Coders say that AI-generated code often contains errors, and checking and fixing it can take more time than writing the code from scratch.

“We are being forced to use AI agents to make large-scale changes across the entire codebase. It is simply impossible to assess the quality and security of that volume, especially given that hundreds of programmers are doing the same thing,” said a UX designer at a technology company.

The expert added that the team is accumulating “a mountain of technical debt” that will be impossible to untangle once models become prohibitively expensive.

Tech company executives have been actively reporting the share of code generated by AI:

  • in April, Google said the figure was 75%;
  • in 2025, Microsoft CEO Satya Nadella cited 30%;
  • at Anthropic, the figure was 90%;
  • Meta CEO Mark Zuckerberg predicted that within 12–18 months, AI systems would write most of the code that improves AI itself.

Against this backdrop, mass layoffs continue across the industry, with companies explaining them as a result of automation and cost optimization.

The productivity boost did not happen

According to 404 Media, the “huge productivity leap” supposedly delivered by artificial intelligence has not translated into an increase in either the quantity or quality of products.

Developers deny that AI is genuinely useful in their work, but they are still required to integrate it into their workflow.

“Using LLMs in one form or another is a mandatory requirement. Their use is part of performance evaluation criteria. We are literally flooded with AI tools, and the answer to any problem is to ‘try AI first,’” said a programmer at one FAANG company.

Because performance assessments are tied to the adoption of the technology, most developers use AI “for show.”

An engineer at a fintech company said that using LLMs is not mandatory at their workplace, but it is encouraged: developers are given access to Cursor.

A programmer at a small web design firm stressed that AI assistants inside IDEs have not increased productivity. The code contains errors, and every line has to be double-checked.

“Another developer works with me on a contract basis. He generates huge amounts of code, leaving me with pull requests of more than 1,000 lines to review, and that takes an enormous amount of time. As a result, I feel more tired and burned out than ever before in my life,” the engineer said.

A coder from the fintech sector added that AI can generate more code than a team has time to review or explain.

“As a result, you either throw it away or ship it while fearing that it may contain very low-quality elements,” he explained.

AI is useful — sometimes

Developers acknowledge that AI does handle some tasks well, for example helping quickly assemble prototypes and implement solutions in unfamiliar areas.

One engineer said that LLMs are useful when working with large volumes of information: they can find where a request is handled on a server, summarize log data, and help search documentation related to code changes.

Skill deterioration

As AI becomes more deeply embedded in the workflow, developers are losing skills they have spent years building. Researchers call this phenomenon “cognitive debt” or “cognitive atrophy.”

A programmer at a small web design company said that at one point he could not remember how to implement an API in Laravel — and that it “scared him to death.”

“It is like when mobile phones appeared and we stopped remembering phone numbers. For me, it turned into outsourcing the thinking process. My critical thinking and my ability to sit down and think through a problem or a project have deteriorated,” said a software developer in the financial sector.

AI is here to stay

Most engineers agree that large language models are here to stay and will continue to play a role in programming. The real question is how the industry will deal with management’s current obsession with the technology, especially when it comes to training the next generation of developers.

“We are hiring junior programmers who rely on AI to complete the simplest tasks. They do not have the knowledge or experience to understand when neural network outputs contain errors or are inefficient,” the UX designer said.

AI coding tools can help with prototypes, documentation search, and unfamiliar tasks, but forced adoption risks creating more review work, technical debt, and skill erosion. For engineering teams, the key issue is not whether to use AI, but how to keep human code review, architecture judgment, and junior developer training from weakening.