Amazon employees are inflating AI usage to top leaderboards and impress managers

[Alfonso Maruccia]

Editor's take: Like many major technology companies attempting to turn "dumb" AI models into profitable tools, Amazon is reportedly engaging in questionable accounting practices. According to internal sources, the company is now actively encouraging employees to inflate usage metrics to sustain momentum in its AI business efforts.

Three Amazon employees told The Financial Times that the company's internal AI usage metrics are likely inflated. According to the sources, employees are using MeshClaw, Amazon's internal AI platform, to perform non-essential tasks in an effort to make a stronger impression on managers.

The OpenClaw-inspired MeshClaw tool gives "Amazonians" a way to run local AI agents on their own hardware. Amazon is currently deploying the technology more broadly, enabling users to automate a larger portion of their daily work.

"Some employees are participating in an internal "tokenmaxxing" competition to consume as many AI tokens as possible"

– Doug Colkitt (@0xdoug) May 12, 2026

Some employees are indeed doing just that, the sources said. They are participating in an internal "tokenmaxxing" competition to consume as many AI tokens as possible, climb company-wide leaderboards, and impress their managers.

Amazon said that employees' AI token usage will not be used to evaluate individual performance. However, the sources claimed that managers are still informally taking tokenmaxxing activity into account. Some managers are reportedly tracking AI usage unofficially, while employees continue expanding MeshClaw-based automation. One source said these "perverse incentives" are fueling increasingly competitive behavior among staff.

One source said, "there is just so much pressure to use these tools. Some people are using MeshClaw simply to maximize their token usage."

MeshClaw can automate repetitive tasks and interact with third-party applications, including Slack and email clients. Amazon has officially stated that thousands of employees are now using the tool on a daily basis. However, the company is pushing for AI deployment that is safe, secure, and responsible – at least when it comes to its own customers.

Amazon had initially shared team-wide statistics on MeshClaw usage but has since restricted access to internal staff only. The company's official policy instructs managers not to consider AI usage when evaluating employee performance, though FT sources suggest a different reality.

Beyond fueling a "perverse" race to climb internal tokenmaxxing leaderboards, MeshClaw's growing adoption is also raising security concerns. Some employees are reportedly uneasy about using generative AI to automate broad portions of their work, particularly due to concerns about hallucinations or factual errors.

One FT source said, "the default security posture terrifies me," and they are "not about to let it go off and just do its own thing."

Permalink to story:

Amazon employees are inflating AI usage to top leaderboards and impress managers

 

[pnntmp]

This article is 100% based on hearsay and concoctions. There isn't even one verifiable claim, everything is 'reportedly' coming from mysterious anonymous sources that may or may not exist.
Why would any company encourage inflating internal metrics? That's something nobody outside the company sees, internal metrics are a very valuable feedback. It's insane the management to deliberately falsify the information based on which they make their decisions.

 

[Dreadcthulhu]

When a measure becomes a target, it ceases to be a good measure.

 

[The Talking Tech]

This article is 100% based on hearsay and concoctions. There isn't even one verifiable claim, everything is 'reportedly' coming from mysterious anonymous sources that may or may not exist.
Why would any company encourage inflating internal metrics? That's something nobody outside the company sees, internal metrics are a very valuable feedback. It's insane the management to deliberately falsify the information based on which they make their decisions.

With literally every company in the world including Amazon encouraging and exploiting their employees into training A.I models...and you really think this is a bridge too far?

Is it really that hard to believe this?

Naiveness have no limits.....

 

[pnntmp]

With literally every company in the world including Amazon encouraging and exploiting their employees into training A.I models...and you really think this is a bridge too far?

Is it really that hard to believe this?

Naiveness have no limits.....

AI companies are not 'encouraging' their employees to train AI models - they are demanding it, that's the employees main job. And believe me, none of them feels exploited - they work a lot, but are compensated even better. The 'exploited' most definitely don't need you to militate on their behalf.

 

[NumberNine]

Well at least everyone involved here is being equally dishonest with each other, which in the corporate world is basically as good as integrity.

 

[The Talking Tech]

AI companies are not 'encouraging' their employees to train AI models - they are demanding it, that's the employees main job. And believe me, none of them feels exploited - they work a lot, but are compensated even better. The 'exploited' most definitely don't need you to militate on their behalf.

Just because they have to choice between keeping themselves employed training their replacement or lose their jobs, it doesn't mean they are not feeling exploited.

 

[pnntmp]

Just because they have to choice between keeping themselves employed training their replacement or lose their jobs, it doesn't mean they are not feeling exploited.

Actually nobody has to choose between these things. Training a model is not training your 'replacement', it's something entirely different.

 

[Mr Majestyk]

A lose-lose for employees. CEO says use AI or your fired. Employee uses AI, CEO says to board, it's apparent AI is essential to our company as so may employees are using it. Board tells CEO to cut out the middleman and sack the employees

 

[ZedRM]

Some have said that AI needs very strong regulations worldwide. Others argued against that idea.

Where we at now people?

 

[DrSuess]

Yeah, I was so busy I let AI order my lunch. It ordered me bean curd with tofu, that was the moment I realized AI sucked.

 

[abactuon]

But how do you check something like this?

88-year-old Donald Knuth, the legendary computer scientist and author of The Art of Computer Programming, published a paper on Stanford's website titled "Claude's Cycles." It begins with the words "SHOCK! SHOCK!": it turns out that Anthropic's Claude Opus 4.6 model has solved an open graph theory problem Knuth had been working on for several weeks.
The process wasn't smooth: Stappers had to restart sessions due to errors, and the model periodically forgot to document its progress. The problem for even m remains open—here, Claude proposed hypotheses but was unable to generalize the results. "It looks like I'll have to rethink my views on generative AI," Knuth wrote. "What a joy—not only to learn that my hypothesis has a beautiful solution, but also to celebrate this impressive breakthrough in automated proof and creative problem solving."

Donald Knuth is too old a man, and I don’t trust his young colleague (who allegedly communicated with Claude) implicitly.

 

[llamapanda]

Plot twist: this article written almost entirely by AI, consuming an obscene amount of tokens in order to prove effectiveness of the tech to upper management