In 2026, developers’ reliance on AI coding tools remains unwavering, according to new research.
Although these tools are undeniably speeding up coding processes, some experts caution that they may not be enhancing the quality of code, potentially leading to significant issues later on.
In a striking update released in February 2026, the prominent AI research organization METR announced that the majority of developers are now unwilling to perform even basic tasks without AI assistance.
METR had initially sought to build on prior research from 2025 analyzing AI’s impact on coding productivity. That study compared task completion times for open-source developers doing work manually versus using AI tools.
Despite claims of improved productivity, developers discovered that AI often hindered their progress. While code generation was quicker, they found themselves spending additional time addressing mistakes, guiding the AI, and awaiting task completions.
Upon attempting to replicate their earlier experiment to assess advancements in AI and coder skills, METR faced an unexpected challenge.
Developers declined to take part, citing a refusal to work without AI even for the study, the researchers admitted.
In May, METR opted for a different approach, releasing a survey enabling technical staff to self-report their perceptions of AI’s influence on productivity. Predictably, many felt that AI had doubled their worth to their companies.
However, recent concerns regarding the costs associated with “tokenmaxxing” and emerging research have cast doubt on these self-assessments.
Tokenmaxxing, referring to the practice of using token counts as a measure of productivity with AI, has been prevalent in 2026, but its efficacy might already be in question.
Recently, Amazon discontinued its internal token-tracking system, Kirorank, after discovering employees manipulated the system through excessive AI usage, leading to soaring costs, as reported by the Financial Times. This incident illustrated that merely using AI does not guarantee enhanced productivity.
Uber’s experience mirrored this trend, having exhausted its 2026 AI budget within just four months, according to The Information. COO Andrew Macdonald noted on a podcast that this expenditure did not translate into measurable improvements in project outcomes or productivity.
The idea that AI-generated code reduced ongoing maintenance needs has also been challenged. In a widely circulated blog, programmer and author James Shore argued that AI might actually create greater maintenance demands.
“If you write code at double the speed, can you also halve your maintenance costs?” he cautioned. “If not, you might be better off in the long haul.”
Additional reports suggest that reliance on AI could complicate code maintenance further.
A tweet from Aiswarya Sankar, founder and CEO of Entelligence AI, revealed that companies are allocating 44% of their tokens to address bugs caused by AI-generated code. Similarly, an analysis by Code Rabbit showed AI-generated code produced problems at a rate 1.7 times higher than human-generated code.
These statistics may seem self-serving from those promoting AI tools.
Yet independent studies echo these concerns. Recently, researchers from Singapore Management University published a report noting that “AI-generated code can accumulate long-term maintenance liabilities in real software projects.”
Given that developers increasingly appreciate their AI tools, what is the way forward?
Proponents of AI coding tools assert that developers should leverage AI for complex tasks like bug fixes, as proposed by Cognition founder Scott Wu.
However, Wu acknowledges that while his AI agent, Devin, can operate autonomously, its skills range from junior to mid-level, depending on the task. This solution is not without its oversight.
Researchers from SMU advocate for a more balanced strategy. Programmers should grasp the strengths and weaknesses of AI as thoroughly as they understand their preferred coding languages. Effective quality assurance designed for AI is critical, and programmers must diligently review AI outputs akin to working with less-experienced team members.
Additionally, these researchers, along with Wu, emphasize that humans must maintain responsibility for overarching areas such as software architecture and security design.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.
#Coders #refusing #work #AIand #bite