AI’s complexity cliff: Stack Overflow CEO on trust in AI, productivity, and the skills disruption

“I think we will see a lot of job displacement in the next two years,” says Prashanth Chandrasekar

As AI tooling turns the software industry upside down, transforming everything from code generation to enterprise workflows, Stack Overflow sits at a key vantage point. Computingasked CEO Prashanth Chandrasekar what its data reveals about AI adoption and for some 2026 predictions.

Stack Overflow, the developer’s source of peer-reviewed technical knowledge for more than 15 years, is now both a licenced data provider for large language models (LLMs) and a useful barometer of developer sentiment.

The Stack Overflow Developer Survey is a definitive report of software development. In its 15th year, the survey received almost 50,000 responses from all over the world. Its results were announced in the summer, but Chandrasekar draws attention to some findings that should cause concern for AI vendors, and enterprises trying to leverage it to boost productivity.

“Generally speaking, people are leveraging AI tools a lot more,” Chandrasekar explains.” 84% of developers in our survey say they’re using some sort of AI tool.”

What should be cause for concern among the makers of such AI tooling is how few developers trust them.

“Only about 29% said they trust the outputs of AI tools. That is down from over 40% in prior years. That’s a meaningful drop-off.”

Stack Overflow’s enterprise customers who run a private, secure version of the platform report the same patterns as the developer community more widely. “When we talk to them, they say people are spending a lot of time trying to debug or unravel what you might call extreme AI slop. Simple things are fine. But the moment the task is complex, it goes around in loops, loses context, or just isn’t reliable.”

45% of respondents told the survey that debugging AI-generated code was frustrating and time-consuming, which is a notable finding given how many software teams are under pressure to adopt AI rapidly.

Security and privacy intensify the hesitation in enterprise IT. The question of trust in output is parallelled by questions of whether companies can trust AI development tools with their data.

“Data residency, privacy, security are very real concerns in enterprise environments,” Chandrasekar says.

Trust is integral to the tricky task of balancing the enterprise arm of Stack Overflow’s business with its community values. The company has faced criticism, not least from the developer community, over the deals it has signed with OpenAI, Google, Microsoft and Amazon. The question of Stack Overflow content being effectively sold without the consent of users has led to disgruntlement, to put it mildly.

Chandrasekar argues these deals ultimately help the developer community because its data can make the tools better.

“We ran our own analysis across all the major models. When you ground these systems in or data which is high quality and human-curated data, the hallucination rate drops dramatically. Accuracy goes up by forty-plus percent,” the CEO says. “The models can be far better. The problem is that most people aren’t interacting with grounded systems day-to-day.”

“We took our data looked at all the LLMs, and we published how accurately these LLMs actually perform relative to the answers on our site, and that is available on Stack Overflow.”

Vibe Coding: Useful for prototypes

Vibe coding has been everywhere this year, even managing to become Collins Dictionary ‘s Word of the Year. Is building interfaces or applications using pure natural language prompts really a useful enterprise tool?

“For simple things, it’s very good,” says Chandrasekar, “We use it internally to prototype. Think of it as a first draft. It gets the conversation going quickly and you don’t have to be a trained programmer to do that. But it runs into the same complexity cliff as AI tooling more generally.

“It may become more mainstream over time, but today? 77% of developers told us vibe coding is not part of their professional workflow. It’s used more by designers, product people, or founders who want to sketch an idea without waiting for engineering support.”

Chandrasekar is reluctant to predict the direction of travel with what could be described as the deskilling of development because he sees both the positive potential for democratisation but also risk.

“This opens the door for people without computer science degrees to build prototypes or start software companies and that’s great. It will dramatically increase the number of people creating software.

“But for anything that needs to scale, you still need real engineering fundamentals. You wouldn’t want to rely on a team that doesn’t know what they’ve written.”

Job displacement

Asked about the likely disruption to the developer workforce in the years ahead and Chandrasekar is blunt.

“I think we’re going to see a lot of job displacement over the next 24 months because three things are happening.

“Firstly, the massive capital investment that we’ve seen. Banks, global enterprises, every big tech company have all invested huge amounts of money in AI. CFOs want to see a return and productivity gains.

“Secondly, companies plan years head. They’re going to look at projected AI improvements and say, ‘Why do we need 300 developers in this group?’ They will reduce future headcount.

“Finally,” he continues, “you have the expectation of productivity gains. If your senior developers can operate at 1.5x productivity with AI, the question becomes: do you still need the same volume of junior developers?”

The entry-level job market isn’t going to get any easier anytime soon, but of course, this will create challenges in a few years’ time when all the mid-level career talent simply won’t exist.

“Junior work is the easiest for AI to handle. But that creates a structural problem. If companies stop hiring juniors, you break the pipeline. The mid-level engineers of tomorrow have to come from somewhere.”

Chandrasekar believes the market will eventually rebalance (although demurs from explaining how) but not before a period of turbulence and an explosion in freelance and startup activity.

“It will be easier than ever to start a tech company. Displaced workers will go independent or build new ventures.”

Scaling AI in the face of resistance

I ask Chandrasekar his thoughts on the MIT study earlier this year, which was the first widely researched piece of evidence that enterprises are not seeing the value they expect from AI pilots and that 95% fail. He acknowledges that enterprise adoption remains slower and messier than public hype leads us to believe.

“When AI pilots succeed, they often do so because they’re tested on self-selecting, pioneering teams. Those groups come back with 20%–30% productivity gains. But when leaders try to roll that out to the whole organisation it drops to single digits.”

Cultural and emotional resistance is a huge factor to enterprise finding the value it seeks from AI, which isn’t terribly surprising given that employees from finance, legal, HR and multiple other lines of business in addition to software development, have been deluged with content for the last two years telling them AI is coming for their job, and their children’s future.

“Telling employees to use AI isn’t a tooling decision—it’s a change-management journey. People fear replacement. They worry about their skills becoming obsolete. That human resistance is one of the biggest reasons this isn’t mainstream yet.”

The CEO describes developer sentiment as a spectrum. “Some are resistant. Some are enthusiastic. But the pragmatic middle is growing. They want the best of both worlds: AI-powered interfaces grounded in human-verified technical knowledge.”

A bubble, but not necessarily a bad one

With tech giants investing billions and sovereign wealth funds pouring oil-reserve money into AI startups, does Chandrasekar think this is sustainable? The short answer is no, but he doesn’t think it’s necessarily a bad thing.

“We’re probably in a bubble but bubbles aren’t purely negative. They create ecosystems. Big companies fund startups, startups buy their chips, and innovation cycles accelerate.

“There’s going to be a lot of companies that are going to be hurt by it and probably won’t make it as in all bubbles.”

The real constraint, he argues, won’t be capital, it will be energy infrastructure.

“The grid is going to be absolutely hammered. You won’t have enough power for every new data centre. At some point, there’s going to be a reckoning.”

Ultimately, he believes AI will reshape the industry, but the transition will be uneven, and leaders must prepare for both capability leaps and structural gaps.

“The direction of travel is clear but the complexity cliff is real. Trust is still too low. And the talent pipeline may suffer before it strengthens. It’s going to take thoughtful management and real evidence to get from where we are to where we want to be.”