Dos and don’ts of writing performance reviews with generative AI

How excited are you that AI can write decent stuff now? It can write emails, meeting notes, borderline-passable blogs (not this one!)—and yes, performance reviews. User beware, though: There’s a lot to take into account before you fire up your chosen AI writing buddy and generate a performance evaluation.

Below, we outline the dos and don’ts of writing performance reviews with generative AI.

Heads up talent developer folks: This would be a good piece to distribute to managers in your org. Or you could use it as inspiration to create your own internal dos and don’ts list.

Here’s what to know.

Good news: You can stop dreading writing performance reviews

Does anyone enjoy writing these things? OK I confess: I‘m everyone’s biggest cheerleader and also I’m a writer, so yes in the past I’ve enjoyed writing reviews. But I still dreaded them! They’re so important and take so much brain space and time. I know I’m not alone in this: As Axios recently reported, “The sensitive work of writing your own self-assessment, or reviewing the work of an employee who reports to you, has become so daunting or monotonous that some would rather turn it over to AI.”

But even beyond the new generative AI tools and performance feedback software taking on some of the writing work for us, there are lots of resources out there today to make the whole process less gnarly. Have you dug around for guidance lately? There’s the SBI framework, the CARE model, Textio’s Equitable Performance Feedback course, and I’ll bet your HR team provides ample content and maybe even training to support you in this. Take advantage!

The point is: It’s a known difficult task and there’s lots of help out there. With some boundaries based on what your company allows, you can build your ideal process and system and knock these things out.

Can AI write my performance review?

An important thing for us to align on here is that AI can indeed help you write a performance review—but it’d be highly irresponsible (and honestly not even effective) for you to completely check out of the work. A quality performance review requires specific context and examples that only you would know how to phrase for your employee. It’s necessary to have a “human in the loop.” At minimum, you should consider yourself a heavy-handed editor in this task.

Is it OK to use ChatGPT for performance reviews?

My friend, it is not. You may think that because I’m a writer and also a Textio employee that of course I don’t like ChatGPT. And you’d be right. But also: Data. Data show that ChatGPT cannot be trusted to write performance reviews.

Plus, think about what ChatGPT and other general-purpose large language models (LLMs) are doing: They are simply predicting likely words. They take your prompt and then spit out an “answer” that’s a bunch of statistically likely phrases and sentences. It’s math, applied to writing. It’s not intelligent; it’s not even strategic. It’s just a remix of words from the internet (and you know what the internet is like). You can and must do so much better for your team.

The Gen AI options for performance reviews

Let’s do a very brief primer on your gen AI options for writing performance reviews. As we get into the dos and don’ts, you’ll see many things depend on what type of gen AI you’re using. Here are the general buckets:

• General-purpose LLMs: This is ChatGPT, Google Gemini, Microsoft Copilot, and the like. These are intended for a wide variety of use cases, and their source material (training data) is broad swaths of the internet.
• Add-on AI features: These are functionalities built into existing tools. For our purposes, these are features built into existing HR technologies—you know how all of your software suddenly has wowcoolamazing AI functionality? That’s what we’re talking about here. Often these are just piping in a general-purpose LLM without doing much to augment the source data for more targeted and trustworthy outcomes. They are typically generic AI features that have been rushed to market without purposeful development; some are literally being built as you, the customer, are asking for it. 
• Purpose-built AI tools: These are AI tools built specifically for a certain task, in this case for writing performance reviews and feedback. Textio Lift was the first and others are now coming to market. Purpose-built is what you want, if that’s not already clear. There are tons of extra safeguards and shortcuts baked into these tools that will make the job of writing performance reviews far faster, safer, and more effective.

Dos and don’ts of writing performance reviews with generative AI

Let’s do the dos, then we’ll do the don’ts.

• DO: Find out your company’s AI policy before you do anything. Maybe you can only use approved tools, maybe you can only put certain information into generative AI, maybe gen AI is completely off the table in your org. Find out.
• DO: Use it to synthesize data from multiple sources. Pop in all the disparate info and feedback you’ve gathered on your employee’s performance and ask the AI to pull out trends and summarize the main themes. This can be helpful for conceptualizing what you want the review itself to ultimately contain. Important: Make sure you read carefully and actually agree with the highlights the AI pulled out. It could have hallucinated, so you definitely need to check it.
• DO: Spend the time to get your prompts right. Tell the AI to pretend it’s your ideal version of a manager. Tell it what knowledge and skills it has (seriously). Tell it to base its recommendations only on very specific input you provide. Tell it what performance review examples, frameworks, and models you’d like it to align with. (It’s probably not worth telling it to avoid bias—it likely won’t, and you’ll get weird results.) Note: If you’re using a purpose-built tool, you probably won’t have to “engineer” your prompts like this because it will all be programmed in.
• DO: Use it to edit. Draft up a rough iteration of the performance review or input an outline of the points and phrases you want to include, and ask AI to help. Ask it to shorten or expand certain parts. Ask it for different ways to phrase things that are stumping you. Ask it to correct your spelling and grammar. Depending on the tool you’re using, you could also ask it for feedback on how clear, actionable, relevant, and unbiased it is—and get suggestions for improving it accordingly. Surprise, surprise: You can only 100% trust these kind of recommendations from a purpose-built tool.
• DON’T: Put sensitive or confidential information in. Unless you’re using a company-approved tool that has been vetted for data privacy and security, and you have confirmation that you can include employee and organizational details, you need to keep it general and anonymous. When in doubt, leave it out.
• DON’T: Fail to fact check. Can’t emphasize this enough. The tricky thing with LLMs is that they sound smart and correct. It’s easy to find yourself nodding along and then later be like, wait what? It can beautifully and confidently write garbage. Check every single word.
• DON’T: Leave it generic. Hate to tell you but you probably already know: People can spot generative AI writing from outer space now. It’s going to feel very weird to your employee—who knows you, and knows your voice—to get a generic performance review “from” you. Worse, it’s going to undermine trust and give them the impression you’re not fully invested in their success. You need to edit. You should also be aligning the review to your org’s leadership principles and company values, which is tough to get an AI to do in a meaningful way.
• DON’T: Leave in bias. Whether they’re your words or the robot’s, they are quite likely to contain bias. It’s nearly impossible to write a bias-free review without performance review software showing you where you’re being unfair, harmful, or perpetuating stereotypes. But even if you don’t have access to such software, you can still take a pass through the review to look for any comments that are based on personality traits (instead of strictly work), any hedging phrases that make your review unclear, any feedback the person can’t actually act on, and any unnecessarily gendered terms. (Another plug for taking Textio’s free equitable feedback course to build the muscle of spotting these things. You can also check out these equity-focused performance review templates.)

Using generative AI to write performance reviews can make your feedback more effective and save you a lot of time and stress. It’s knowing how to use it, and which tools to use, that gets you those promised AI results that people shout about on LinkedIn. And now you know!

Textio Lift detects biases in performance feedback as you write, and helps you write a more fair and effective review in real-time. Learn more about our free trial for Textio Lift.