Meta Platforms has told its engineers that how much they use artificial intelligence will no longer determine how they are rated. In an internal memo this week updating expectations for engineers, executives Maher Saba and Santosh Janardhan wrote that the company will not use AI adoption dashboards or token-usage figures to assess employees’ impact, according to people who received the message. The memo added a number intended to put the policy in context: 93 percent of code modifications at Meta are now assisted by AI agents. And it carried a caution that sounds obvious after a year of watching it be ignored: using AI, the executives wrote, is not itself the end goal.
The shift closes an experiment in measuring people by machine usage. About a year ago Meta told employees they would be graded on AI-driven impact, and in practice that standard translated into labels such as AI Native, AI First and AI Enabled, according to employees and reporting by Wired. Inside the company, the policy produced what some workers called tokenmaxxing, a race to consume as many AI tokens as possible so the numbers on a dashboard looked right, whether or not the underlying work improved.
The new guidance is meant to point managers back at the work. Employees are to be judged on what they produce and whether it matters, with AI treated as one way to get there rather than as the measure of the journey. The change drops the language that tied evaluations to AI usage and to the internal AI Native designation. Meta said the update simply emphasizes what was always true, that employees are assessed on their contributions.
The timing is awkward in a way Meta’s leadership acknowledges. The company spent the past year pushing adoption: it made AI tools broadly available and expected, ran company-wide AI Transformation Weeks, and told engineers that AI-assisted output was the new normal. At the same time, the cost of that enthusiasm became a management problem. Internal AI usage at Meta has grown so quickly that the company is tracking toward spending billions of dollars on it in 2026, and it has begun building systems to monitor and ration token consumption. Meta is telling engineers that tokens will not measure them even as its own dashboards watch every token they spend.
The earlier metrics also drew legal challenge. In July, roughly two dozen employees sued Meta, arguing that AI-usage labels were used to target workers on health and family leave during layoffs in May. Meta has said employees are assessed on their contributions and that the labels were not used in evaluations. The policy change does not settle that case, but it removes the most visible symbol of the disputed system.
Underneath the memo is a measurement problem the whole industry is circling. When machines write most of the code, what does an engineer contribute? Usage can be gamed, as Meta’s own year demonstrated; output volume can be inflated by agents that generate code in seconds; and outcome quality is hard to score at scale. Meta’s answer is to fall back on managerial judgment of product results, and to accept that evaluating engineers now means evaluating the judgment they apply to work an agent produced.
The cautionary lesson extends beyond Meta. Usage statistics collected under institutional pressure are close to worthless as signals of real utility. If a company as sophisticated as Meta could not stop employees from gaming token counts with the incentive made explicit, vendor-reported engagement figures across the industry deserve the same skepticism. Founders citing adoption dashboards as proof of product-market fit, and executives treating token consumption as a proxy for productivity, are measuring the same thing Meta just stopped measuring.
The 93 percent figure deserves its own reading. Meta’s own engineering leaders have said AI agents now assist with nearly all code modifications, which means the ordinary work of writing software has already been reshaped. What survives as distinctively human is deciding what to build, reviewing what a machine produced, catching the errors that look plausible, and owning the outcome when something breaks. A performance system built for that reality has to judge judgment, not keystrokes.
Meta is not alone in recalibrating. Technology companies have cycled through internal AI metrics over the past year, from mandatory usage quotas to per-employee token budgets, and several have quietly walked the policies back as employees learned to game them. The difficulty is that AI genuinely changes what workers produce, so abandoning a bad metric leaves managers without any metric at all, and the industry has not yet agreed on what the right one looks like.
For Meta’s engineers, the memo removes one incentive to churn tokens and restores a cleaner question: did the work matter? For the company, it separates two ideas that its own dashboards had fused, that AI should be used and that using it is an achievement in itself. The expectation that engineers work with AI is not going anywhere; 93 percent of code modifications is evidence of that. What Meta has decided is that the number of tokens consumed tells it less than the work those tokens produced.
