AI in Performance Reviews: What Managers Can Use It For, Legally

AI in performance reviews means using a language model to draft, summarize, check or score the written evaluation of an employee. Drafting from a manager's own notes is lawful everywhere and useful. Letting the model's output decide a rating, a raise, a promotion or a dismissal is where the risk sits, and in Illinois, New York City, California and, from January 1, 2027, Colorado, it is where the law attaches notice, record and human review duties. The dividing line is whether the model's output substantially assists or replaces the decision, and the safe practice is to keep it on the drafting side of that line.

The pages that rank for this topic are written by the HR software vendors whose products do the drafting. They promise that the tool flags bias, cite no statute and no study, and quote time-saving figures that cannot be traced. This guide uses the Pew, Gallup and Resume Builder surveys with their samples and dates, the one published test of model bias in performance feedback, and the statutes, to give a manager a rule for each use of AI in a review cycle.

Five uses of AI in a review, from safe to risky

How many managers are doing this

Managers use AI more than the people they manage. Gallup's fourth-quarter 2025 survey of 22,368 US employees found 69% of leaders, 55% of managers and 40% of individual contributors using AI at work at any frequency, and 44%, 30% and 11% using it a few times a week or more. Gallup's June 2025 reading had 40% of all employees using AI at least a few times a year and 19% weekly or more.

The one survey that asked managers specifically about decisions on their reports is Resume Builder's, fielded June 24, 2025 through Pollfish with 1,342 US full-time managers who have direct reports, published June 30, 2025. Among managers who use AI in managing people: 91% used it to assess performance, 78% for raise decisions, 77% for promotions, 66% for layoffs and 64% for terminations. Asked whether AI made final decisions without human input, 5% said all the time, 16% often and 24% sometimes. On training, 32% had formal training in using AI for people decisions, 43% informal and 24% none. The sample was limited to managers aged 25 and over with household income above $75,000, an associate degree or higher and at companies with more than 11 staff, recruited online by a resume company. Read it as a signal about the direction of practice, not as a national rate.

No published figure isolates performance management as a share of HR AI use. SHRM's December 2025 survey of 1,908 HR professionals found 39% of organizations using AI in HR, with recruiting the top function at 27%. Our HR technology statistics page has the full breakdown.

Where the bias comes in

Performance reviews have a long-documented bias problem without AI: women receive more feedback about personality and less about outcomes, and reviews of some groups are shorter and vaguer. The question is whether a language model reduces or reproduces that pattern, and the only published test on performance feedback specifically is small and old. Textio's analysis published January 25, 2023 prompted the then-current ChatGPT to write feedback for people in different occupations and traits, then ran 150 pairs of prompts that differed only by pronoun. The model presumed a nurse was a woman about nine times in ten. Asked to write about a "confident" employee with no gender given, it produced neutral text 60% of the time, male 30% and female 10%. Feedback it wrote for women was about 15% longer than for men, with the extra length mostly critical. More than 80% of the pronoun-swapped pairs produced substantially the same text, so the difference was concentrated in a minority of outputs. This was a vendor's informal test of an early model, not a peer-reviewed study, and current models differ. It is the reason the consistency check in the table above is worth running on AI drafts as well as human ones.

The broader evidence on model bias comes from hiring, where a 2024 University of Washington study found three large language models preferring white-associated names over Black-associated names on identical resumes 85% of the time against 9%. A model that ranks employees from text about them is exposed to the same effect. Our AI resume screening guide covers that study and the lawsuits.

One structural risk is specific to reviews. Models reward volume and fluency. A manager who feeds the model three pages of notes about one employee and four bullet points about another will receive a longer, warmer review for the first, whatever the two people did. The model amplifies the manager's attention, which is itself the most common source of unfairness in reviews.

What employees think

Pew Research Center's survey of 11,004 US adults, fielded December 12 to 18, 2022 and published April 20, 2023, is the only large probability sample on these specific questions, and it is now four years old. Asked about AI being used to evaluate job performance, 39% opposed, 31% favored and 29% were not sure. Asked about AI being used to decide who gets fired, 55% opposed and 14% favored. For promotions, 47% opposed and 22% favored. For final hiring decisions, 71% opposed. The pattern is the same as the risk table: people are split on AI in evaluation and against AI in the decisions that follow from it. A Gartner survey of about 3,500 employees in October 2024, reported in trade press in March 2025, found 87% thinking algorithms could give fairer feedback than their managers; we could not read the question wording and the two findings may not conflict, since fairer feedback and a final decision are different things.

