Why You Shouldn't Use LLMs to Draft Employee Performance Reviews (And What to Do Instead)

The Temptation of the Clean Slate

It is late on a Thursday evening, your coffee has gone cold, and you are staring at a blank HR portal. You have seven performance reviews to write by tomorrow morning. Your team is brilliant, but condensing a year of their triumphs, near-misses, and subtle career growth into three boxes of "constructive feedback" feels like pulling teeth.

Then, a thought strikes. You have a tab open to /platforms/openai or /platforms/claude. Why not paste in a few bullet points, ask the model to "professionalise the tone," and let the LLM do the heavy lifting?

It feels harmless. It feels efficient. It is also a fast track to eroding the trust of the people who make your team tick.

While using AI to optimise code or write boilerplate is a developer’s superpower, using it to write human feedback is an ethical misstep. Let’s look at why machine-generated performance reviews fail your team, and how you can use AI as an editor rather than a proxy author.

The Linguistic Uncanny Valley

Humans are remarkably good at detecting synthetic empathy. When you use an LLM to draft a performance review, the model does what it does best: it predicts the most statistically probable sequence of corporate buzzwords. It replaces your genuine, quirky, specific appreciation with bloated phrases like "demonstrates a proactive approach to cross-functional synergy."

When an employee reads this, they do not feel valued. They feel processed.

Performance reviews are already fraught with anxiety. If an engineer suspects that their manager could not find fifteen minutes to write three sentences from the heart, the relationship suffers. Feedback is a trust currency; paying it in counterfeit bills depreciates its value instantly.

The Privacy Nightmare You Are Ignoring

Before you paste your team's achievements into a prompt, think about where that data is going.

If you are using the free, consumer-facing tiers of popular web interfaces, your inputs may be used to train future models. By pasting a summary of an employee's performance—complete with their project names, specific mistakes, and personal development goals—you are potentially feeding sensitive corporate and personal data into a public model.

Even if you are using an enterprise API with strict data-retention policies, there is a fundamental ethical issue with uploading an individual's professional assessment without their explicit consent. Your employees have a right to know who, or what, is processing their career trajectory.

For a deeper dive into how different providers handle your inputs, check out our guide on /platforms/claude/articles.

The Homogenisation of Talent

LLMs are trained to find the average. When you ask them to write feedback, they round off the sharp, interesting edges of your team's personalities.

An engineer who is brilliantly chaotic—producing stunning, creative architecture but occasionally forgetting to update their Jira tickets—will be nudged by an LLM towards a generic profile of "balanced performance." The model will spit out standard advice on organization, stripping away the nuanced appreciation of their unique strengths.

If we let LLMs write our reviews, we end up managing to the mean. We encourage a culture of compliance rather than celebrating the eccentric outliers who actually build the innovative systems we rely on.

How to Use AI Ethically as a Manager

This does not mean you must banish AI from your management workflow entirely. Writing is hard, and staring at a blank page is a genuine bottleneck. The key is to shift the AI’s role from writer to editor and sounding board.

Here is a practical, ethical framework for using LLMs in your review cycle:

1. Reverse the Workflow: Write First, Edit Second Never start with an empty prompt and a list of bullet points. Instead, write your raw, messy thoughts down first. Do not worry about spelling, grammar, or corporate tone. Write exactly what you mean:

> "Sarah killed it on the legacy database migration. She stayed up late twice to fix the replication lag. She needs to speak up more in planning meetings because she has great ideas but lets others talk over her."

Then, ask the LLM to review your draft for clarity, bias, or constructive framing. Your prompt should look like this:

> "I have written this performance feedback for an engineer. I want you to act as a neutral editor. Check if my criticism is constructive and clear. Do not add new information, do not use corporate jargon, and preserve my original meaning."

2. Check for Hidden Bias We all have unconscious biases. LLMs can be incredibly useful for spotting gendered language or unfair expectations in your writing. You can run your raw draft through a model with a prompt designed to highlight these issues:

> "Review this feedback for potential gender, age, or cultural bias. Are there words here that are statistically used more often to criticise specific groups (e.g., 'abrasive' vs 'assertive')?"

This uses the machine as a mirror for your own judgement, rather than a replacement for it.

3. Keep the Data Local If you must use an AI tool to help structure your thoughts, ensure you anonymise the input. Replace "Sarah" with "Developer A" and remove specific, identifiable project codenames. This protects your team's privacy while still allowing you to get help with structural layout.

Keeping It Human

Your team wants to hear from you. They want to know that you noticed when they saved that production deployment at 2 AM, and that you care enough about their career to sit down and write about it in your own words.

Keep the LLMs in your terminal and your build pipelines. When it comes to evaluating the human beings behind the code, leave the machines out of the loop.