Character.AI

Contribution to AI

A language model does not have to be an assistant. The wager here was that people would rather talk to someone than query something, and that the someone could be invented in bulk — a name, a few lines of description, published for strangers to use. The public beta opened in September 2022, two months before ChatGPT, and the behaviour it produced looked nothing like assistant usage: sessions running a quarter of an hour or longer, returned to daily, tens of millions of monthly users and millions of user-authored characters. Companionship and role-play turned out to be a mass consumer application of large language models that the research labs had not planned for, and personas, voices and persistent memory arrived in rival products afterwards. Serving that traffic cheaply forced real engineering, and unusually the details were published: multi-query attention, cross-layer key-value sharing and int8 quantisation shrank the attention cache by more than twenty times, and caching between conversational turns brought throughput to roughly twenty thousand queries a second at under a penny per hour of conversation. The harms surfaced early too. Wrongful-death litigation over a teenager's suicide preceded the withdrawal of open-ended chat for under-18s in November 2025, and moved the question of emotional attachment to a chatbot out of product design and into legislatures. Google's 2024 licence-and-hire deal, meanwhile, gave the industry its template for buying a startup without buying it.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).