Amazon staff and AI tools: : Goodhart`s law & cobra effect in real life

As detailed in a new report by the Financial Times, Amazon employees are reportedly using the company’s new internal AI tool, MeshClaw, to create extraneous AI agents — not to increase productivity, but just to drive up AI activity.

Also, Amazon is tracking employees` consumption of AI tokens, incentivizing some of their colleagues to prioritize quantity over quality when it comes to the technology. Some Amazon employees said that rising AI expectations are changing their workplace for the worse.

Their colleagues were using the software to automate additional, unnecessary AI activity to increase their consumption of tokens. They said the move reflected pressure to adopt the technology after Amazon introduced targets for more than 80 per cent of developers to use AI each week, and earlier this year began tracking AI token consumption on internal leader boards.

Though Amazon apparently told employees that their AI usage stats wouldn’t come up in performance evaluations, not all workers believe it. “Most managers are looking at it,” employees said. “When they track usage, it creates perverse incentives, and some people are very competitive about it.”

The employees claim that the company has a target of 80% of developers using AI each week, and that employees’ token consumption is tracked on an internal leaderboard. But a representative for Amazon said that there is no such company-wide metric for AI usage, nor are there internal leaderboards where employees are measured against each other. Rather, employees are able to view their own AI usage on personal dashboards.

MeshClaw, the tool some Amazon employees are using to inflate their AI usage, takes inspiration from OpenClaw, another AI tool that’s infamous for its potential productivity—and for its potential risks. Unlike other AI models, OpenClaw and MeshClaw run locally on users’ own hardware, giving them unprecedented independence. Earlier this year, the director of alignment at Meta Superintelligence Labs went viral when OpenClaw nearly nuked her entire email inbox, proving the potential danger of giving too much access to AI.

mazon isn’t the only company reportedly asking its employees to ramp up their AI usage. At companies like OpenAI and Anthropic, individual employees are processing billions of tokens a week, while managers at Meta and Shopify are factoring workers’ token consumption into their performance reviews. At Google, even nontechnical employees are being told to use artificial intelligence in their workflows.

What interesting conclusions can we make from this case from DEXIMES point of view? Here they are:

1. 1. Reckless usage doesn`t immediately improve workflow or output “per se” (despite some managers still think so). If business simply requires employees to use AI tools at least Х times a week, employees will definitely start making some weird and pointless things just to meet the usage quota.

• • Goodhart`s Law always works. It`s known by some specialists that any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes. Under Goodhart’s law, “when a measure becomes a target, it ceases to be a good measure”. And employees start using “tell me how you measure me, and I will tell you how I will behave” scenario. In the 90s companies used to track engineer productivity by monitoring lines of code as an output. They ended up with engineers purposefully writing terrible code turning loops, methods, etc. into hundreds of unnecessary repeated lines.

• • Top executives shouldn`t forget about “cobra effect”. The results of a perverse incentive scheme are also sometimes called cobra effect, where people are incentivized to make a problem even worse. Once upon a time, the British government, concerned about the number of venomous cobras in Delhi, offered a bounty for every dead cobra. Initially, this was a successful strategy; large numbers of snakes were killed for the reward. Eventually, however, people began to breed cobras for the income. When the government became aware of this, the reward program was scrapped, and the cobra breeders set their snakes free, leading to an overall increase in the wild cobra population.

• • Many blue-chip companies have implemented strict AI usage quotas or internal leaderboards just to prove their digital transformation to investors and executives. To meet these unrealistic metrics, employees always resort to “token maxing”. If this pattern persists, firms may normalize performative AI use, reward the wrong behavior, distort compensation schemes and miss really valuable gains from AI adoption. The healthier alternative is to focus on quantifiable business outcomes, quality metrics, real error rates, and real-life customer impact instead of raw token counts.
    
    
  • Generative AI can significantly transform some business outputs. Really. Nevertheless, it`s not an all-in-one solution for every case, everywhere and all the time. It still requires thoughtful special guidance, thorough tune-up and valuable feedback from workers in order to produce useful outputs on complex or ambiguous work. Recent research from BetterUp Labs and Stanford Social Media Lab unveiled the phenomenon of “workslop” – AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task. It creates the illusion of progress – slick slides, lengthy reports, overly tightened summaries, or code without context. Rather than saving time & costs, it leaves colleagues to do the real thinking and clean-up.