Article · 8 min read

AI coding tools were supposed to free developers. Many say they feel trapped instead.

Published September 2026

The story AI companies told about coding assistants was appealingly simple: give developers a smart autocomplete, and they ship more software, faster, with less tedium. GitHub, Cursor, Anthropic, and OpenAI have all published numbers to back this up. The tools, by their own account, are transforming software development.

So it is worth paying attention when a large share of the people actually using these tools say something quite different.

A new Coddy Developer Survey found that four in five developers, 80%, say their use of AI has felt more like a dependence than an advantage. That is not a fringe complaint. And it sits alongside a growing pile of independent data that complicates the tidy productivity narrative.

What developers are actually reporting

Respondents to the Coddy survey cite the loss of natural stopping points, such as waiting on reviews or hitting mental walls, as a driver of daily fatigue and longer work sessions. A separate survey adds weight to this: LeadDev's own 2026 leadership survey found that 45% of engineers now work more hours per week than the prior year.

The numbers on trust are striking too. 84% of developers use AI coding tools daily, but only 29% trust the output. That gap, between constant use and consistent distrust, is a strange place to be. It suggests developers are not using these tools because they are confident in the results. They are using them because stopping feels costly, or because their employers expect output volumes that only make sense if AI is in the loop.

Survey data reveals that 43% of developers keep coding with AI after hours despite intending to stop, and 39% report that these tools make achieving mental detachment exceptionally difficult. Some researchers frame this in neurological terms, pointing to the reward loops that form around watching automated code generation succeed or fail. Whether or not the neuroscience holds up, the behavioural pattern it describes is real enough.

There's more where this came from. New articles most weeks.Browse all articles →

The productivity numbers are more contested than they look

Here is the part that tool vendors would prefer you not focus on. A METR randomised controlled trial found that experienced developers were 19% slower with AI tools, despite perceiving themselves to be 20% faster. That gap between felt speed and actual speed is significant. It suggests the tools create a convincing sensation of productivity that does not always match what the clock records.

Vendor-conducted studies from companies like GitHub and Cursor consistently report 50 to 100% productivity gains, but typically use self-selected early adopters and controlled task environments. Independent research shows more nuanced outcomes. This is worth keeping in mind every time a press release cites a headline number. The methodology matters enormously, and companies have an obvious interest in the results looking good.

One of the clearest structural shifts is where developer time is now going. Developers now report spending 11.4 hours per week reviewing AI-generated code versus 9.8 hours writing new code, a reversal of the 2024 pattern. Review overtook writing as the single largest time sink. Respondents who classified themselves as heavy agentic-tool users reported review hours climbing to 14 to 16 per week while writing hours stayed flat or dropped modestly.

That is a meaningful shift in what the job actually involves. Writing code, at least for many people, carries a sense of authorship and problem-solving. Reviewing code someone else wrote, let alone code a machine produced, is a different kind of work entirely. Write-in comments consistently flagged review fatigue as an underreported productivity drag. When the AI produces more code than a developer can meaningfully review, teams either merge under-reviewed work or queue pull requests indefinitely.

The productivity plateau nobody advertises

Perhaps the most practically important finding from the independent survey data is this: self-reported productivity jumps 34% in the first 60 days of using AI coding tools, then flattens, with gains concentrating in specific task types rather than across-the-board velocity. The initial boost is real. The sustained transformation is harder to find.

This pattern makes intuitive sense. AI coding tools are genuinely good at certain things: boilerplate code, standard patterns, writing tests for existing functions, looking up unfamiliar syntax. For that category of work, they save real time. The problem is that companies, having seen the early productivity spike, restructure team sizes and output expectations around it. When gains plateau, developers are left carrying higher workloads with tools that are no longer delivering the early returns.

While 74% of surveyed developers believe heavy reliance on AI tools boosts their promotion odds, 51% warn of severe burnout risks. Those two things are not incompatible. You can believe a behaviour is professionally rewarded and personally unsustainable at the same time. Plenty of industries have demonstrated exactly that.

Who gains and who carries the cost

It is worth being direct about the incentive structure here. The companies selling AI coding tools benefit when developers use them heavily and when that usage is framed as productivity. The companies employing developers benefit when output volumes rise. The developers themselves get a tool that may help them, but also one that is reshaping their working day in ways that were not fully advertised.

Compounding psychological strain is the phenomenon of verification debt, requiring meticulous line-by-line debugging. Developers often spend hours sifting through generated blocks that look correct at a glance. This is not the experience the marketing implies. The experience implied is a confident assistant that accelerates your work. The experience reported is, often, a plausible-looking colleague whose output you cannot quite trust but cannot afford to ignore.

38% of employees have shared confidential company data with unapproved AI systems, a phenomenon known as shadow AI. This happens partly because official, sanctioned tools do not always fit the workflow, so people reach for whatever works. It is a security problem that flows directly from the gap between what AI tools promise and what approved enterprise deployments actually deliver.

The honest version of what these tools are

None of this means AI coding tools are worthless. They are not. The fastest-growing primary tool in early 2026 was not the benchmark leader but the tool that slotted cleanly into existing workflows. That is actually a healthy signal: developers choosing based on fit rather than hype. And there are genuine use cases where the tools save meaningful time, particularly for repetitive or unfamiliar work.

But the 80% dependence figure is a reasonable prompt to ask harder questions. When a tool is used by nearly everyone in a profession, but the majority of users describe their relationship to it as dependency rather than advantage, something about that dynamic deserves examination. Whether it is the tools themselves, the pace at which they were adopted, the expectations layered on top of them by employers, or some combination of all three is not yet clear.

What remains contested is whether AI coding tool adoption is translating into durable engineering value or accelerating technical debt. The honest answer, based on the evidence available right now, is probably both, and the balance likely depends on how carefully individual teams are managing their use.

The productivity story was always going to be more complicated than a press release could contain. We are now at the point where the complications are becoming visible.

From Telltale

Keep reading

If this one was useful, there's plenty more on the site. Pieces on how AI works, plus coverage of AI news, the downsides included. All free to read, no account needed.

See all articles →

References

Published September 2026 · telltale-ai.com
All articles · Privacy · Terms