Synthesia and Amazon Overhaul Code Reviews to Fight AI Slop

AI coding tools are shifting the engineering bottleneck from writing code to reviewing it. In a Sonar survey of over 1,100 developers, AI generated 42 percent of code, yet 96 percent of respondents do not fully trust it. Additionally, 61 percent call AI code unreliable, and 38 percent say reviewing it requires more effort. This trust gap has boosted startups like CodeRabbit, which raised $143 million at a $1.5 billion valuation to run 2 million weekly reviews for 17,000 customers, including Nvidia, Indeed, and BMW.

At Synthesia, 118 engineers adopted Claude Code in November 2025, causing pull requests to surge 120 percent year-over-year, with 95 percent containing AI code. This led to issues like finding 10 duplicate versions of one function. At Bonterra, with 290 engineers, proposed changes tripled within three months of adopting AI, causing code entering review to rise tenfold and tripling review times. To manage this, companies use AI agents for initial checks, though fewer than 5 percent of changes bypass human review at Synthesia.

Practitioners are also moving upstream. At Amazon Stores, a 70-person team supporting over 1,000 developers uses AI to modernize 17 years of mobile shopping app code. Engineers now focus on writing precise specifications before generation. In one case, a missing instruction caused an agent to generate 25,000 lines of code in the wrong Swift version, producing 600 errors. Rewriting the specification and restarting the agent resolved the issue in 15 minutes.

This shift redefines engineering careers. To prevent superficial approvals, Temporal now requires developers to explain their AI agent's design choices or face rejection. Meanwhile, IBM uses AI to help junior engineers perform 70 to 80 percent of tasks once reserved for seniors, pushing them to audit AI outputs rather than writing basic code from scratch.