The article explores the possibility of generating AI code that is free from errors, or "hallucinations." The author references a July 2024 talk given at the International Mathematical Olympiad, where DeepMind's AI models AlphaProof and AlphaGeometry successfully solved four out of six problems, marking a significant achievement in AI capabilities. The following year, an improved version called Gemini reached gold-medal standards under exam conditions. These models utilize the Lean software language to convert natural language problems into structured mathematical proofs, which can then be verified step-by-step. The main challenge lies in translating this approach from mathematical proofs to programming. Writing code is less structured than creating proofs, complicating the task of defining "correct" code. The author provides the example of a gravity model used to simulate population movements, illustrating how AI-generated code can be evaluated. Five levels of correctness are identified, starting with the basic requirement that the code runs without errors. More advanced checks include verifying code style and formatting, ensuring internal consistency, and validating input/output relationships. The article emphasizes the limitations of these checks. For example, while a code’s structure can be validated, ensuring that it produces sensible outputs is more complex. Users often need prior knowledge of the model’s expected behavior to effectively assess the results. The discussion highlights the broader implications of AI-generated code in various fields, underscoring the need for rigorous verification methods to ensure reliability and accuracy in AI outputs.