Why Developers Are Becoming Dependent on AI Tools — And the Risks That Follow

AI coding assistants have become an essential part of software development in 2026, with new research suggesting that many programmers are now reluctant to work without them.

Earlier this year, AI research organisation METR reported that developers were unwilling to participate in parts of a productivity study because they did not want to complete tasks without AI assistance. The finding emerged as researchers attempted to revisit a well-known 2025 study that compared software development with and without AI tools.

That earlier research produced a surprising result. Although developers believed AI was making them more productive, the data suggested the opposite. While AI accelerated code generation, programmers often spent extra time reviewing outputs, fixing mistakes, directing the AI, and waiting for tasks to finish.

Unable to repeat the original experiment, METR instead conducted a survey in May that relied on self-reported experiences. Unsurprisingly, many technical workers said they believed AI had roughly doubled their productivity and value to their organisations.

However, several recent developments have cast doubt on whether increased AI usage automatically translates into better results. One example is the rise of “tokenmaxxing,” a trend in which the number of AI tokens consumed became an informal measure of productivity.

According to the Financial Times, Amazon recently shut down its internal AI leaderboard, known as Kirorank, after employees reportedly gamed the system by overusing AI agents and driving up costs. The situation highlighted that heavy AI usage does not necessarily lead to greater efficiency.

Uber faced similar questions. The Information reported that the company exhausted its 2026 AI budget within the first four months of the year. Uber COO Andrew Macdonald later said the spending had not produced a measurable increase in projects or overall productivity.

Critics also argue that AI-generated code may create new maintenance challenges. Programmer and author James Shore warned that writing code faster is only beneficial if long-term maintenance costs do not increase at the same pace. Otherwise, organisations may create more software that requires ongoing attention and resources.

Both industry leaders and academic researchers have raised additional concerns. Entelligence AI founder Aiswarya Sankar claimed that companies spend a significant portion of their AI resources fixing bugs introduced by AI-generated code. Meanwhile, code-review platform CodeRabbit reported that its analysis found AI-generated code produced more issues than human-written code.

Although those findings come from companies that sell code-review solutions, independent researchers have reached similar conclusions. A study published by researchers at Singapore Management University warned that AI-generated software can introduce long-term maintenance costs into real-world projects.

Supporters of AI coding agents argue that future AI systems will help address these problems by automatically handling testing, debugging, and maintenance tasks. Cognition CEO Scott Wu, whose company created the coding agent Devin, believes AI can remove much of the repetitive work developers dislike. At the same time, he acknowledges that current AI coding agents still perform more like junior- or mid-level engineers than fully autonomous experts.

Researchers say the solution is not to abandon AI but to use it more effectively. Developers should understand both the strengths and limitations of AI systems, implement strong quality-control processes, and carefully review AI-generated code. They also agree that humans should continue leading critical areas such as software architecture, security planning, and overall engineering strategy.

As AI becomes deeply embedded in software development, the biggest challenge may not be adopting the technology but ensuring that reliance on it does not create new problems in the future.