Six months ago, the story around AI coding tools was almost entirely about speed and adoption. In July 2026, a second story has caught up with it: what that speed costs. CodeRabbit’s review of 470 open-source pull requests found AI-generated code carries 1.7 times more defects than code written by humans. Stack Overflow’s 2025 Developer Survey found trust in AI tool accuracy fell to 33%, down from 43% the year before. GitClear’s latest maintainability research shows copy-pasted code has nearly doubled since 2022. None of these numbers come from AI skeptics. They come from the same research groups, code-review vendors, and developer surveys that spent 2023 and 2024 documenting the productivity gains. The tools got faster, and they kept getting faster still: Anthropic’s Claude Mythos Preview reached 93.9% on SWE-bench Verified by May 2026, according to Codersera’s AI agent benchmarks, a fresh high-water mark that landed in the same window trust kept falling. The acceleration had not let up five months later, either: by October 2026, Claude Opus 5.5 posted 66.4% on the newer, tougher Terminal-Bench 4.0, according to CodeSota. The code got harder to trust. What changed is not that AI stopped writing usable code. It is that enough of it has now shipped, and enough of it has now broken quietly in production, for researchers to measure the pattern instead of just guessing at it.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The Numbers Behind 2026’s AI Code Quality Reckoning
The headline figures are easy to state and hard to ignore. AI-generated code shows 1.7 times more defects per pull request than human-written code, according to CodeRabbit’s December 2025 analysis. Only 33% of developers trust the accuracy of AI tool output, per Stack Overflow, down from 43% a year earlier. Sonar’s 2026 State of Code survey puts the number lower still: 96% of more than 1,100 professional developers say they do not fully trust AI-generated code, and only 48% always check it before committing, according to Sonar’s research.
Put those together and a pattern appears. Reliance on AI-generated code keeps climbing, and the newest adoption data suggests it is accelerating even faster than Sonar’s own forecast anticipated. Sonar found AI already accounts for 42% of committed code industry-wide and expects that share to reach 65% by 2027, while Jellyfish’s engineering-benchmarks research found code-assistant adoption jumped to 72.8% of companies by August 2025, up from 49.2% just seven months earlier, and by December 2025 roughly half of organisations reported that AI now generates 50% or more of their code, up from only 20% in January of that year. Confidence in that code is moving the opposite way. That widening gap, more code shipped with less verification, is what researchers now call a verification bottleneck, and it sits at the centre of every report cited in this piece.
What ‘Silent Failures’ Mean, and Why They’re Worse Than a Crash
IEEE Spectrum gave this problem a name in a widely discussed January 2026 report: silent failures. A silent failure is code that runs without throwing an error, passes its tests, and still does not do what it was supposed to do. Nothing crashes. Nothing gets flagged. The bug sits there until it surfaces somewhere else, often in production, often much later than a normal crash would have appeared.
IEEE Spectrum’s reporting describes coding tasks that took roughly five hours with AI assistance in 2024 now taking seven to eight hours once 2026-era debugging time gets added in. That is not a rounding error. It suggests the speed advantage AI coding tools were sold on has partly reversed for a meaningful share of tasks, once someone has to locate and fix what the model got quietly wrong.
A Simplified Example
The pattern researchers describe is easiest to see in a small, illustrative example rather than an abstract description.
def get_discount(user_tier):
discounts = {"gold": 0.20, "silver": 0.10}
return discounts.get(user_tier, 0.10) # silently defaults instead of raisingThe function never crashes. A typo such as a trailing space on “silver “, or a new membership tier the table was never updated for, quietly returns the wrong discount instead of failing loudly. Static analysis generally will not flag it, because nothing about the code is syntactically wrong. That is the essence of a silent failure: correct-looking code, wrong behavior, no alarm.
Inside CodeRabbit’s Study: AI Code Carries 70% More Bugs
CodeRabbit, an AI code-review company, published the most detailed breakdown of the problem in its State of AI vs Human Code Generation report. Researchers there compared AI-generated and human-written code across 470 open-source pull requests using a structured issue taxonomy, and the gap held across nearly every category measured.
AI-authored pull requests averaged 10.83 issues each, compared with 6.45 for human-authored ones, a 1.7 times difference. Logic and correctness errors came in 1.75 times more common in AI code. Quality and maintainability problems ran 1.64 times higher. Security findings appeared 1.57 times more often, and performance issues 1.42 times more often.
