Ai promised coders superpowers. inside amazon, it’s eating their day.
Amazon’s engineers are spending more time debugging machines than shipping features, according to leaked accounts that shred the gospel of friction-free ai development.
The whistleblowers, quoted in a The Guardian investigation, describe a daily grind where auto-generated snippets arrive swollen with hallucinations, stale dependencies and security land-mines. One senior developer at the company’s Seattle headquarters put it bluntly: “We clock more hours correcting garbage than we ever spent writing the thing ourselves.”
Metrics, not mastery, drive adoption
Amazon tracks how many lines an ai assistant spits out and how quickly engineers merge them. Performance reviews, the engineers say, reward volume, not viability. The result: coders accept broken pull requests knowing they will have to triage the fallout later, because pushing back is flagged as “low ai enthusiasm” in internal dashboards.
The numbers mock the premise. A backend team handling Alexa’s voice-commerce module saw defect tickets rise 42 % after mandatory ai rollout, while story-point completion flat-lined. “We essentially hired an intern who never sleeps but lies constantly,” another engineer noted.
Amazon’s spokesperson declined to comment on specific metrics, citing “ongoing optimisation efforts.” Yet redundancy rounds that trimmed 27 000 roles last year coincided with a 40 % budget bump for machine-learning infrastructure, feeding suspicion that the long game is fewer humans, more heuristics.
The hallucination tax
Modern large language models excel at remixing public code, but Amazon’s proprietary libraries are opaque to the internet corpus the models ingested. Ask for an API that routes grocery orders and the ai hallucinates parameters that never existed. Engineers must then archaeology-dive through decade-old repos to disprove the fantasy.
Compound that with legal risk: if the model regurgitates GPL-licensed code, the entire module can be tainted. One lawyer inside Amazon estimates 8 % of AI-generated commits now trigger a mandatory licence-scan, adding another hour to what was sold as a ten-second shortcut.
Interns feel it hardest. New hires fresh from university, already grappling with Amazon’s legendary onboarding fire-hose, are told to “trust the copilot.” They spend nights learning why the copilot tried to import a package deprecated in 2016. Mentorship hours evaporate; senior devs are too busy fire-fighting.
A cultural loop that feeds itself
Managers who question the ROI risk career stagnation. The AI task-force sits outside normal reporting lines and reports directly to VP level, giving it immunity from push-back. “Saying ‘this tool slows us down’ is interpreted as ‘I refuse to innovate,’” an anonymous principal engineer wrote on an internal forum post later scrubbed by moderators.
Meanwhile, Amazon’s outward narrative stays polished. CEO Andy Jassy’s last shareholder letter hailed generative AI as “the largest productivity swing since cloud computing.” Investors cheer; staff tighten their laces.
The contradiction is sharpening. Teams quietly maintain parallel codebases: an AI-blessed branch for metrics, and a hand-written version that actually ships. Double work, zero fanfare.
What happens when the babysitter outnumbers the baby?
Amazon is not retreating. Job ads for “AI code-curation specialists”—humans whose sole task is to babysit models—have ballooned 300 % year-over-year across its careers portal. The company is automating the automation oversight, layering meta-tools that flag suspect suggestions, creating a hall-of-mirrors where every fix invites another algorithmic review.
Outside observers see a cautionary tale. “If the sharpest tech giant on Earth can’t make AI coding net-positive, that’s a signal flare for the whole industry,” says RedMonk analyst Rachel Stephens.
Amazon’s engineers, for now, comply. They toggle the AI assistant, paste the mangled output, and file another ticket. The share price hovers near record highs. Somewhere in a server farm, a model logs the interaction as another success, training itself on the corrections it forced humans to write. The loop is closed, and it is very profitable.