Guidelines to give employees when they’re using AI in performance reviews (it’s a “when” not “if”). Help them use AI safely and effectively with these pointers.
April 4, 2024
Many people dread writing performance reviews. Lots of people dread writing all together. For these folks, and even the people who feel confident in their feedback and writing skills, generative AI beckons.
ChatGPT and the many other generative AI tools on the market promise easier, faster, and passably professional writing. And for a task as difficult as giving performance feedback, the idea of getting a robot to do the heavy lifting can feel like a no-brainer.
Of course, many AI options out there pretty clearly have no brains themselves.
Even so, your employees are gonna do it. They are going to see what they can get from AI as they start on performance reviews, even if they suspect it’ll be weird. And even if they clean up the weirdness, there’s bound to be bias and unproductive feedback baked into the review.
Here are some things you can proactively encourage and discourage to help managers use AI in a safe, effective way. You could even take the ideas below and convert them into your own internal checklist to present and distribute. We’re all about responsible AI here, so please feel free to steal.
Note: You’ll likely want to modify these items based on whether you already offer AI support for performance feedback as a company (and which kind, if you do). Some tools have a lot more safeguards and capabilities built in, so you may not need to outline parameters like this if you have a tool you trust that you can point people to.
Things to encourage
Here are some best practices to share with your team for using AI in performance feedback.
Encourage them to use it to synthesize data from multiple sources. It can be helpful to have an AI pull out trends and themes across different sources of performance data on an employee. This can speed up the review-writing process by helping managers quickly see commonalities across project notes and feedback from colleagues in terms of the behavior and impact of the employee. Be sure to emphasize that they shouldn’t assume the AI got it 100% right; they should evaluate the themes themselves to confirm they’re correct and modify from there.
Also! Very important to ensure people know not to include sensitive employee or company information if they’re using a publicly available tool. Everything should be anonymized, and anything proprietary should be completely left out.
Encourage them to spend time to get their prompts right. “Prompt engineering” is so important for getting back usable results. Tell your people to make sure they’re crafting appropriate prompts to generate relevant performance review content. They can instruct the AI to, for example, pretend it’s a highly qualified manager who uses the SBI framework and CARE model and wants to balance positive and constructive feedback and also provide sufficient growth opportunities. The particulars will vary depending on your company’s philosophy on performance management, but in general, help people know what to put in their prompts.
Encourage them to use it to edit. Most people already think of using AI this way, but you can help them get even more out of it with editing suggestions. They can, for example, drop in a rough draft of a review, or an outline of key points, and ask AI to shorten, expand, or form points into paragraphs. This can help people give more succinct or more robust feedback. AI can also correct spelling and grammar.
Caveat: When you’re asking an AI to write or edit feedback itself, as opposed to helping you think through what you’d write on your own, the output is going to need to be thoroughly fact-checked, evaluated for “hallucinations” and bias, and aligned to your organizational principles and/or performance management best practices. Make absolutely sure that people know they need to put AI-generated performance review text through multiple rounds of screening before they can adopt it, especially if their prompts are insufficient. What those screens look like depends on the tool you’re using—Textio Lift, for example, generates pre-screened content as it is—and your company standards.
Encourage them to look for bias and hallucinations. Hopefully, hallucinations—completely made-up “facts” and points in generative AI text—are easy to see and correct. Bias is much, much harder. You really need software for it, but if that’s not something that’s in the cards for you right now, you can, at minimum, encourage everyone to take Textio’s Equitable Performance Feedback course. It’s free! And it teaches you how to begin to see bias in feedback, and how to write more equitable reviews. Have folks check out this piece on common biases in feedback as well. Make sure everyone knows they are on the hook to avoid bias in any way they can.
Things to discourage
These are some things to make sure people are not doing when they’re using AI for performance feedback.
Discourage them from putting sensitive or confidential data in. We already covered this a little above, and it’s probably very much on your radar already, but: Lay this out clearly for folks. Unless you’re using a tool you’ve vetted and approved for internal use, employees should not be putting employee performance data or company info into an AI tool. Give them examples of what is and isn’t okay. Have them use, for example, “employee X” instead of names, and instruct them to remove all names of internal projects, financial specifics, etc. Your Legal team could be a great partner in developing these guidelines.
Discourage them from leaving it generic. Your average AI-written performance review is easy to spot, especially when it’s “from” someone you know well, like your own manager. Getting feedback in language that is clearly not your manager’s can make it feel like they checked out of the process and aren’t invested in your performance and growth. Tell people to modify what they get back from AI to fit their own voice and to include the context, examples, and ideas only they have; that will make the feedback a lot more relevant and effective.
Discourage them from using ChatGPT if you have an alternative. A lot of people have become comfortable with using ChatGPT in their day-to-day lives, and their instinct will likely be to use it for performance reviews, too. If you offer any other feedback-writing support, make sure people know it and know how to access it. It’s not that ChatGPT is all bad; it’s that it wasn’t designed for feedback, doesn’t protect against bias, and doesn’t save you from yourself when you’re writing a bad prompt and don’t know it. If you can steer people to a vetted and more purpose-built tool or resource, everyone will benefit.
What’s the best AI for performance feedback? We think it’s Textio—and you and your team can try it for free! Sign up for access here.