The Security Breakdown
The security figures are where the numbers turn genuinely alarming. AI-generated code was 2.74 times more likely to introduce cross-site scripting vulnerabilities, 1.91 times more likely to create insecure object references, 1.88 times more likely to mishandle passwords, and 1.82 times more likely to use insecure deserialization, per CodeRabbit’s findings. A separate 2025 benchmark from Synchsoft and SusVibes puts the same problem in even starker terms: across 200 coding tasks, the best-performing agent produced functionally correct code 61% of the time but secure code only 10.5% of the time, a gap that shows passing tests and shipping safe code are not the same achievement. Even purpose-built security models are still catching up: Google’s Flash Cyber variant of Gemini 3.8, released in September 2026, scored just 47.2% on CWE-Bench, the benchmark built specifically to measure how well models avoid known vulnerability classes. None of these are exotic bugs. They are exactly the defect types a security review is built to catch, which is why AI-heavy codebases need more review, not less.
| Defect Category | AI-Generated Code | Human-Written Code |
|---|---|---|
| Average issues per pull request | 10.83 | 6.45 |
| Logic and correctness errors | 1.75x higher | Baseline |
| Quality and maintainability issues | 1.64x higher | Baseline |
| Security findings | 1.57x higher | Baseline |
| Performance issues | 1.42x higher | Baseline |
| Critical-severity issues | 1.4x higher | Baseline |
| Major-severity issues | 1.7x higher | Baseline |
| Cross-site scripting flaws | 2.74x higher | Baseline |
| Insecure deserialization | 1.82x higher | Baseline |
| Insecure object references | 1.91x higher | Baseline |
GitClear’s Maintainability Gap: Copy-Paste Code Nearly Doubles
GitClear tracks code health across millions of commits every year, and its 2026 report, The Maintainability Gap, documents a shift in how code actually gets written once AI assistants enter the workflow. The firm had already flagged the direction of travel back in February 2025, when it found AI assistants were driving four times more code cloning across a 211-million-line sample, according to DevClass‘s reporting on the research. The 2026 follow-up shows the trend has only deepened: copy-pasted code rose from 9.4% of new code in 2022 to 15.7% in the first half of 2026, while moved, properly refactored code fell from 21% to just 3.8% over the same stretch.
Within-commit copy-paste is up 41% and duplicated code blocks are up 81%, GitClear found. Constructs that mask errors instead of handling them, catch blocks that swallow exceptions rather than deal with them, rose 47%. Two-week code churn, the share of code rewritten within fourteen days of being written, climbed another 15%.
The reverse trend is just as telling. Cross-file function calls, a rough proxy for genuine code reuse, dropped 35%. Refactoring line moves fell 70%. Work on legacy code maintenance is down 74% compared with 2022. Developers are increasingly pasting new code instead of reusing or cleaning up old code, and GitClear’s data suggests AI assistants are accelerating that habit rather than correcting it.
| Metric | 2022 | H1 2026 |
|---|---|---|
| Copy-pasted code (share of new code) | 9.4% | 15.7% |
| Moved / refactored code (share of new code) | 21% | 3.8% |
| Within-commit copy-paste | Baseline | +41% |
| Duplicated code blocks | Baseline | +81% |
| Error-masking constructs | Baseline | +47% |
| Two-week code churn | Baseline | +15% |
| Cross-file function calls | Baseline | -35% |
| Refactoring line moves | Baseline | -70% |
| Legacy code maintenance activity | Baseline | -74% |
Developer Trust Is Collapsing, and the Surveys Agree
Three separate surveys published over the past year point to the same conclusion from different angles. Stack Overflow’s 2025 Developer Survey is the most direct measure of sentiment.
“33% of developers trust the accuracy of AI tool outputs in 2025, down from 43% in 2024.”
Stack Overflow, 2025 Developer Survey
Distrust grew even faster than trust fell.
“46% of developers say they do not trust the accuracy of AI tool outputs in 2025, up from 31% in 2024.”
Stack Overflow, 2025 Developer Survey
That distrust is not just an attitude, it shows up as lost time. Stack Overflow’s researchers found that “45% of developers say debugging AI-generated code takes longer than writing it themselves”, and the broader mood has soured alongside it: “positive sentiment for adding AI tools into developer workflows fell from 72% in 2024 to 60% in 2025.”
Sonar’s 2026 State of Code survey, covering more than 1,100 professional developers, found the sharper end of the same trend. 96% do not fully trust AI-generated code, and only 48% always verify it before committing. 61% agree that AI often produces code that looks correct but is not reliable, which is close to a plain-language definition of a silent failure.
