Programmers already rely on artificial intelligence to help them write code faster. Tools like GitHub Copilot or models like GPT-4 promise miracles, but the reality is more complicated. According to an article in IEEE Spectrum, these AI assistants are actually getting worse. The reason? They are trained on data they generate themselves, leading to so-called model collapse. This means that output quality declines because AI learns from the errors and simplifications it produces itself.
Research shows that when models are trained on synthetic content—that is, text and code generated by other AI systems—they lose the ability to handle complex or unusual tasks. For example, in tests involving databases for financial systems, which use simple entries such as debit or credit, AI fails in terms of security. Anjan Saha, who commented on the issue, emphasizes that for complex systems such as power grids or medical applications, we need data in multidimensional matrices, which AI systems like ChatGPT or GitHub Copilot cannot reliably handle. Instead, they focus on common tasks but fail when faced with unusual problems.
Bias Toward Popular Frameworks
Another major problem is bias toward popular technologies. AI coding tools favor popular languages and frameworks, such as Python or React, because they have the most training data for them. This creates a popularity paradox: lesser-known frameworks, such as those used for specific engineering applications, remain overlooked. As a result, programmers working with specialized tools receive poor-quality suggestions, causing frustration.
This bias causes AI to ignore edge cases. If you need code for an unusual scenario, such as integrating a legacy system with a new one, AI often suggests a solution that does not work. According to analyses from 2025 to 2026, such as one by Faros AI that examined 10,000 developers, this manifests itself in a productivity paradox: AI helps with routine tasks but complicates difficult ones that require human expertise.
Decline in Data Quality
The decline in high-quality data also has a major impact. Sites like Stack Overflow are experiencing a drop in questions—by as much as 84%—because AI solves simple problems. Only the difficult ones remain, and they go unanswered. This means that new models are trained on less diverse data, increasing the risk of collapse. Synthetic content is flooding the internet, causing AI to learn from its own mistakes and reducing its reliability.
For example, in tests with CSV data for banking schemas, AI generates data that looks good but conceals security vulnerabilities. Anjan Saha points out that systems such as social networks or power grids need to process complex parameters in Oracle databases, which AI cannot do without human oversight. As a result, developers spend more time making corrections than they should.
Impact on Developer Productivity
Although 84% of developers, according to Stack Overflow surveys, plan to use or already use AI tools, many report frustration. Tools like Claude Code or Cursor may help with small tasks, but they fail on large projects. For example, Replit's Agent was able to build an entire application, including deployment, in 2025, but earlier versions were chaotic and required extensive intervention. This leads to a situation where AI increases speed at the beginning but causes problems later on.
Research warns of model collapse, in which training on AI-generated code reduces performance on new problems. For non-experts, this means that even though AI appears intelligent, its errors accumulate and can lead to serious failures in applications such as financial systems or medical devices.
Overall, these problems suggest that AI coding assistants need better training data and human oversight to prevent them from deteriorating further.
