Coders Seek AI Assistance – Navigating Potential Risks

By 2026, researchers found that developers have a strong grasp of AI coding tools.

While AI helps programmers write code more efficiently, some researchers warn it may not necessarily improve code quality, which could lead to problems down the line.

In February 2026, the prestigious AI research institute METR disclosed a surprising finding: most developers would not tackle even minor tasks without AI assistance.

METR aimed to update groundbreaking research from early 2025, which had centered on AI coding efficiency. Their study compared the time open-source developers spent on tasks manually versus with AI help.

Even though developers claimed that AI boosted their productivity, they found it surprisingly slowed their progress. Although it generated code swiftly, they spent more time identifying and correcting errors, guiding the AI, and waiting for it to complete tasks.

When METR tried to replicate the experiment to evaluate advancements in both AI and developer skills, they hit a snag.

Developers were disinclined to participate “because they preferred working with AI,” the researchers noted.

Instead, METR conducted a survey in May allowing tech professionals to self-report productivity gains from AI. Unsurprisingly, they believed AI had doubled their value to their companies.

However, recent reports about the high costs associated with tokenmaxxing, along with certain studies, raise questions about these self-assessments.

Tokenmaxxing, a trend in 2026 where productivity is measured by the number of tokens used, is beginning to fade.

Recently, Amazon closed its internal token-monitoring leaderboard, Kirorank, after employees exploited it by excessively employing AI tools, escalating costs, according to the Financial Times. This demonstrated that increased AI usage does not inherently lead to heightened productivity.

Uber reportedly surpassed its 2026 AI budget within just four months, as noted by The Information. COO Andrew Macdonald stated on a podcast that this overspending had not yielded a significant improvement in project results or productivity.

Moreover, AI-generated code might not alleviate ongoing maintenance demands and could even escalate them, as programmer and author James Shore pointed out in a viral blog post on Hacker News.

“Sure, you can write code twice as fast now? Better hope your maintenance costs are halved,” he stated. “Otherwise, you’re in trouble. You’re trading temporary speed for enduring obligation.”

Further evidence suggests that AI can exacerbate code maintenance problems.

A tweet from Aiswarya Sankar, founder and CEO of the reliability engineering startup Entelligence AI, highlighted that companies are dedicating 44% of their tokens to address bugs generated by AI. Additionally, the code-reviewing firm Code Rabbit examined open-source pull requests and found that AI produced 1.7 times more issues than human-written code.

While these metrics may appear self-serving, they are promoted by those marketing AI code-review tools.

Nevertheless, independent researchers have noted similar issues. An April report from Singapore Management University warned that “AI-generated code can introduce long-term maintenance costs into actual software projects.”

Given programmers’ attachment to their AI tools, what might the resolution be?

Advocates for AI coding agents maintain that developers can leverage these tools to manage the monotonous tasks of code correction as swiftly as AI generates outputs. Scott Wu, founder and CEO of Cognition and creator of the AI coding agent Devin, supports this perspective.

However, he acknowledges that while Devin can operate independently, he assesses its capabilities to be between those of a junior and mid-level developer, depending on the tasks at hand. This isn’t a simple “set it and forget it” solution.

The SMU researchers recommend a more human-focused approach. Programmers should deeply understand the tasks AI excels at versus those it does not, as well as their preferred programming languages. They need solid quality assurance systems designed for AI and must carefully review the AI’s output, treating it as if it were produced by a junior developer.

Additionally, the researchers, along with Wu, suggest that humans should concentrate on overarching responsibilities like software architecture and security design.