METR’s Productivity Paradox: Feeling Faster, Measuring Slower
METR, the research group best known for testing how much AI actually speeds up software engineers, published an updated study in February 2026 that complicates the productivity story further. An earlier version of the research suggested AI tools were slowing experienced developers down on complex tasks, even though those same developers believed they were roughly 20% faster.
The February update corrected for a selection problem: a large share of developers refused to take part in tasks that required working without AI. METR reported that 30 to 50% of developers declined to submit tasks without AI assistance, even when offered $50 an hour to do so, as TechCrunch reported in May 2026. Once METR adjusted its cohort, to 57 developers across more than 800 tasks, the measured slowdown narrowed to just 4%, with a confidence interval wide enough to include a small speed-up.
The detail that matters most here is not the final percentage. It is that developers would rather skip a paid task than do it without AI. That says as much about dependency as the defect-rate numbers say about quality.
The Governance Gap: Why 92% of Firms Can’t Control Their Own AI Code
If the quality data explains the problem, GitLab’s 2026 AI Accountability Report explains why so few companies are catching it early. The report, based on a Harris Poll survey of 1,528 developers and technology buyers, found that 92% of organisations report governance challenges with AI-generated code, and 80% admit they adopted AI coding tools faster than they built policies to manage them, according to reporting on the study.
The most striking single figure may be this: 43% of organisations cannot reliably tell which code in their own systems was written by AI and which was written by a person. Without that basic visibility, targeted review of higher-risk AI output is close to impossible. GitLab also found 82% of respondents believe AI-generated code is creating a new category of technical debt, and 91% plan to invest in AI code governance tools within twelve months.
Tech Insider covered this governance gap in detail in June, and the quality data that has emerged since then reads like the follow-up chapter. It is not just that companies cannot govern AI code. It is that the code itself is measurably harder to govern.
How Claude Code, GitHub Copilot, and Cursor Compare on This Problem
The three tools fighting hardest for the AI coding assistant market, Claude Code, GitHub Copilot, and Cursor, are not named directly in most of the quality research, which tends to treat AI-generated code as a single category rather than break results out by vendor. That is notable on its own, given how much attention has gone into comparing them on adoption, pricing, and now raw revenue instead: Second Talent reported that Claude Code’s run-rate revenue had already topped $2.5 billion by February 2026, a scale that makes the absence of vendor-specific defect data even harder to justify.
Tech Insider’s own coverage has tracked Claude Code pulling level with Cursor at 18% workplace adoption in early 2026, while GitHub Copilot still leads at 29%. Cursor’s commercial trajectory has outrun that adoption share, though: Second Talent reported the company surpassed $1 billion in annualized revenue and raised a further $2.3 billion in November 2025, a war chest well out of proportion to its workplace-adoption percentage alone. GitHub, for its part, has already shifted Copilot to usage-based billing, with GitHub AI Credits replacing premium requests on June 1, 2026. Business now costs $19 per seat per month and Enterprise $39, and the model list includes GPT-6.1 Sol and Gemini 4 Argon, a change that indirectly puts a price on the extra review work heavier AI usage now demands. The underlying model race has kept accelerating regardless of the trust numbers: by July 2026, Anthropic’s Claude Opus 5 was rated 99 on Swfte’s leaderboard with a 1522 Elo score and roughly 74 tokens per second, while OpenAI’s GPT-5.6 Sol posted 88.8% on Terminal-Bench at around 750 tokens per second, according to BuildFastWithAI and Swfte. Two months on, Anthropic pushed Claude Opus 5 further still, to 97.0% on SWE-bench Verified by September 2026, according to DataCamp, and OpenAI answered that same month with an upgraded GPT-6.1 Sol, per MarkTechPost’s reporting. Claude Code itself has moved fast within that race: MarkTechPost clocked Claude Code running on Opus 4.7 at 87.6% on SWE-bench Verified in May 2026, BestAI’s own testing that same May 1-12 window put Claude Code at 91.2% on HumanEval+, 52.3% on SWE-bench, and 87.3% overall, and by June 2026 Claude Opus 4.6 was scoring 80.8% on SWE-bench single-attempt and 65.4% on Terminal-Bench 2.0, according to Tech Insider’s reporting citing Spectrum AI Labs. Earlier in the year the same sprint was already visible: GPT-5.5 posted 88.7% on SWE-bench Verified in April 2026 per the Andrew.ooo leaderboard, and by May Anthropic’s own Claude Mythos Preview had reached 77.8% on the tougher SWE-bench Pro while Claude Opus 4.5 scored 89.4% on the 225-task, six-language Aider Polyglot benchmark, according to Codersera’s AI agent benchmarks. None of those figures measure defect rates. Separately, Cognition’s Devin has pushed further into autonomous territory. The company says Devin now writes 89% of its own code without direct human authorship, which raises the stakes on exactly the review question CodeRabbit and Sonar are measuring. Cognition kept its own model line moving too, releasing a new coding model called SWE-2 in September 2026 that landed within one point of Anthropic’s Claude Fable 5.1 on benchmark scores, according to AI/TLDR.
