48-Hour AI Stress Test Reveals How Free Models Drift Over Repeated Tasks

A developer ran the same support-ticket classification task against a free AI model every hour for 48 hours to test behavioral consistency rather than raw accuracy. The experiment used ten support tickets and three labels, with each ticket appearing roughly twelve times across the test period. Around the 22-hour mark, the model began misclassifying a ticket and reinforcing its own errors because the prompt fed it recent outputs as memory, causing it to trust stale context over the actual input. Response hashing proved critical, flagging format changes and anomalies that a lenient label parser had silently masked as successful runs. The author concluded that long-running automation depends more on output stability and honest logging than on benchmark performance scores.