Nearly half of AI‑written code needs debugging

Writing software with artificial intelligence is now common, but getting that code to behave after release is still a manual job. A 2026 Lightrun survey found 43% of AI-generated code changes still need debugging in production. (lightrun.com) Lightrun said it surveyed 200 senior site reliability engineering and DevOps leaders at large enterprises in the United States, United Kingdom, and European Union. The company released the findings in its 2026 State of AI-Powered Engineering report, published in April 2026. (lightrun.com; venturebeat.com) The report says 88% of organizations need two to three redeploy cycles to publish a single AI-generated change, and 11% need four to six. VentureBeat also reported that zero respondents said they were “very confident” AI-generated code would behave correctly once deployed. (lightrun.com; venturebeat.com) Production is the live version of software that real users touch, and debugging there is the expensive part because failures hit customers, revenue, or both. Lightrun said 44% of artificial intelligence site reliability engineering or application performance monitoring tool failures happened because the tools did not capture execution-level data, meaning the step-by-step behavior of running code. (lightrun.com) That leaves a gap between code that passes quality assurance and staging tests and code that holds up in the real world. Lightrun framed that gap as a “trust wall,” with teams shipping more machine-written code faster than they can verify it in live systems. (lightrun.com) The backdrop is rapid adoption. Google Cloud’s 2025 DevOps Research and Assessment report said 90% of technology professionals now use artificial intelligence at work, and more than 80% reported productivity gains. (services.google.com) That same Google Cloud report said higher artificial intelligence adoption was associated with higher software delivery throughput but also with lower delivery stability and more burnout. It also found that about 30% of respondents had little or no confidence in AI-generated code. (services.google.com) Some of the operational risk has already shown up in public. VentureBeat reported that Amazon had two major storefront outages on March 2 and March 5, 2026, and said both were traced to AI-assisted code changes deployed without proper approval. (venturebeat.com) After those incidents, VentureBeat said Amazon started a 90-day code safety reset across 335 critical systems and required senior engineer approval for AI-assisted code changes before deployment. The immediate lesson from Lightrun’s survey is narrower: writing code faster has not removed the work of proving it works. (venturebeat.com)