Meta “Token Legends” AI leaderboard controversy: using usage as performance could turn it into performative work

Meta is measuring employees’ AI adoption in an unexpected way—according to a report by The Information, the company has created an internal leaderboard called “Token Legends,” where employees compete with each other using AI compute and token usage as stand-in metrics for status and productivity. Ethan Mollick, a well-known professor at the University of Pennsylvania, quoted the classic management paper “The Folly of Rewarding A, While Hoping for B” on X and issued a sharp warning: when companies use the wrong metrics to measure the effectiveness of AI adoption, AI may become the next generation of “performative work.”

Token usage contest: Meta’s AI adoption comes with new rules of the game

According to reports, Meta has set up an internal “Token Legends” leaderboard that allows employees to view each other’s AI compute consumption. This mechanism has created a competitive culture internally, and employees have begun to use token usage as proof that they’re “embracing AI.” However, this approach raises a fundamental question: does usage equal value?

In another tweet, Mollick went further and provided a surprising figure: Meta’s daily AI compute consumption reaches two trillion tokens a day. This scale is not only evidence of investment in technical infrastructure, but also clear proof that enterprise-level AI adoption has entered a stage of large-scale, institutionalized rollout.

Classic management warning: the AI version of “rewarding A while hoping for B”

Mollick draws on the classic management paper “On the Folly of Rewarding A, While Hoping for B” to analyze this phenomenon. This widely cited paper reveals a common problem in organizations: when incentive mechanisms are disconnected from actual goals, employees will optimize the metrics being measured rather than the results the organization truly needs.

Applied to Meta’s situation: the company wants employees to use AI to improve work quality and efficiency (goal B), but it measures them by token usage (reward A). The result may be that employees use AI heavily to climb the leaderboard, even if those uses do not translate into real productivity improvements. This is exactly like the performative labor in past corporate settings—where “being seen in the office” was treated as “working hard.”

2025 feels seamless, but 2027 will be completely different

Mollick also presents an important time-based perspective: in 2025, GenAI may not have a significant impact on jobs at large enterprises, because at that time there are no truly agentic tools, adoption takes time, and everyone is still in the experimenting stage. But this situation is changing rapidly.

He warns that studies showing AI had no impact in 2025 cannot tell us what 2027 will look like. As agentic AI tools mature and organizational processes are fully reshaped, companies will officially move from the “experiment phase” to the “scale deployment phase.” And how to design the right incentive mechanisms during this transition will determine who can truly gain a competitive advantage from AI.

Lessons for the industry: the real challenge of AI adoption isn’t technology

Meta’s “Token Legends” case reveals the deeper problem behind enterprise AI adoption: technical deployment itself is no longer the bottleneck—organizational behavior and incentive design are. When companies make “how much AI you used” a KPI, they are effectively rewarding a behavior that has nothing to do with output. Truly effective metrics should measure the real outcomes that AI delivers—project completion speed, code quality, customer satisfaction—rather than simply usage volume.

For Taiwan-based companies driving AI transformation, Meta’s experience offers an important warning: while rushing to roll out AI tools, it is even more important to think carefully about how to design the supporting performance evaluation system. Otherwise, AI will only become a new kind of performative work tool, not a genuine productivity transformation engine.

This article Meta “Token Legends” AI leaderboard controversy: using volume as performance could lead to performative work first appeared on Lien News ABMedia.