How Copilot’s Model Lineup and Billing Stand Now
GitHub replaced the former premium-request system with usage-based GitHub AI Credits on June 1, 2026, and one AI credit equals $0.01. The model menu has kept moving since. GPT-6.1 Sol was added to GitHub Copilot on September 29, 2026, billed at provider-list pricing under usage-based billing. Claude Haiku 5.5 followed on October 7, 2026, and is available to Copilot Pro, Pro+, Max, Business, and Enterprise users. Gemini 4 Argon was added on October 13, 2026, for Copilot Pro+, Max, Business, and Enterprise users.
Costs for teams are easier to read than they used to be. Business is $19 per seat per month and includes 1,900 GitHub AI credits per user per month, while Enterprise includes 3,900 credits per user per month, with additional usage reported at $0.01 per AI credit. GitHub’s own page says a $10 budget covers 1,000 credits, and that alerts fire at 75%, 90%, and 100% of any configured budget. GitHub also lists a deprecation of several models across Copilot Chat, inline edits, ask and agent modes, and code completions on November 16th, 2026. For reviewers, that means the model behind a given pull request can change quickly, so teams that label AI-written code should also record which model produced it.
None of this means one tool is measurably worse than another on defect rates, since no public study has broken the numbers out that way. It means the competitive fight has expanded. Speed and pricing got these tools this far. Verification and reliability look set to decide what happens next.
From 2023 Hype to 2026 Reckoning: A Short History
The arc here is short but sharp. GitHub Copilot’s general availability in 2022 and its rapid enterprise rollout through 2023 set the terms of the conversation: AI pair programmers that made individual developers faster. Early GitHub-backed research found developers completing tasks up to 55% faster with Copilot than without it, and that framing carried through most of 2023 and 2024.
The first warning signs came from GitClear, which flagged rising code churn and falling refactor rates as early as its 2024 research. Stack Overflow’s 2025 survey was the first major, developer-reported signal that trust was moving the other way. Cursor and Claude Code’s rise through late 2025 added a second complication: newer, more agentic tools that write larger chunks of code at once, which means a single silent failure can now touch far more of a codebase than a single autocomplete suggestion ever could. The raw capability curve tells its own version of that story: Presenc AI found coding agents climbing to 70-78% on SWE-bench Verified by May 2026, up from just 49% in early 2025, a jump in headline competence that arrived well before the review and governance tooling needed to keep pace with it. Then came the January 2026 IEEE Spectrum report on silent failures, followed within months by GitClear’s 2026 maintainability data, CodeRabbit’s defect-rate study, GitLab’s governance survey, and Sonar’s verification-gap findings. The capability race never slowed down to match the caution: Google’s Gemini 3.1 Pro topped LiveCodeBench Pro with a 2887 Elo score in May 2026, per Codersera, the same month SWE-bench’s own maintainers released ProgramBench, a new benchmark for grading whole software artifacts rather than single-function patches. Anthropic followed with Claude Fable 5 reaching 95% on SWE-bench Verified, that same May Logical Intelligence’s Aleph model hit 94% on VeriSoftBench against rival Aristotle’s 69% and turned in a perfect 100% on the Verina formal-verification benchmark, according to PRNewswire and Yahoo Finance, OpenAI’s GPT-5.5 had already posted 82.7% on the newly launched Terminal-Bench 2.0 that April, per MarkTechPost, and Z.ai’s MIT-licensed GLM-5.2 scored 82.7% on Terminal-Bench that same July, per BuildFastWithAI. Snorkel AI added its own yardstick that same month, July 2026, with Senior SWE-Bench, a 100-task benchmark drawn from 12 repositories built to test senior-level engineering judgment rather than isolated bug fixes. By September 2026, even the benchmarks themselves were being recalibrated: Scale AI’s SWE-Bench Pro V2 trimmed its public task set from 731 to 642, a sign the underlying problem set keeps shifting as models saturate it. The saturation kept compounding the following month: Claude Opus 5.5 scored just 66.4% on the newly introduced Terminal-Bench 4.0 in October 2026, CodeSota reported, evidence that benchmark designers were raising the difficulty bar about as fast as models cleared the old one. The field has also grown crowded enough to strain any single scoreboard: an October 2026 LeanZero tally tested 44 models across 43 agentic-coding benchmark entries. All of it, benchmark gains and benchmark revisions alike, says nothing about the defect rates piling up behind them.
