Developers in a Dilemma: Unable to Code Without AI

By 2026, many software developers find themselves unable to code without AI assistants. However, a series of research studies warn that this dependency may lead to long-term negative consequences. A shocking study published in February by the U.S.-based research institute METR revealed that the majority of developers admitted they “could not work without AI.” When METR attempted to replicate a pioneering experiment conducted in 2025, participants refused to work without AI support. This phenomenon, termed “token-maxing,” highlights the disconnect between excessive AI consumption and actual productivity gains.

Developers Who Can’t Work Without AI In its

2025 experiment, METR’s research team measured the time it took open-source developers to complete the same tasks manually versus using AI assistance. Although participants felt that AI increased their productivity, the study found that time spent correcting errors and adjusting prompts often made AI-assisted work slower than manual tasks. When METR attempted a follow-up experiment in 2026, researchers encountered an unexpected hurdle: developers outright refused to participate unless AI assistance was provided. Unable to execute the experiment as planned, METR conducted a self-reported productivity survey instead, which revealed that developers perceived their value to have doubled thanks to AI. However, researchers remain skeptical of these self-assessments, as there is often a significant gap between perceived and actual performance outcomes.

The Rise and Fall of “Token-Maxing”

By 2026, the software industry faced a widespread phenomenon known as “token-maxing.” This refers to the practice of measuring productivity by the number of tokens (processing units) consumed by AI. Developers would frequently engage AI tools to generate vast amounts of code, creating the appearance of efficiency and productivity. But this trend is now on the decline. According to the UK-based Financial Times, Amazon recently shut down its internal token tracking leaderboard, “Kirorank,” after employees manipulated their rankings by overusing AI agents, inflating costs in the process. Amazon came to realize that increased AI usage does not automatically translate to higher productivity. Similarly, ridesharing giant Uber faced severe challenges. As reported by The Information, Uber exhausted its entire AI budget for 2026 within the first four months. COO Andrew Macdonald admitted during a podcast that this expenditure did not lead to measurable improvements in projects or productivity. The situation has unveiled a troubling pattern where AI-related costs squeeze profits without delivering tangible benefits.

Maintenance Debt from AI-Generated Code James

Shore, a programmer and author, sounded the alarm on his blog, stating, “Even if AI enables you to write code twice as fast, if you can’t halve the maintenance cost, you’re essentially signing a perpetual servitude contract.” Shore’s post gained significant attention on Hacker News. AI-generated code tends to contain more bugs than human-written code, leading to increased maintenance costs over time. Aiswarya Sankar, CEO of the startup Entelligence AI, tweeted that companies spend 44% of their AI tokens on fixing bugs in AI-generated code, sparking widespread discussion. Code Rabbit, a provider of code review tools, also presented data confirming the quality issues of AI-generated code. Their analysis showed that AI-generated code requires more revisions during reviews compared to traditionally written code.

The Illusion of Productivity Gains

Researchers acknowledge that AI can speed up developers’ workflow, but only during the code generation phase. Subsequent stages—debugging, testing, code review, and documentation—often require more time, potentially offsetting the initial gains. Additionally, over-reliance on AI tools raises concerns about declining fundamental programming skills. If problem-solving abilities and algorithm design thinking deteriorate, this could lead to a long-term decline in the technical capabilities of entire organizations.

Striking the Right Balance in AI Utilization

Industry experts emphasize the importance of using AI strategically rather than eliminating it entirely. For instance, AI can be employed for routine tasks like automated code generation and documentation, while human judgment should be prioritized for architectural design and complex business logic implementation. Rather than measuring productivity by token consumption, organizations should focus on actual outputs such as the number of releases, bug rates, and customer satisfaction. The examples of Amazon and Uber highlight the risks of setting metrics that employees can exploit, leading to inflated costs without meaningful results. While AI coding tools will continue to evolve and hold potential for genuine productivity improvements, it is crucial to recognize the risks of excessive dependency. Combining human decision-making with strategic AI utilization is essential for sustainable progress.