AI-written code is spreading faster than teams can verify it, report finds

Most software development teams have merged AI-generated code or tests they did not fully understand, according to a new TestMu AI report. The company says 44% of surveyed teams are seeing more missed defects or production incidents as AI-written code becomes more common.
- 81% of teams merged AI-generated code or tests they did not fully understand.
- AI writes more than half of production code in 40% of teams.
- Only 41% perform automated checks on AI-written code before merging.
- 44% report more missed defects or production incidents.
- Only 13% have an incident response plan for AI errors.
AI is taking a larger role in software development
TestMu AI, formerly LambdaTest, published its State of AI in Quality Engineering 2026 report on 13 November 2026. The study found that 81% of software development teams had integrated code or tests generated by AI that they did not fully understand.
AI now writes more than half of production code for 40% of teams, according to the report. At the same time, 97% of the teams surveyed use AI in testing or plan to do so. This means AI is increasingly involved in both producing software and checking it, while questions remain about how those outputs are validated before deployment.
Developers appear particularly exposed to the problem. TestMu AI says 93% of developers have merged AI outputs they did not fully understand.
Verification often comes after the code is written
Only 41% of teams perform automated checks on AI-written code before it is merged. More teams rely on manual review, at 55%, or pass the code to quality assurance after it has been written, at 51%.
The report links separate continuous integration testing with fewer missed defects. Teams that do not conduct separate CI tests reported rising missed defects more often than other teams, at 49% compared with 38%.
Using one AI model to judge another also appears to create a verification gap. Teams using an LLM as a judge regularly merged AI outputs they did not understand at twice the rate of other teams, 29% compared with 15%.
AI-generated tests were described as “almost correct, but not quite” in 65% of cases on average, according to TestMu AI.
AI agents and incident planning remain weak points
The report says 77% of teams are building AI agents that need testing. However, 37% have no systematic way to evaluate those agents, and only 28% test them against users who try to break or defeat them.
Spending is also weighted toward generating AI outputs rather than verifying them. Half of the teams allocate most of their AI testing budget to producing AI results, while 17% allocate most of it to verification. Only 13% have an incident response plan for AI errors.
TestMu AI co-founder and CEO Asad Khan said the company believes AI outputs should be checked by a system independent of the model that produced them, before merging and whenever changes are made. The report suggests that teams adopting this approach see fewer defects reach production, according to the company.
For organizations using AI to write code, create tests or operate agents, the findings point to a growing need for independent checks before deployment. TestMu AI says its report is available at testmuai.com/research/state-of-ai-in-quality-engineering.