Four years after Copilot’s debut, the industry has moved from asking how much faster AI makes developers to asking how much of that speed was borrowed against quality it has not yet paid back.
The Market and Business Impact for Engineering Teams
The business consequences are already visible in where money is moving. AI code-review and verification tooling, CodeRabbit, Sonar, and similar products, sits in one of the few corners of the developer-tools market where demand is rising alongside the concerns driving it, and ACR reported in October 2026 that AI code-review pricing ranged from $12–$160 per developer monthly. GitLab found 91% of organisations plan to invest in AI code governance tools within a year, and 98% have already allocated or expect to allocate budget for it.
That is a direct cost that was not in most engineering budgets two years ago. Review time is climbing to offset defect rates that run 1.7 times higher than human-written baselines, and some of the raw time savings from AI generation are being eaten by exactly that extra review and debugging work, per IEEE Spectrum’s reporting.
There is a second-order effect worth watching too. As usage-based billing spreads across the market, the cost of an AI-heavy workflow is becoming visible line by line instead of hidden inside a flat subscription. By July 2026 that line-item pricing spanned a wide range: OpenAI’s GPT-5.6 Sol billed at $5 per million input tokens and $30 per million output tokens (a price since undercut by GPT-6.1 Sol, which launched in September 2026 at $2/$10 per 1M tokens), Anthropic’s Claude Opus 5 came in slightly cheaper than GPT-5.6 Sol at $5 in and $25 out, and budget-tier models such as DeepSeek’s V4 Flash undercut both of those at just $0.28 per million output tokens, according to Capital & Compute, ThursdAI News, and Morph. That budget tier is not just cheap, it is competitive on raw coding benchmarks too: DeepSeek-Coder V3 reached 91.5% on HumanEval and 40.5% on BigCodeBench in May 2026, while Alibaba’s Qwen2.5-Coder-32B posted 92.5% on HumanEval and 38.2% on BigCodeBench over the same period, according to Presenc AI, figures that undercut the pricier frontier models on cost without conceding much on capability. Open-weight releases pushed the same way through late summer: Alibaba launched Qwen3.8-Max in August 2026 and its 2.4T-parameter weights followed 9 days later, while Xiaomi released MiMo-V2.6-Pro and Flash under the MIT license in September 2026. Not every vendor is playing that pricing game in the open, though: MiniMax launched M3.1-Flash-Preview inside its own Code agent in September 2026 without publishing a public price at all, Startup Fortune reported, leaving buyers to negotiate costs that competitors post upfront. A team generating more code that then needs more review and more regeneration is not just slower to ship. It is now measurably more expensive to run, in ways finance departments can finally see on an invoice, with the choice of model itself becoming a budget lever.
Hiring is starting to reflect the same shift. Job listings for senior code-review and platform-engineering roles increasingly mention AI-output verification as a core responsibility rather than a side task, and engineering leaders now cite reviewer capacity, not generation speed, as their main scaling constraint. A market that spent three years hiring for who could produce the most code is quietly starting to hire for who can be trusted to catch what the machines get wrong.
Insurers and enterprise buyers are asking sharper questions too. Software procurement teams increasingly want to know whether a vendor’s codebase includes AI-generated components and what review process those components went through, mirroring the due-diligence questions that followed earlier waves of open-source dependency risk.
What This Means for Developers and Employers in Ireland
Ireland’s position in this story is bigger than its size would suggest. Dublin hosts European headquarters or major engineering hubs for several of the companies at the centre of this data, including Microsoft, which owns GitHub, alongside a dense cluster of software employers that have spent the past three years pushing AI coding tools into daily workflows.
That concentration cuts both ways. Irish-based engineering teams get early access to whatever governance tooling and review processes vendors build in response to this data. They are also more exposed than most to the EU’s regulatory direction, since the EU AI Act’s transparency and risk-management obligations push in the same direction as GitLab’s governance findings, even though the Act does not target AI-generated source code specifically.
