Microsoft explains why AI loses track in long conversations

The team’s research, including the papers “LLMs Get Lost In Multi‑Turn Conversation” (May 2025) and “LLMs Corrupt Your Documents When You Delegate” (April 2026), shows that performance drops sharply when tasks are split across several exchanges. Small deviations accumulate silently, so a document may remain grammatically correct while key meanings drift. The researchers stress that what feels easy for humans can be hard for AI, and vice‑versa, because transformers rely on pattern matching rather than true understanding.

To mitigate these issues, they recommend pausing a confused session, summarizing all objectives, constraints, and prohibitions, and restarting the conversation with that summary. Adding human‑verified checkpoints—mid‑task summaries, difference checks, and final human approval—helps keep the output aligned with the original intent.