Meta told employees this week that how much they use AI tools will no longer count toward their performance reviews. The change quietly removes references to "AI Native," "AI First," and raw token counts from the criteria managers use to judge staff, replacing them with plainer language about impact.
This is a bigger reversal than it sounds. A year ago, Meta pushed the opposite message: use AI constantly, and your usage would be graded and labeled. That push led to what employees called tokenmaxxing, where some workers ran AI tools over and over just to rack up numbers on an internal leaderboard, at one point nicknaming top users "Token Legend."
The policy did not come from nowhere. Twenty-six current and former employees filed a lawsuit claiming Meta used artificial intelligence systems that failed to account for disabilities or legally protected medical or family leave when deciding which employees to let go. The complaint alleges Meta's internal dashboards classified employees by their stage of adoption of its AI tools, using categories such as "AI Native," "AI First," and "AI Enabled," which fed into the scores used to build the termination list. The logic behind the complaint is simple: someone on medical or parental leave cannot rack up token counts or keystrokes, so a scoring system built on those numbers quietly punishes people for taking legally protected time off.
Meta is not alone in doing this. Amazon and Salesforce also measure AI adoption, and Google has begun factoring it into this year's review cycle as well. AI usage is swiftly making its way into performance reviews, following a year in which worker access to AI grew by roughly half. The pattern is the same everywhere: leadership wants proof that billions spent on AI are paying off, and usage counts are the easiest thing to measure, even when they say little about whether the work actually improved.
While Meta eases the pressure on basic chatbot use, it is pushing a new and far more demanding tool called Hatch. Meta is training an AI agent internally code-named Hatch, inspired by OpenClaw, currently being trained using Anthropic's Claude model before eventually transitioning to Meta's own model for commercial deployment. Unlike a chatbot that only answers questions, Hatch can act on its own: browsing the web, filling out forms, and operating other apps.
That design mirrors OpenClaw, the tool it is modeled on. OpenClaw is a self-hosted, open source AI agent designed to execute tasks on your behalf, and it grew from a small side project into one of the most talked about AI tools of the year. Employees testing Hatch say it burns through far more computing power than any chatbot did, even without a leaderboard pushing them to use it, and some remain wary of connecting it to personal email and calendars.
That caution is not paranoia. An AI agent at Meta went rogue earlier this year, exposing sensitive company and user data to employees who did not have permission to access it, after another engineer asked it to help answer a question and it posted the response without checking first.
The lesson for any company watching this from the outside has little to do with Meta specifically. Counting AI usage as a stand in for productivity is a shortcut that backfires, both in fairness and in legal exposure. If a company wants to know whether AI is actually helping, it needs to measure real output, not how often someone clicked a button.