For Irish employers, the practical takeaway from this wave of research has less to do with which coding assistant to choose and more to do with whether a review process exists that assumes AI output needs checking, rather than one built for the assumption that it mostly does not. Irish software exporters selling into the EU single market face an added incentive to get this right early, since compliance requirements built around one jurisdiction’s rules tend to become the de facto standard for a product sold across all of them.
How Engineering Teams Are Responding Right Now
The response so far splits into three categories. The first is tooling. Dedicated AI code-review products are adding rules specifically aimed at the defect categories CodeRabbit and Sonar have documented, from error-masking constructs to the security flaws that showed up 1.4 to 2.74 times more often in AI-generated code.
The second is process. Some engineering organisations are moving toward mandatory human review for any AI-generated pull request above a certain size or risk level, rather than treating AI suggestions as equivalent to human-written code for review purposes. GitLab’s finding that 43% of companies cannot even identify which code was AI-written suggests most teams are still a step behind this shift, since labelling has to come before targeted review can work at all.
The third is cultural, and slower to measure. Sonar’s finding that only 48% of developers always verify AI code before committing suggests review habits have not caught up with the defect data, even at teams that already know the numbers. Closing that gap looks less like a tooling problem and more like a discipline problem, which is typically the harder one to fix.
Five Predictions for AI Code Quality Through 2027
Based on the direction of this data, here is where the AI code quality story looks likely to go next.
- AI code-review spending outpaces AI code-generation spending. With 91% of firms already planning governance investment, per GitLab, verification tooling looks set to grow faster than the generation tools it checks.
- Code provenance tracking becomes standard, not optional. Once 43% of firms admit they cannot tell AI code from human code, tagging the source of every commit turns into a compliance requirement as much as an engineering one.
- Vendors start marketing reliability over raw speed. Expect GitHub, Anthropic, and Cursor to lean harder on defect rates and verification claims in their marketing, echoing how cloud providers eventually sold uptime rather than raw compute power.
- A named production incident becomes the industry’s reference case. Every past software reliability shift has had one high-profile failure that made the abstract numbers concrete. AI-generated silent failures are overdue for theirs.
- Regulatory language starts referencing AI code accountability directly. The EU AI Act’s existing transparency requirements are a plausible template for future rules covering AI-generated source code, building on the direction GitLab’s research already points toward.
Frequently Asked Questions
What does “silent failure” mean in AI-generated code?
A silent failure is code that runs without crashing or throwing a visible error, yet does not do what it was supposed to do. IEEE Spectrum’s January 2026 report on the topic found these failures are becoming more common in newer AI coding models, and they are harder to catch than a crash because nothing in the program’s output signals that anything went wrong.
Is AI-generated code really less reliable than human-written code?
On the metrics researchers have measured so far, yes. CodeRabbit’s analysis of 470 open-source pull requests found AI-generated code averaged 1.7 times more issues than human-written code, with logic errors, security flaws, and maintainability problems all elevated.
Why don’t developers trust AI-generated code anymore?
Trust has fallen because the defect data has started to catch up with the speed claims. Stack Overflow’s 2025 survey found trust in AI tool accuracy dropped to 33% from 43% the year before, while Sonar’s 2026 survey found 96% of developers do not fully trust AI-generated code.
Does this mean AI coding tools are getting worse?
Not exactly. IEEE Spectrum’s reporting suggests the underlying models plateaued in quality through 2025, even as their fluency and confidence increased, which makes flawed output look more convincing rather than less common.
Which AI coding assistant has the fewest bugs?
Public defect-rate research so far treats AI-generated code as one category rather than breaking results out by GitHub Copilot, Claude Code, or Cursor individually, so there is no verified, vendor-specific ranking on bug rates as of mid-2026. Buyers comparing tools today are largely relying on adoption figures, pricing, and benchmark scores rather than defect data, which is exactly the gap the next generation of research is likely to fill.
How are companies responding to the AI code quality problem?
GitLab’s 2026 research found 91% of organisations plan to invest in AI code governance tools within twelve months, and many engineering teams are introducing mandatory human review for larger AI-generated pull requests.
What should Irish and EU-based developers watch for next?
Expect governance and provenance-tracking requirements to firm up alongside the EU AI Act’s broader transparency rules, particularly at the multinational engineering hubs based in Dublin and elsewhere in the EU.
