A new prototype chatbot lets you rewrite what you already said, instead of just bolting on corrections.
Researchers at QxLab Ireland built ReChat, a chat interface where prior turns are editable rather than fixed. Today's chatbots treat a conversation as an append-only log: correct yourself and the outdated request just sits there, still shaping every answer that follows. The paper calls this context pollution. In a controlled study with 17 participants, people rated ReChat's editable transcripts higher than standard chat on clarity, confidence, and ease of use, and were less likely to want to restart the conversation entirely. A follow-up analysis of the study's transcripts found that editing also shortened conversations and cleared out stale context that a normal chat log would have kept feeding the model.
The idea targets a real annoyance: anyone who has corrected a chatbot mid-conversation knows the model often half-remembers the original, wrong framing anyway. Treating conversation history as a mutable state, closer to a document you revise than a transcript you append to, is a small but sensible reframing of how chat UX should work. It's also a reminder that most prompt-engineering pain isn't about clever phrasing, it's about state management.
Seventeen participants and a working prototype is early evidence, not a product. Whether mainstream chat products ever adopt editable history, and whether letting users rewrite the record creates its own problems, like quietly erasing evidence of a model's earlier mistakes, is the more interesting question this paper doesn't answer.