Artificial intelligence (AI) has yet to deliver the sweeping productivity gains in software development that would support predictions of major economic disruption within the next five years, according to Ghanaian technology analyst Bright Simons.
Simons argues that software development offers the clearest early test of AI’s economic impact because the industry is highly digitised, with code that can be generated and tested automatically and AI tools already widely adopted by developers.
“If AI is going to have a crazily disruptive run, we should see it first in software,” Simons said. “It is the economic and organizational sector most amenable to AI.”
His assessment challenges increasingly aggressive forecasts from technology leaders including Elon Musk, who expects AI to surpass humanity’s collective intelligence by 2030 or 2031, and Bill Gates, who has warned of major labour-market disruption within four to five years.
Simons said the evidence emerging from software development instead suggests that AI is increasing the volume of code produced without delivering comparable gains in useful output.
A 2026 working paper tracking more than 500,000 GitHub developers estimated that AI tools generated 240% more saved code changes across successive generations, while releases increased only 30%, according to Simons.
A separate analysis of 718 firms found that adoption of AI agents was associated with 30% more code, but only small and statistically insignificant gains in completed software issues and projects. Review time increased 49%, effectively offsetting part of the production gains.
The quality and reliability indicators are also raising questions about whether higher AI-assisted output translates into better software.
Telemetry covering 22,000 developers showed incidents per pull request increasing 242.7%, while bugs per developer rose 54%, Simons said. An analysis of 623 million code changes found 81% more duplicated blocks, with refactorings accounting for less than 10% of changes, down from 25%.
Developer sentiment also deteriorated as AI adoption increased, with positive sentiment falling from more than 70% to about 60%. Only 3% of developers reported high trust in AI-generated output, according to Simons.
The findings point to a broader problem for companies deploying AI: producing more digital work is not necessarily the same as creating more economic value.
“Google’s DORA research also showed higher delivery throughput alongside lower stability,” Simons said. “It is always the surrounding organisation that determines whether faster production became useful progress.”
That distinction matters for businesses betting on AI to reduce labour costs, accelerate product development and increase productivity. If additional code requires more review, creates more bugs or produces less stable software, some of the apparent efficiency gains can be absorbed elsewhere in the organisation.
Simons said consumer-facing technology provides another test of whether AI has already transformed the quality of digital services.
J.D. Power’s 2026 banking studies, he noted, found virtual assistants struggling with fraud, disputes and problem resolution. Airline website satisfaction increased 3% in 2026, while ratings for mobile-app quality declined sharply.
The growth in AI-assisted software supply has also not necessarily translated into stronger demand. GitHub researchers found roughly twice as many new applications entering major marketplaces, Simons said, without a corresponding increase in measured engagement with those new entrants.
That suggests AI may be lowering barriers to software creation and enabling more developers and companies to launch products, while the market has yet to demonstrate that a larger supply of applications produces proportionately greater user value.
For investors and companies, the distinction between more output and more productivity could become increasingly important as AI spending expands and businesses assess whether automation is generating measurable returns.
“We have had a dress rehearsal of AI at its most intense: the software that shapes the net,” Simons said. “And so far, so underwhelming.”
