You’ve been shipping code faster than ever. Copilot is writing your boilerplate, ChatGPT is fixing your bugs, and your sprint velocity is through the roof. You feel like a 10x developer. But you’re actually building a bomb.
Every line of code an AI writes today is a withdrawal from an account we don’t know how to repay.
We’ve convinced ourselves that AI is reducing our technical debt. The narrative is seductive: AI handles the tedious refactoring, writes the tests, and clears the backlog. It feels like negative-interest tech debt—debt that actually pays you back over time. But this completely ignores the dark, exponential reality of how these models actually learn.
Think about it. Where does the next generation of AI get its training data? From the internet. From GitHub. From public repositories. And what are those repositories rapidly filling up with right now? AI-generated code.
We aren’t training the next generation of AI on human ingenuity; we’re training it on its own regurgitated echoes.
This is the phenomenon of model collapse. When AI models are trained on data generated by previous AI models, the quality degrades. The variance drops. The weird, brilliant, highly specific edge cases of human code get smoothed out into generic, mediocre mush. The optimists insist that AI will just keep getting smarter. The realists know that without a fresh supply of human signal, it might just get dumber with every iteration.
The immediate promise of AI reducing developer burden is clashing violently with the long-term risk of self-reinforcing code degeneration. We are inadvertently poisoning the data source for tomorrow’s AI. It’s an invisible, compounding problem. You can’t see it in your IDE. You can’t measure it in your CI/CD pipeline. But it’s there, growing quietly in the background.
Speed isn’t free. We are trading the long-term intelligence of our tools for a short-term spike in our sprint velocity.
If you are a developer, an engineering leader, or an AI tool adopter, you need to rethink how you integrate these tools. Don’t let AI write your core business logic unchecked. Don’t blindly merge its pull requests. Keep the human signal strong. If we don’t, the quiet dread of this invisible debt will become a deafening crash.
The ultimate irony of AI coding tools is that they might be the very thing that stops AI from ever reaching true intelligence.
FAQ
Q: Won't AI companies just filter out AI-generated code during training?
A: They'll try, but it's practically impossible to distinguish good AI code from human code at scale. The contamination is already happening, and the synthetic data is diluting the human signal.
Q: So what, should I stop using Copilot entirely?
A: No, but be ruthless. Use AI for boilerplate and repetitive tasks, not core architecture. Treat its output as untrusted debt that requires intense human review, not a finished product.
Q: Is this just an excuse for bad developers to write bad code?
A: No, it's a warning to leadership. The developers aren't the problem; the system is. We are incentivizing speed over sustainability, and the AI is just the accelerant burning down the foundation.