Skip to main content

The Hidden Costs of AI Coding, Revisited

Faster code doesn’t mean better software. MIT IDE Digital Fellow and Dartmouth Professor Geoffrey Parker on what companies should do to get lasting value from AI coding.

November 17, 2026

lines of code on a computer screen
  • A year after warning that AI coding’s early gains may not hold up inside complex legacy systems, Dartmouth College Professor of Engineering Innovation and MIT IDE Digital Fellow Geoffrey Parker says AI coding clearly has benefits, but the risks are still real.
  • Technical debt and hidden dependencies can quickly pile up when companies move fast, especially when junior developers use AI without the experience to spot the complexities in legacy code.
  • Companies should view AI coding as part of a wider AI digital transformation, Parker says, and it won’t deliver results without the right infrastructure and trained people behind it.

Whether you’re at an agile start up or a legacy tech company, AI coding tools are now standard issue. According to Stack Overflow’s 2025 Developer Survey, 84% of respondents said they use or plan to use AI in development, and 51% of professional developers use it daily. But researchers who examined how AI was being integrated into coding tasks found that speed has come with costs that many companies are only now starting to count.

MIT Initiative on the Digital Economy Digital Fellow and Dartmouth College Professor Geoffrey Parker, as well as Edward Anderson and Burcu Tan, cautioned in the August 2025 issue of MIT Sloan Management Review that the benefits of AI coding also come with risk.

The reported early productivity gains of AI coding came from controlled studies of isolated tasks. In real companies, where new code sits on top of complex existing systems, the risks of using AI to write code are greater and harder to manage.

In a recent interview with Parker, we asked if the advancement of AI over the past year has changed his assessment, especially as AI coding becomes part of everyday work. The co-author of Platform Revolution, the definitive book on digital platforms, noted that the risks of AI coding still exist, and companies have acknowledged this challenge.

“AI implementation is another digital transformation technology,” said Parker. “It will not give you the benefits you’re looking for if it doesn’t match the right infrastructure.”

Are the findings on AI coding in 2025 still relevant in 2026?

The 2025 research acknowledged that using AI for coding tasks was a huge advancement. But the researchers also argued that faster code doesn’t automatically mean better software. Companies must also be aware of the hidden costs that surface as companies scale beyond pilots.

Their assessment is based on interviews with people across the software development chain, from junior developers to lead engineers and CIOs, in industries including insurance, web hosting, social media, defense, management consulting, and fintech. They paired those interviews with a review of the trade press and their own economic modeling.

What they found was that technical debt is eroding some of the productivity advancements of AI. Technical debt, the short-hand term for the cost of additional technical work needed to address shortcuts and quick fixes made during development, can have lasting implications at an organization. Using AI for coding has shown to increase output, speed and productivity, but when AI code is built upon legacy systems, it can lead to setbacks later, they argue.

Parker said that technical debt is a common and persistent problem at established organizations because it’s hard to get leadership excited about fixing it. Now that AI use is widespread, and in the hands of less experienced developers, that problem could compound.

“If junior developers are working in a legacy environment, AI may go badly because they’re creating dependencies they can’t see,” said Parker. “It’s adding complexities your junior person doesn’t know how to look out for.”

Some of the largest tech companies seem to agree. In 2026, Amazon began requiring senior engineers to sign off on AI-assisted production changes made by junior and mid-level staff, and Apple blocked App Store updates for some AI “vibe coding” apps.

“There is some recognition this problem is still salient and real,” Parker said.

More AI coding but less confidence in the results

In the year since that article, AI coding has gone mainstream. It moved from early adopters into the daily routine of developers at companies of every size. JetBrains’ 2026 Developer Ecosystem Survey of more than 15,000 professional developers found that 90% use AI coding agents at work at least weekly, and 68% use them daily. Among 127 mostly small and midsize software firms surveyed by Techreviewer, 89% use AI to write or assist with code, and the median firm says AI touches 26% to 50% of its code.

Confidence hasn’t kept pace, and the risks Parker and his colleagues flagged in 2025 are now showing up in the data. In Techreviewer’s survey, a third of firms had found security problems in AI-generated code, and 37% reported over-reliance on AI or declining developer skills, especially among junior developers. SmartBear’s 2026 State of Software Quality and Testing report, which surveyed 1,436 U.S. and U.K. leaders and practitioners, found that 46% of teams have shipped AI code that later failed in production.

Even so, the people furthest from the code remain the most confident: 73% of leaders have a lot or complete confidence that AI code works as intended, compared with 52% of practitioners.

AI agents accelerating workflows

Developers are not only using large language model tools in the coding process. In the past year, the use of AI agents has grown as well.

“The agentic flow is making hours and days of work possible without you intervening,” said Parker.

A year ago, a developer might ask a chatbot for a snippet of Python. Today, agentic AI can carry out complex, well-specified workflows that run for hours or days without anyone stepping in. In an experienced professional’s hands, that’s powerful. In a newcomer’s, Parker said, he wouldn’t trust it for an enterprise implementation.

Parker said companies must train workers on the risk of building hidden dependencies and growing more technical debt when writing code with AI. These training programs include upskilling junior developers and creating mentorship programs with senior developers.

AI made the individual contributor a manager

For Parker, the lesson of the past year is that AI coding isn’t just a tooling choice, a workflow tweak, or a technical debt problem. It’s a digital transformation, and like past transformations, it will redefine roles and processes rather than simply speed up existing ones. Technology won’t deliver the benefits companies are looking for, he said, if it doesn’t match the right infrastructure.

The effect on roles is already visible. Parker pointed to Blue Health Intelligence, a consortium-owned analytics company that serves individual member organizations for which building their own analytics products makes little sense. In a visit to Parker’s classroom, the company provided a case study of its analytics agent. After spending a year building it, the agent was about as good as a first-year analyst. The agent still makes mistakes and needs senior review. But when Medicaid and Medicare codes changed, analysts needed two weeks to understand the implications. The agent took 40 minutes.

That entry-level analyst job effectively no longer exists, Parker said, and the agent will only get better. Software teams are seeing the same pattern. Work that once went to entry-level staff now goes to AI, and workers are taking a more managerial role, overseeing work completed by AI. The larger question is how quickly companies reshape work around these tools, and whether they use them mainly to cut costs rather than raise productivity.

Taking AI beyond coding

These issues reach well beyond IT, software development and coding. Siloed data limits the extent to which companies can reap gains from AI in all industries, from finance to operations to marketing.

Parker continues to examine these questions through his research on technical debt and AI’s influence on platforms and logistics, and he and his coauthors are currently writing a follow-up to Platform Revolution, titled Platform Domination.  The new book covers many of the ways the platform economy has evolved over the past decade, the power that the major players are able to exert, and how AI is changing the platform landscape.