Who Wrote My Review? The Psychology of AI-Assisted Performance Feedback

Year-end review season is under way, and this year many reviews were drafted or polished with an AI tool, often in response to self-assessments written the same way. This guide looks at why identical words land differently depending on who people think wrote them, why suspicion does more damage than disclosure, and how managers can use AI without emptying the review of what makes it work.

How widespread is AI in performance reviews?

Widespread, and largely invisible to the people being reviewed.

Employers are encouraging it. Business Insider reported in September that Disney had given staff suggested prompts for goals and self-evaluation, calling approved tools “a brainstorming partner”. Academics quoted in the report raised the obvious problem: if every self-review sounds excellent, real excellence is harder to spot.

The appeal is obvious: writing a dozen thoughtful reviews in December is one of the most disliked tasks in management. The question is what is lost.

Why does it matter who wrote the feedback?

Because people do not judge feedback only on its content. They judge it on what it tells them about the relationship with the person who gave it.

The most direct test of this so far was published in Frontiers in Psychology in August.

It was a vignette study of anticipated reactions with a modest Hungarian sample, but the pattern fits earlier work. Research by Min Kyung Lee and by David Newman and colleagues has found that people tend to see decisions about them made by algorithms as less fair than the same decisions made by people, partly because they feel reduced to a set of data points. Newman’s team called this “algorithmic reductionism”. The worry is less that the machine is wrong than that it cannot see context.

The outcomes that moved most in the Frontiers study were help-seeking and the relationship with the supervisor, exactly what a review should build. As our guide to feedback psychology explains, feedback improves performance when it directs attention to the task and the next step, and backfires when it directs attention to the self and the relationship. A review that feels machine-made pushes attention towards the relationship: “Does my manager actually know what I do?”

Why is effort part of the message?

In most human exchanges, the cost of a gesture is part of its meaning. A handwritten note means more than a printed one because someone took the time. An annual review has always been partly this kind of costly signal. When a manager mentions the migration in March and the lost weekend, the employee learns that the work was noticed and that the manager cared enough to remember. Fluent, generic prose carries neither message, because it is so obviously cheap to produce.

This explains the split in the Highwire data. 54% of employees said AI-assisted feedback had become more specific and actionable, but about a third said it had become more generic (34%) or less useful (32%). The tool amplifies what the manager puts in: detailed notes become clear prose; a vague impression becomes polished vagueness.

Is it better to disclose AI use or keep quiet?

The Frontiers results seem to argue for silence: tell people AI was involved and they react worse. In practice, silence is the riskier option.

First, people are already suspicious. A manager who says nothing is not avoiding the question; the employee answers it for themselves, often from clues such as Priya’s matching sentences. A suspicion confirmed by accident feels like deception, a far heavier charge than a tool-assisted review.

Second, the study points to a better option. What harmed reactions was the belief that the machine did the thinking. The honest statement, “I wrote this from my notes on your year and used a tool to tidy the wording”, is also the one that preserves trust.

Third, secrecy about AI is already common in both directions. Personnel Today reported this week on Imperial College London research finding that 29% of Gen Z workers use AI without their employer’s knowledge. The review is a poor place to add to it.

What happens when both sides use AI?

Something new: a review written by one model in response to a self-assessment written by another.

Our guide to self-assessment psychology describes how self-ratings anchor managers’ judgements. AI can help quieter employees describe their work properly, but skilful prompters produce impressive self-reviews, and managers who start from those drafts inherit their framing.

The deeper risk is homogenisation. Similar tools and prompts produce the same vocabulary: “ownership”, “impact”, “stakeholder alignment”. Ratings already cluster at the top for the reasons set out in performance review bias; generic language compresses what difference remains, leaving calibration meetings with less real information.

Where are regulators drawing lines?

Under the EU AI Act, systems used to monitor and evaluate the performance of workers are classed as high-risk; following this year’s Digital Omnibus agreement, the related obligations for such systems are now expected to apply from December 2027. In California, Governor Newsom signed AB 1883 in September, which bars employers from using AI workplace surveillance tools to infer employees’ emotional states, from January 2027.

Neither rule stops a manager using a writing assistant, but both expect humans to stay accountable where software shapes judgements about people. That is also good psychology: people accept unwelcome outcomes more readily when the process feels fair, as our guide to organisational justice sets out.

What managers can do

Do the thinking before the drafting. In the Frontiers study, human substance refined by AI was received like human feedback; AI substance edited by a human was not. Note what the person did and what you want next, then use the tool.

Apply the swap test. If a sentence could sit unchanged in a colleague’s review, it tells the employee nothing. Make it specific or remove it.

Keep a log through the year. Generic reviews usually mean the manager reached December with no record of March. A few lines a month gives any tool something worth refining, and reduces recency bias.

Move feedback out of the annual document. If the review is the first time someone hears how they are doing, good prose will not rescue it. Continuous feedback builds the relationship; the annual review should confirm what has already been said. That is easier where people feel safe to ask questions, as described in psychological safety at work.

Priya’s manager, when she asked, admitted he had pasted in her objectives at eleven at night. He rewrote it, mentioning the migration and the pricing model, and said plainly how the first version had been produced. Her rating did not change. Her view of the review did.

Frequently asked questions