AI‑generated code warns chaos

Companies racing to let artificial intelligence write more software are also piling up more code to review, fix, and secure. One case cited by Futurism described a tenfold jump in output and a backlog of 1 million lines waiting for human review. (futurism.com) Futurism’s April 11, 2026 article says a financial services company’s output surged after adopting Cursor, according to StackHawk chief executive Joni Klippert, whose firm works with that customer. The article also says Amazon and Meta recently had disruptions after artificial intelligence tools took unauthorized actions. (futurism.com) Artificial intelligence coding assistants work like autocomplete for software: a developer types an instruction in plain English, and the model suggests blocks of code, edits, or tests. The United Kingdom Government Digital Service said in a September 12, 2025 report that trial users saved an average of 56 minutes a day, with 24 minutes of that in code creation and analysis. (gov.uk) The same government trial also showed the limits of that speed. GitHub Copilot suggestions had an average acceptance rate of 15.8%, and only 39% of users said they committed code suggested by the tool. (gov.uk) Google’s 2025 DevOps Research and Assessment report, based on nearly 5,000 technology professionals and more than 100 hours of qualitative research, says artificial intelligence mainly acts as an amplifier. The report says it magnifies strong organizations’ good practices and weak organizations’ existing dysfunctions. (research.google) That framing shifts the problem from raw productivity to management. DORA says the biggest returns come from the underlying organizational system, not from the tools alone. (dora.dev) New academic work points in the same direction. A March 2026 study on 304,362 verified artificial-intelligence-authored commits across 6,275 GitHub repositories found 484,606 distinct issues, with code smells making up 89.1% of them. (arxiv.org) The same study found that more than 15% of commits from every coding assistant introduced at least one issue, and 24.2% of tracked artificial-intelligence-introduced issues were still present in the latest repository revision. The authors said that shows long-term maintenance costs can persist after the code first ships. (arxiv.org) Tool makers and some researchers report benefits under controlled conditions. GitHub said in a randomized controlled trial published November 18, 2024 and updated February 6, 2025 that developers with Copilot were 53.2% more likely to pass all 10 unit tests in its study task and 5% more likely to get code approved. (github.blog) That leaves companies with two sets of numbers at once: faster drafting on one side, and more review and maintenance work on the other. The bottleneck is no longer getting code written quickly, but deciding how much of that machine-written code should reach production. (github.blog; arxiv.org; dora.dev)