Meta's internal leaderboard pulled
Meta reportedly shut down an internal AI leaderboard called 'Claudeonomics' within days after employee usage data began circulating publicly, and the company is separately investigating an engineer over alleged unauthorized downloads of 30,000 private images. Those incidents underline how metrics-driven internal tooling and loose data access controls can create governance and privacy failures faster than teams expect. For large ML orgs, the risk is not model quality but the incentives and observability around who can run what experiments and who sees what telemetry. (indiatoday.in)(cybernews.com)
Meta built an internal scoreboard that ranked employees by how much artificial intelligence they used, and then pulled it within days after screenshots and usage data started escaping the company. The tool was called “Claudeonomics,” and reports on April 9, 2026 said Meta shut it down after internal numbers began circulating publicly. (theinformation.com) (indiatoday.in) The board did not measure whether a project shipped faster or whether a model answered better. It measured tokens, which are the small chunks of text an artificial intelligence system reads and writes, like counting every word-sized piece that goes into and out of a chatbot. (theinformation.com) (gizmodo.com) That made the system feel less like a productivity dashboard and more like a step counter that rewards movement whether or not you are going anywhere useful. The leaderboard reportedly showed the top 250 employees and handed out labels such as “Token Legend,” “Session Immortal,” “Model Connoisseur,” and “Cache Wizard.” (theinformation.com) (aol.com) The numbers were huge enough to turn an internal joke into a governance problem. Reporting said Meta employees used more than 60 trillion tokens over a recent 30-day period, and one top user logged about 281 billion tokens by themselves. (theinformation.com) (the-decoder.com) Once a company turns one metric into status, people start optimizing for the metric instead of the job. Reports said some employees let artificial intelligence agents run for hours mainly to inflate their token totals, which burns computing resources and makes the leaderboard look like output when it may just be activity. (the-decoder.com) (theinformation.com) The shutdown landed next to a separate Meta problem that is much less funny. A former Meta employee in London is under criminal investigation after allegedly downloading about 30,000 private Facebook images while working at the company. (cybernews.com) (news.sky.com) Court documents reported by British outlets say the engineer allegedly wrote a script to bypass Meta’s internal detection systems, which is the digital equivalent of building a side door around the building’s badge scanner. The Metropolitan Police cybercrime unit is investigating, and Meta said it discovered the issue more than a year ago, fired the employee, notified affected users, and referred the matter to police in the United Kingdom. (telegraph.co.uk) (news.sky.com) (mashable.com) Put those two stories together and the pattern is not “Meta used too much artificial intelligence.” The pattern is that internal systems can expose the wrong information to the wrong people, or reward the wrong behavior, long before a company notices the incentives it has created. (theinformation.com) (cybernews.com) That is why the real risk in large machine learning groups is often not the model itself but the plumbing around it. A leaderboard can turn token use into office prestige, and a weak internal control can turn employee access into a privacy breach affecting thousands of personal photos. (theinformation.com) (news.sky.com) Meta’s week shows how fast internal tools can move from clever to corrosive. If a company can measure every prompt and every click, it also has to decide who sees those numbers, who can act on them, and what behavior those numbers quietly pay people to chase. (theinformation.com) (indiatoday.in)
