Why AI Fails at OAuth and How I Got Unstuck

The author recounts their experience using AI to set up OAuth with Supabase, Google, and Expo. While AI excels at coding tasks, it struggled with the complexity and configuration requirements of OAuth. The core problem lies in OAuth's fragmented nature, involving interconnected systems with hidden configurations that the AI cannot access. Tiny mismatches in these configurations, invisible to the AI, lead to persistent "invalid_grant" errors and frustrating debugging. Different OAuth implementations depending on versions and setups further confuse the AI. Runtime issues, such as link handling and memory loss, also contribute to the problems AI struggles to identify. The author initially doubted their abilities, but realized the limitations of AI in solving such a multi-faceted configuration issue. They reframed their approach, focusing on understanding the entire OAuth process rather than solely relying on AI-generated code. This shift led to more informed questions, better understanding, and ultimately, a successful resolution. The author concludes that AI is effective for code-centric problems, but struggles with configuration-heavy tasks where failure points reside outside the code. They learned to recognize and work around these limitations, leveraging AI's strengths while avoiding its pitfalls.