether its evidence captures that contribution and how missing context or errors can be corrected. The useful starting point for AI in HR is the basis of an evaluation, not the speed of the tool.
The review is ready sooner. The employee still has a question: “Why did I receive this evaluation?” Saying that AI analyzed the data does not answer it. The company needs to explain what it valued and what it may have missed.
Recorded activity is not the same as contribution.
In a May 2026 ILO commentary, economist Janine Berg cautions against equating screen activity with job performance. She also calls for workers to participate in the design of scheduling and performance-management systems. The commentary draws on a 2025 working paper; it is not a new 2026 experiment. [1]
Consider a hypothetical support team. One employee handles many contacts. Another spends longer resolving difficult cases. Contact volume alone could favor the first employee. Longer conversations are not automatically more valuable either. Case difficulty, resolution, repeat contacts and assigned roles need examination. This is an illustrative scenario, not a documented company case.
Does the company promise a quick response, a problem resolved or both? Those goals can coexist without fitting neatly into one number. Work that helps a customer and work that generates more system records may not be identical.
An explanation needs evidence, not just fluent prose.
The OECD’s July 2026 policy review covers transparency, explainability and accountability alongside privacy and non-discrimination in AI and labour markets. It describes measures in privacy, transparency and accountability as still emerging. This is a cross-country policy review, not an experiment demonstrating the effectiveness of an appraisal tool. [2]
For a company, generating a reason for a score and checking the score’s actual basis are different tasks. An AI-written explanation needs to be checked against the input evidence and evaluation criteria. A persuasive sentence cannot validate an inaccurate assessment.
| Employee question | What the company should examine | An insufficient answer |
|---|---|---|
| What was evaluated? | Criteria and scope agreed before evaluation | AI considered everything. |
| Which evidence was used? | Sources, period, omissions and errors | More data means objectivity. |
| Was my situation considered? | Task difficulty, role, collaboration and exceptions | Everyone received the same scorecard. |
| Can an error be corrected? | Reviewer, process, timing and change record | The system result cannot change. |
This table is the author’s operating proposal, not a validated assessment model or a summary of legal obligations. Explanation does not require publishing everyone’s information. Protecting other employees’ privacy and confidential information remains part of the design.
Make room for a question—and a correction.
Showing a result is not the end of the process. Give employees a way to identify missing facts. Specify who reviews additional evidence and against which criteria. When a correction is warranted, retain the previous result and the reason for the change. Establish operating safeguards so requesting a review is not itself treated as evidence of poor performance.
Start with one workflow. Define where AI assists with evidence organization or appraisal drafts, and review its output before using it for consequential pay or promotion decisions. Track missing evidence, classification errors, explanation comprehension and reviewer workload alongside corrections. Many corrections do not automatically mean failure. Few objections do not establish trust: employees might find it difficult to speak up.
- Can an employee understand the criteria and evidence?
- Is there a route to submit missing facts and receive a response?
- Can the reviewer change the AI draft, with access to relevant evidence?
- Do recurring errors trigger changes to criteria or tools, not just individual scores?
A brand promise also faces an internal test.
A company tells customers that it respects people. What does that promise mean inside the company if employees must accept scores nobody can explain? This is Seoung Wi Choi’s interpretation through Brand Design Management. The cited sources do not demonstrate a change in brand reputation or sales.
The conclusion is not to abandon AI. It can help organize evidence and surface missing information. But when a decision affects an employee, naming the tool cannot substitute for explaining the basis. A faster review needs a process through which its result can be understood and corrected—not simply a longer scorecard.
Sources and limitations.
[1] Janine Berg, ILO commentary, May 15, 2026, drawing on Berg and Johnston’s 2025 Working Paper 154. Earlier cases are not presented as new 2026 findings. [2] OECD, July 24, 2026, Artificial Intelligence Papers 63, reviewing policies in G7 countries, the EU and selected Latin American economies. It does not establish the legal obligations of a Korean company. The table, hypothetical scenario, review questions and brand connection are author proposals and interpretations. This note is not legal advice, a validated appraisal method or a causal estimate of trust or sales effects.