The law by decision type

Title VII, the Americans with Disabilities Act and the Age Discrimination in Employment Act apply to an AI-influenced rating or dismissal exactly as to a human one, and disparate impact liability does not require intent. Beyond that, the duties depend on where the employee is and what the output decides.

Two federal documents that pages still cite are gone. The Equal Employment Opportunity Commission's technical assistance on AI and the ADA (May 2022) was removed from its website in January 2025. The National Labor Relations Board General Counsel's memorandum 23-02 of October 2022, which targeted algorithmic management and surveillance, was rescinded by the Acting General Counsel on February 14, 2025. The underlying laws are unchanged: the ADA still requires accommodation in evaluation, and the National Labor Relations Act still protects employees who discuss their reviews with each other.

The practical reading of the table: drafting and checking text is outside every rule. A model that proposes a rating which the manager adopts is inside Illinois and California now and Colorado from 2027, and inside New York City if the rating feeds a promotion. The full comparison is on our AI hiring laws by state page.

What not to feed the model

• Health, leave and accommodation information. The ADA limits disability-related inquiries and the Family and Medical Leave Act makes retaliation for leave unlawful. Text about an employee's absence for surgery should not be in a prompt about their performance.
• Protected characteristics and proxies. Age, pregnancy, religion, national origin, and the zip codes Illinois names as a proxy.
• Monitoring data as evidence. Activity scores, keystroke counts and screenshot summaries describe presence, not output. A model given them will treat them as performance. Our productivity tracking policy guide explains why monitoring data should not feed ratings.
• Other employees' reviews as context. Comparing one employee against named colleagues in a prompt discloses the colleagues' records to a vendor.
• Anything in a consumer account. Employee names and performance text are restricted data under any sensible AI use policy, and a free account may train on them.

Rules for a review cycle that uses AI

1. The manager writes the evidence; the model writes the prose. Notes, examples, outcomes and goals go in. If the manager cannot supply them, the problem is the manager's attention, and the model will hide it.
2. Equal input, equal effort. Set a minimum of specific examples per employee before any drafting starts, so the model's output length and warmth track the record rather than the manager's familiarity.
3. Run the consistency check across the whole team. Ask the model to list, for each draft, the number of specific outcomes cited, the number of personality descriptors, and the length. Re-read the outliers.
4. No ratings from the model. If the tool proposes a rating or ranking, turn the feature off until legal has mapped the Illinois, California, New York City and Colorado duties and the notice has been issued.
5. Use the company account only. Enterprise plans that do not train on inputs, under the AI use policy, with employee data classified as restricted.
6. Tell employees. A sentence in the review process guide that managers may use an approved AI tool to draft text from their own notes, and that every rating and decision is the manager's, costs nothing and removes the surprise.
7. Keep the notes. The manager's source notes are the primary evidence Colorado's human review standard and any Title VII defense will turn on. Keep them for four years with the review.
8. Train the managers first. Resume Builder found 24% of AI-using managers had no training and 43% only informal training. Our AI training for employees guide has a manager module.

Key takeaways

• Drafting a review from the manager's own notes and checking drafts for consistency are low-risk uses. Letting a model propose ratings or shape raises, promotions or dismissals is where notice, audit and explanation duties attach.
• 69% of leaders, 55% of managers and 40% of individual contributors used AI at work in late 2025 (Gallup, 22,368 employees).
• Among AI-using managers in a June 2025 survey of 1,342, 91% used AI to assess performance, 78% for raises, 77% for promotions, 66% for layoffs and 64% for terminations; 21% said AI made final decisions often or always. The sample was online and skewed.
• Textio's January 2023 test found early ChatGPT presumed nurses were women nine times in ten and wrote feedback about 15% longer for women, mostly critical. It was an informal test of an early model.
• Pew (December 2022, 11,004 adults): 39% oppose and 31% favor AI evaluating performance; 55% oppose AI deciding firings; 47% oppose AI deciding promotions.
• Illinois (since January 1, 2026) and California (since October 1, 2025) regulate AI in promotion, discipline and discharge now; Colorado from January 1, 2027; New York City when a tool substantially assists a promotion decision.
• The EEOC's 2022 ADA guidance and the NLRB's 2022 algorithmic management memo were both withdrawn in early 2025. The ADA and the National Labor Relations Act are unchanged.
• Keep health, leave, protected characteristics, monitoring scores and colleagues' reviews out of prompts; use a company account; keep the manager's notes for four years.