The Performance Review Nobody Wrote: When AI Talks to AI About Your Year

An employee opens ChatGPT and asks it to turn a year of scattered notes into a polished self-review. Their manager, staring down a stack of a dozen of these, opens their own AI tool and asks it to read the self-review, pull out the highlights, and draft a response. Somewhere in the middle, a third AI tool scores the whole exchange for tone, bias, and consistency before it goes into the employee's file.

At no point in that sequence did the employee and the manager actually sit down and talk about the year. And increasingly, that's not an edge case. It's the process.

How we got here

This is happening because both sides have real reasons to reach for the tool. In a 2025 survey of more than 1,300 managers by ResumeBuilder, 94% said they use AI to create employee development plans, 91% use it to assess performance, and 88% use it to write performance improvement plans. The time pressure behind that is real too: managers report spending three to six hours per review gathering notes and crafting feedback, and separate research found 49% of managers struggle to review a year of feedback while 42% find the process outright burdensome, according to data compiled by Windmill.

Employees have their own version of the same math. Writing a self-review that accurately sells a year of work, without sounding either falsely modest or uncomfortably self-promotional, is a genuinely hard writing task most people only do once or twice a year. Handing the first draft to AI is the same impulse that has AI writing cover letters: it's the blank page that's unbearable, not the honesty.

Some companies have simply made this official. Meta rolled out an internal "AI Performance Assistant" to help staff draft review content, and JPMorgan encouraged employees to use in-house AI tools to draft year-end review text, though the bank's guidance was explicit that the output should be a starting point, with the employee still responsible for the final version. At Tailor Brands, a business platform company, managers no longer read through customer support tickets themselves. As cofounder Tom Lahat put it, AI now analyzes every ticket every day and management review has become "basically that the AI is giving us a summary of the actions," with managers building feedback off that summary rather than the underlying work itself.

The cost of the loop

This is where the pattern your instinct is picking up on has a name: workslop. BetterUp Labs and Stanford's Social Media Lab coined the term in a 2025 Harvard Business Review article to describe AI-generated content that looks like real work but doesn't hold up once someone actually needs to use it. Their research found 41% of workers have encountered workslop at their job, and each instance costs the person who receives it nearly two hours of rework, sorting out what the document actually means and whether it's accurate.

The trust damage compounds from there. Follow-up research on the same phenomenon found that when colleagues recognized a document as workslop, trust in the person who sent it fell by 42%, perceived effort dropped by 49%, and the clarity of the interaction fell by 31%, regardless of whether the AI origin was disclosed upfront. Of people using AI at work, 18% admitted they had sent along AI-generated content they knew was unhelpful, low effort, or low quality. That's not a hypothetical about some other department. That's the exact exchange happening inside a review cycle: a self-review that reads well but is thin on real specifics, met with manager feedback that reads well but was never actually written by the manager.

It also helps explain a number that should worry any executive currently pushing AI adoption for its own sake. MIT Media Lab research found 95% of organizations that have piloted generative AI report no measurable return on that investment. Holweg and Davenport, writing in HBR in 2026, argue the mechanism is what they call "knowledge decay": when AI-generated work replaces real judgment across an organization, colleagues downstream spend their time verifying, correcting, or quietly working around it instead of building on it. A performance review built this way isn't neutral. It's actively degrading the one document that's supposed to be a record of what actually happened and what should happen next.

It's not all bad. That's the point.

None of this means AI has no place in a review cycle, and pretending otherwise would be its own kind of dishonesty. The underlying problems AI is being asked to solve are real. Reviews genuinely do suffer from inconsistency between managers, from vague language that gives an employee nothing to act on, and from a manager's memory of a year that's shaped more by the last two weeks than the first ten months. AI tools that pull together goals, project data, and peer feedback across a full year can catch things a manager's memory simply won't. And employees aren't reflexively opposed to any of this: 75% support AI-generated review content, as long as a human actually reads and stands behind it before it reaches them.

The line is the same one that matters in hiring. A tool that helps a manager organize twelve months of scattered notes into a clear draft, which the manager then edits, checks against specific memory, and puts their own name behind, is doing the job a good assistant would do. A tool that reads an AI-written self-review, generates AI-written feedback in response, and gets checked by a third AI for tone before anyone human has read either document closely, isn't assisting the review. It's replacing the only part of the review that was ever supposed to be human: one person telling another person, in their own words, what they saw this year and what they think should happen next.

Bringing the human back

The fix isn't banning AI from the review cycle. That fight is already lost, and the time-saving case for using it to organize information is legitimate. The fix is drawing the line in the same place it belongs in hiring: AI can gather, summarize, and organize. It should not be the thing doing the judging, and it should not be standing in for the conversation itself.

Practically, that means a manager can use AI to pull together a first draft from real notes and real project data, but the specific examples, the judgment calls, and the tone need to come from someone who was actually paying attention during the year. It means an employee can use AI to help structure a self-review, but the manager reading it should be able to tell the difference between a document that reflects genuine reflection and one that's technically accurate but hollow, because that difference is exactly what workslop research shows people already notice, whether or not anyone discloses it.

Most of all, it means protecting the one part of a review that AI genuinely cannot do: the actual conversation, where a manager says something true about the year that isn't in any document, and an employee gets to respond in real time. That's not a productivity loss. That's the actual review. Everything AI touches around it should be in service of making more room for that conversation, not a replacement for having it at all.