According to Decrypt, physicists at George Washington University have developed a formula intended to estimate when an AI chatbot will transition from providing good answers to bad ones. The researchers tested the formula on small models and found that early results support its predictive ability.
The approach aims to identify the point at which performance degrades in chatbot responses. It is based on observable characteristics during model operation, though specifics of the underlying mechanism are not detailed in the summary. The test phase involved small models only, and no timeline or threshold for “bad” answers was provided.
The finding reflects ongoing interest in monitoring and anticipating reliability shifts in AI systems. Further validation would require additional testing under varied conditions, which has not yet been reported.
