Why AI Models Like Gemini and ChatGPT Can Feel Less Capable Over Time

Developers have increasingly reported that AI models such as Gemini and ChatGPT seem to perform worse on tasks they once handled reliably. Researchers from Stanford and UC Berkeley have documented this as 'model drift,' where retraining a model to fix one issue can unexpectedly degrade performance in unrelated areas. Safety updates and stricter guardrails can also cause 'over-refusal,' making models appear blander and less capable. Additionally, cost-cutting optimizations like quantization and Mixture of Experts routing prioritize speed over depth, potentially affecting output quality. Rising user expectations also play a role, as tasks once considered impressive are now treated as baseline requirements.