AI Models Drop 83% of User Instructions, Study Finds

Ask ChatGPT to remember something for the rest of a conversation, and there's a decent chance it'll forget within a few exchanges. Researchers at Penn State just measured how bad the problem is, and the number is stark: only 17 percent of user instructions survive the compression process AI systems use to manage long conversations.

The mechanism behind this is called "compaction." When a chat runs long enough, the model's context window fills up, so the system summarizes everything that's happened so far and discards the rest to make room for new input. It's a practical fix for a real constraint, and it works fine for preserving the broad strokes of a task. What it doesn't do well is hang onto the specific rules a user attached along the way, things like "check with me before you change anything" or "don't use my name." Those instructions aren't part of the core task and aren't baked into the system's permanent settings, so when compression runs, they're the first casualty. The Penn State team set out to document exactly which kinds of instructions get lost, and how consistently it happens.