Ai promised faster coding. inside amazon, it’s slowing engineers down.
First the bots wrote the code. Now the humans debug the bots. A Guardian investigation into Amazon’s forced roll-out of generative ai inside engineering teams shows the tools once billed as turbo boosters are becoming parking brakes: developers spend entire sprints cleaning hallucinated functions, patching security holes the model introduced, and writing long explainers for auditors who no longer trust a single green diff.
The numbers leaked to the newsroom are blunt. One Alexa-division engineer logged 37 hours re-writing an autocomplete suggestion that was supposed to handle voice-conversation context. “It looked confident, the syntax compiled, but it forgot every user after turn three,” the engineer told the paper. The final time stamp: four hours longer than the hand-coded baseline he had thrown away in week one.
How did “optional” become “non-negotiable”?
Internal dashboards track “ai leverage” the way sales teams track quarterly revenue. Managers receive weekly heat-maps showing who accepted the most machine suggestions. Miss the team’s 70 % adoption target and your performance review carries a red flag, according to screenshots shared by five staffers across Prime Video, Alexa and the retail platform. The result: engineers click “accept” on garbage snippets just to keep the metric green, then open a second tab to repair the damage before the commit reaches production.
Amazon is not alone. Similar scorecards circulate inside Google and Meta, but Seattle’s culture of writing press-ready “press release before the code” memos makes the contradiction starker. A 2022 mandate decreed that every new feature “must demonstrate ai acceleration.” Translation: even if the algorithmic route is slower, riskier and twice as buggy, you ship it and pray.
The knock-on effect is a quiet but growing revolt in pull-request comments. “LGTM after ai detox” has become shorthand for “I refactored everything the model hallucinated.” Veteran engineers schedule “review parties” where juniors parade their ai-generated branches and the room competes to spot the most creative error. Winners get coffee vouchers; losers stay until midnight fixing memory leaks.
Headcount fell 27 % while ai spend tripled
Amazon trimmed 27,000 jobs since 2022 even as it poured billions into Bedrock, CodeWhisperer and custom Titan models. The implicit deal: smaller, smarter teams amplified by silicon colleagues. Yet the productivity graphs presented at all-hands are flat. Defect rates, however, are up 14 % year-over-year in repos where AI penetration exceeds 80 %, according to internal QA slides obtained by The Guardian.
Outside observers see a familiar Silicon Valley loop. “We are watching the enterprise edition of the Diderot effect,” says Prof. Arvind Narayanan at Princeton. “New tools create new chores, and the old chores never leave.” First you need a copilot; next you need a copilot-review copilot; soon you need three headcount just to insure the first robot.
Amazon declined to comment on specific metrics, but spokesperson Brad Glasser reiterated that AI adoption “remains optional where appropriate.” Engineers scoff at the qualifier. “Optional the way a 7 a.m. email from your director is optional,” one told me over Signal.
What happens next is uncertain. The company still promotes “AI-written” as a marketing talisman to investors, yet quietly re-introduces senior-only code gates for payments and security modules. Internally, a proposal to add a “human-authorship” toggle to Git commits is gaining up-votes. The irony: developers would use an automation script to flip that toggle on every file, restoring the old workflow under a new label.
Bottom line: the bots are not taking the jobs; they are taking the joy. Engineers entered the field to solve fresh puzzles, not to babysit probabilistic parrots that invent APIs which never existed. Until the metrics change, the most valuable skill in Seattle is not coding—it is patience.