Meta has taken AI use out of the criteria it uses to rate engineers, after the metric did what metrics do. With adoption written into performance reviews, engineers began spending large volumes of tokens for no reason other than to rank high on internal leaderboards — tokenmaxxing. It was not cheap: Meta's spending on internal AI use in 2026 is approaching billions of dollars. From 2027 the company plans to issue AI budgets and put the spending on a single dashboard.
The figure is a consumption number — what Meta's own staff run up using AI inside the company — and it grew because of how those staff were being graded. A company can absorb a large bill for tools that work. This was a bill for a scoreboard.
The 2027 fix is a spending control. Budgets cap how much a team can burn, and a dashboard shows where it went. Neither addresses what made tokenmaxxing possible in the first place: nobody could tell whether the tokens bought anything. Meta could measure usage and could not measure value, so it graded usage. Deleting the criterion ends the incentive; it does not supply the missing measurement.
Separately, Meta is testing an AI tool called Hatch. According to WIRED, Hatch is meant to carry out tasks on a computer by itself. Inside the company the test has met resistance: some employees do not want to connect Hatch to their personal accounts, on privacy grounds.
These read as one story. Meta wants agents doing real work on real accounts, and the honest way to learn whether they do is to look at what comes out. Instead the company looked at what went in, got what it asked for, and is now installing meters. The engineers who inflated their token counts and the engineers declining to hand Hatch their personal logins are behaving consistently: give them a measure that pays, they will feed it; ask them to absorb a personal risk with no measure attached, they will pass. Neither group is being irrational. The grading scheme was.
What has not surfaced is what happens to the ratings already handed out while the criterion was live. Engineers were ranked, at least in part, on a measure the company has now judged bad enough to remove. Performance reviews turn into compensation and promotion. Retiring a metric is cheap; unwinding what it decided is not, and nothing that has come out suggests anyone intends to.
The timing is its own tension. The budgets arrive in 2027. Hatch is being tested now — and the tool Meta most wants adopted is the one employees are refusing to plug into their own accounts, with the single lever that would have pushed them to do it anyway just taken out of their reviews.