Meta has quietly taken AI usage out of the way it rates its own employees. New performance-review rules drop references to criteria such as "AI usage" and to AI-native status, replacing them with a looser formulation: the required results can be achieved with AI or by other means. Engineers across the company were told Meta will not use AI adoption dashboards or token counts to assess contribution, The Information reported. The rollback comes close to a year after Meta announced it would evaluate staff on their "impact achieved with AI" — a policy that produced a measurable internal culture of burning tokens for the sake of burning them.
In practice, the original policy meant tracking how actively employees used chatbots and agents in everyday work, Business Insider reported. Usage was tied to labels applied to staff: AI-native, AI-first, and AI-supported.
Those labels appear in a lawsuit filed by about two dozen employees, who allege Meta violated US anti-discrimination law during a round of job cuts. Employees who were on medical or family leave say they could not accumulate the required volume of AI usage and were disadvantaged as a result. Meta denies the allegations and the case is continuing.
For several months, employees say, they ran AI tools over and over — sometimes with no practical need — to push internal usage metrics up. One of the things counted was the number of tokens consumed, the notional unit of AI work. The inflation peaked when a Meta employee built an internal leaderboard showing how much AI colleagues were using and handing out titles such as "token legend". After news of the dashboard leaked, the employee deleted it. A couple of months later, Meta imposed limits on how much AI employees could consume.
Meta spokesperson Tracy Clayton told WIRED the updates merely underscore the existing principle that the company evaluates employees' contribution, and that labels such as AI-native were never used in performance reviews. That is a narrow denial. It can be true that no manager wrote "AI-native" on a review form and still true that the company built adoption dashboards, counted tokens, and eventually had to cap usage because staff were gaming a number that supposedly did not count. Several employees told WIRED the changes are small but were received well, because people no longer have to use AI where it makes no sense.
Meanwhile Meta has started encouraging — but not requiring — employees to use its most advanced experiment, an AI agent called Hatch. It can carry out actions on a computer by itself, including browsing the web and operating other applications, putting it in the same category as the viral OpenClaw. Employees have been testing it on corporate devices for several weeks, and the tool is expected to be released publicly later. Meta declined to comment on the Hatch testing.
Some workers are in no hurry to connect Hatch to personal email, calendars and other non-work accounts. Two current Meta employees familiar with the situation told WIRED the reasons are privacy concerns and the risk of accidentally breaking something important. That caution has history behind it: Meta previously ran a project that tracked keystrokes and other activity on work devices, with the collected data intended to train AI systems. The company later paused it, but sources say it noticeably damaged employee trust in Meta's AI tools.
Removing the penalty is the smartest thing Meta has done in this sequence, and it is also self-interested. A company that grades people on tool usage learns nothing from the resulting adoption numbers, because everyone has a reason to inflate them; a company that stops grading on usage finally gets to see which tools people reach for when nothing is riding on it. Hatch is the test that matters — an agent people are asked to point at their own calendars and inboxes — and it only produces a real signal if refusing carries no cost. Some employees are already using Hatch to book personal appointments and organise things outside work. They find it effective, but worry about the environmental and economic consequences of the token and compute load that automating such tasks involves.
That worry is the part the announcement is quiet about. Even with the metric gone, employees say Hatch consumes more tokens than ordinary chatbots and the earlier coding assistants did. Meta has removed the incentive to inflate consumption at exactly the moment it is promoting the tool that consumes the most, so the internal spend curve keeps climbing either way — only now without a leaderboard to blame it on. And employees still believe management wants to see whether they can use AI tools effectively, which is the same expectation as before with the scoreboard taken down.
The fear underneath all of this is headcount. If Hatch genuinely raises productivity, some employees expect another round of cuts. Reuters reported that Meta had prepared a further large wave of layoffs after eliminating 8,000 jobs earlier, then abandoned the plan amid a rise in software bugs linked to over-reliance on AI and slower-than-expected progress on AI agents. Chief executive Mark Zuckerberg has said no new mass layoffs are expected this year.
So the company that shelved a layoff round because AI dependence was producing bugs is now handing employees an agent that acts on its own, and telling them their usage of it will not be counted. Both statements are on the record. Only one of them will survive the next quarter in which Hatch works.