AI speeds coding but software complexity persists
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
A METR study found experienced developers using AI tools took 19% longer on tasks.
Faster coding shifts bottlenecks to review and maintenance, not finished software.
Background
AI coding assistants are marketed as a way to write software faster, and several recent studies and industry write-ups examined whether that speed translates into better software. METR, an organization that ran a study on early-2025 AI and experienced open-source developer productivity, found the tools slowed developers down even as they believed they were faster. Researchers publishing in the journal AI Perspectives and writers at Boston University's College of Engineering and AWS's Executive in Residence blog reached a similar split verdict: routine work speeds up, but architecture, security and maintenance stay human responsibilities.
How it unfolded
- Jul 10, 2025METR's study of experienced open-source developers found they took 19% longer to complete issues when allowed to use AI tools, despite expecting a 24% speedup and later believing AI had sped them up by 20%.
- Oct 29, 2025An AWS Executive in Residence blog argued that speeding up coding shifts bottlenecks elsewhere, such as product-manager requirement clarification and senior-developer review of AI-generated code.
- Nov 4, 2026Tri-Force Consulting Services wrote that AI coding tools accelerate routine work but do not guarantee faster projects, and Boston University's College of Engineering said AI-generated code should be treated as a draft that may introduce security weaknesses or unnecessary complexity.
- Nov 7, 2026A study in AI Perspectives by Marco Barenkamp, Jonas Rebstadt and Oliver Thomas, reviewing more than sixty peer-reviewed papers plus five practitioner interviews, found machine tools give substantial speedups but remain tethered to human-engineered architectures and lack creative capacity.
- Nov 9, 2026A Medium engineering article described AI-assisted development generating unnecessary complexity before anyone established it was useful, citing a discount calculation service that received an interface, factory, strategy implementations and a generic extension framework for what might need two simple rules.
- Nov 10, 2026An arXiv paper on cognitive load in AI-assisted code comprehension reported the AI-induced cost remained largely constant across code sizes.
Key terms
- METR
- A nonprofit research institute in Berkeley, California that evaluates frontier AI models' capabilities and risks.
- open-source
- Software whose source code is freely available for anyone to use, study, modify, and distribute.
- arXiv
- An open-access online repository of scientific preprints in physics, mathematics, computer science and related fields.
Source: Wikipedia
Who’s saying what
- Caution
- The Medium engineering article argues senior engineers should review generated complexity with the same skepticism as human-written complexity, since customers don't care whether a bug was written manually or generated in eight seconds.
- Analysts
- Researchers Barenkamp, Rebstadt and Thomas found machine tools provide substantial speedups and cost reductions but remain tethered to human-engineered architectures and lack the creative capacity to define original problem frameworks.
- Expert
- TechRadar opinion author Itay Grushka writes that coding is faster, but the winners will be those that create more value from it.
- Official
- The AWS Executive in Residence blog compares focusing only on code to making one machine in a production line 30% faster: without optimizing every process, customers don't receive products sooner.
Still unverified
A Quora answer claimed AI code is cheaper and requires less development time than hand-written software but is 'certainly NOT faster and is almost unmaintainable'; METR states it does not claim its developers or repositories represent a majority or plurality of software development work, and the arXiv paper says future research should test whether its constant cognitive cost holds under greater complexity.