The Self-Review Paradox: Why AI Models Rationalize Their Own Mistakes

When AI models review their own code, they apply motivated reasoning rather than objective analysis—defending their initial decisions instead of critically examining them. This essay demonstrates how Claude both generated and then approved a subtle bug (accepting empty strings when specs required non-empty validation). The proposed solution is 'model racing': running multiple AI models in parallel on identical prompts and comparing results, which provides independent perspectives that expose blind spots inherent to self-review.