AI code generators produce working code that wastes energy — researchers test whether they can be retrained to care about efficiency

Researchers tested whether large language models that write code can be taught to prioritize energy efficiency alongside correctness, using a cheaper retraining method that doesn't require rebuilding the entire model. The method showed mixed results: it improved code accuracy for some models but energy savings were inconsistent, suggesting that teaching AI to write efficient code isn't automatic even when you try.

As AI-generated code moves from research projects into production systems, the energy cost of that code matters at scale. Right now, AI writes functionally correct code that runs slowly and wastes power — which undermines corporate sustainability targets and increases data center costs. This paper shows that the problem is harder to fix than expected: you can't just point the model at efficient code examples and expect it to learn. The inconsistency suggests energy efficiency in code generation might require fundamentally different training approaches, not just better prompting.