Published August 6, 2026 · Updated September 23, 2026

Two studies, two vantage points on AI code

The software industry has gone all-in on AI. Executives expect a productivity leap and a large return, and widespread adoption has followed. With AI now authoring the majority of the code at most tech organizations, the increased output feels like progress. Whether it performs like progress is the question two recent research efforts set out to answer, and they came at it from opposite directions.

New Relic’s 2026 State of AI Coding report surveyed 200 technology leaders about how AI-generated code is performing in their organizations. Faros’s latest AI Engineering Report, The Speed Trap, analyzed the most recent 12 months of telemetry from 22,000 developers across 4,000 teams. One study captures what leaders perceive; the other captures what engineering systems record.

Read together, they reveal a more complicated quality story. AI is accelerating delivery, and some of the worst transition-era effects are easing. But larger changes, skipped reviews, QA bottlenecks, and rising incident volume show that faster delivery has not eliminated the downstream cost of verifying and operating AI-generated code.

{{cta}}

Where the two studies agree

Faros’s earlier Acceleration Whiplash research documented what happened as organizations moved from low to high AI adoption. Engineering throughput increased faster than the systems around it could adapt, driving sharp increases in bugs, incidents, churn, and workflow delays.

The latest data shows what happens next. Engineering organizations are getting better at turning AI-generated output into shipped software, and the incident-to-PR ratio is no longer deteriorating at the rate observed during the initial transition. But the pipeline remains leaky. More work is reaching production while verification, remediation, and operational pressure continue to accumulate downstream.

New Relic found the same strain in its survey results. It reported that 78% of organizations saw production incidents climb, 86% saw senior-engineer rework grow, and 74% of leaders said at least a quarter of their organization’s AI-generated code needs significant rework.

Faros’s latest telemetry adds important context. The incident-to-PR ratio increased just 14.5%, compared with 242.7% in the April dataset, suggesting that per-change quality has stabilized significantly. Yet monthly incidents rose 125.4%, mean time to remediate increased 25.8%, and the bugs-to-PR ratio increased 43.9%. Individual changes may be safer than they were during the initial adoption transition, but the total cost of operating the system continues to rise as more changes move through it.

Oversight has also weakened. New Relic reported that 62% of teams often ship AI-generated code without line-by-line verification. Faros found that unreviewed merges increased 76.3%, up from 31.3% six months earlier. The latest data therefore sharpens the quality mirage: code can become safer per change while the overall engineering system becomes more difficult and expensive to verify, remediate, and operate.

The perception-reality gap found in New Relic’s survey

New Relic’s most striking survey finding highlighted leaders’ attitudes toward AI-generated code. A remarkable 94% rated AI-generated code as higher quality than human-authored code at the point of review; 61% of leaders rated AI-generated code as “somewhat higher quality” than human-authored code, and the other 33% rated it as “much higher.” That confidence held steady even as the same leaders reported the increased incidents, rework, and firefighting above. 

The people closest to the budget believe AI-generated code is excellent, even as their teams report more incidents, rework, and firefighting downstream.

Why AI code looks better than it runs

The explanation lives in how AI writes. AI-generated code is both verbose and polished-looking. It is idiomatic, well-named, and pretty consistent with the code around it. It reads like the work of someone who understands the system, which makes it harder for engineers to spot issues during review. The gaps stay beneath the surface until the code runs against real traffic, real dependencies, and real edge cases—only to emerge later as a spike in incidents, bugs, and rework. 

AI’s impact on senior engineers and time in QA

The latest data shows that the bottleneck has shifted again. Time in progress at the authoring stage increased 50.7%, down sharply from 225.2% in the previous dataset. But time in QA increased 300.6%, the largest deterioration among the major flow metrics Faros measured.

The work arriving downstream is also harder to verify. Average PR size increased 71.8%, files changed per PR increased 52.0%, and unreviewed merges increased 76.3%. When larger changes bypass review, the verification burden does not disappear. It moves into QA, remediation, and production, where failures are more expensive and senior engineers are often the ones equipped to diagnose them.

The senior-engineer tax therefore persists, but it is no longer confined to the review queue. Senior engineers are absorbing the combined cost of reviewing larger changes, resolving QA failures, responding to incidents, and reconstructing work that took the wrong path.

The AI code quality mirage

This is the mirage. AI-generated code looks flawless at review and reveals its true state in production. And the same trick plays out one level up: increased AI usage looks like increased productivity, so executives keep pushing more AI into more of the work, while the delivered business value stays somewhere off in the distance. Even as AI coding costs are skyrocketing, both engineers and executives are reading the shimmer on the surface, and the surface looks superficially convincing.

The industry has placed its bet on the appearance of progress. AI has become a tool tech companies depend on, and it now sits squarely at the center of how software gets built. It is also, today, a tool still leaving much to be desired in terms of quality, consistency, and reliability. 

Faros’s data pinpoints these challenges as an authoring problem, not a review problem, so the fix belongs upstream at code generation; adding reviewers, gates, or QA just treats the symptom without addressing the root cause. Engineering companies should first pinpoint how AI is being used and where it could offer the most value. Then, they can improve code quality right at the source by equipping AI with richer context and guardrails, before finally monitoring and governing the entire AI-augmented pipeline.

The distance between how good AI code looks and how well it holds up is the space where engineering organizations are paying the cost right now, in incidents, in rework, and in the senior engineers spending their hours cleaning up after it. The measure that matters is the one underneath—what the code does once it ships, and whether the business is actually achieving better outcomes as a result.

Faros can help you maximize what your AI ships. Reach out for a demo to see how.