As AI chats drag on, does the bot's behavior shift because it's less capable, or because long conversations just trip it up?

This explores why chatbots act differently as a conversation gets longer, and whether that comes down to raw model power or to something else. The corpus doesn't contain a study that directly separates length from capability. Its notes do point the same way, though: the effects of length mostly come from how models are trained and how relationships play out over time, not from model size. The clearest number comes from Why do AI assistants get worse at longer conversations?. Models that score about 90% when an instruction arrives in one message fall to about 65% when the same information comes in gradually over a natural conversation. The same models are involved in both cases, so the drop isn't a capability gap. When information arrives bit by bit, models commit to early guesses and can't back out of them.

The explanation the corpus offers is the training signal. Why do language models respond passively instead of asking clarifying questions? shows that standard RLHF rewards the response that seems most helpful right now. That teaches models to answer straight away instead of asking a clarifying question that would pay off three turns later. A bigger model trained on the same reward inherits the same short-sighted habit. Two notes describe what is missing. Could proactive dialogue make conversations dramatically more efficient? finds that offering relevant information before being asked can cut conversations by up to 60%, yet this behavior hardly appears in AI training data. When should AI agents ask users instead of just searching? borrows a term from linguistics: 'insert-expansions' are the short side-questions people use to check what someone means before going on. It shows that tool-using agents drift when they skip them. Linguists who study conversation describe a further layer in Why don't language models develop conversation maintenance skills?: fixing misunderstandings and handing topics back and forth is social work, not information. Models trained to predict information never learn it, however large they get.

A second, less obvious pattern: length also changes the human side. Do chatbot relationships lose their appeal as novelty wears off? tracked people talking with the Mitsuku chatbot over many sessions. The social pull that drove early engagement faded in a predictable way, so findings from a single session don't hold over weeks. Does chatbot personalization build trust or expose privacy risks? finds a ratchet effect: each good interaction raises the user's expectations, so the same mistake feels worse later than it would have early on. Over many conversations, then, the same chatbot can seem to get worse without its output changing at all.

The strongest evidence that capability isn't the main factor comes from an unexpected place. Can models learn to abstain when uncertain about predictions? shows that small models trained to know when they're unsure, and to decline to predict, match models ten times their size at forecasting how conversations will go. The pattern across these notes is that what decides how a chatbot handles a long conversation is mostly which habits it was trained into: asking, checking, admitting uncertainty, repairing misunderstandings. Size matters less. Scaling up a model trained to please on every single reply mostly scales up how confidently it goes wrong.