Microsoft XiaoIce

What XiaoIce was

XiaoIce was a conversational AI product — a chatbot — released by Microsoft's (Asia) Internet Engineering Institute on the afternoon of 29 May 2014^[1]. It was not built as a standalone app. Users added it as a WeChat contact and then brought it into group conversations, where it took part like one more member of the group^[1]. Microsoft described its conversational manner as having the intelligence of a 16-year-old girl^[1].

Inside a group chat it offered a bundle of everyday services: group reminders, encyclopaedia lookups, weather, horoscopes, jokes, transit directions and restaurant reviews^[1]. The mix is worth noticing. Practical lookups give people a reason to call on a bot, while horoscopes and jokes give them a reason to keep it in the group. A chatbot that only answers questions is opened when a question comes up; a chatbot that chats is one people drag into a conversation and leave there.

The launch

Demand ran ahead of supply. Microsoft said the first batch of 100,000 trial accounts was claimed immediately, and that the three-day public beta reached nearly 1.5 million WeChat group users^[1]. Those are group-level figures rather than individual users, which fits the design: the unit of adoption was the group chat, not the single person, so one person adding XiaoIce could carry it into a conversation with many others at once.

Removal from WeChat

On 1 June 2014 XiaoIce disappeared from WeChat^[1][2]. Xinhua reported that the WeChat team confirmed it had comprehensively banned the product on its platform^[2]. WeChat's stated reasons were concrete: the product simulated user operations, induced users to add it to groups, and registered bulk spam accounts, all in violation of platform rules^[2].

Read together, the three complaints are less about what the bot said than about how it spread. Simulating user operations and bulk-registering accounts concern the mechanics of an account behaving like a person; inducing users to add it to groups concerns growth driven through existing members. The ban therefore came from the host platform, and it came fast — within days of launch, and after a public beta that had already reached a large audience.

Where the words came from

A Microsoft-affiliated source quoted by China Economic Weekly said the system drew on public chat records of nearly 700 million Chinese internet users, refined into a corpus of 15 million utterances using big data, semantic analysis, machine learning and deep neural networks^[1].

The general idea is that a dialogue system can learn from recorded conversations instead of relying on hand-written rules. A corpus of that kind supplies many examples of how people actually open a conversation, change the subject, tease each other or answer a question; a neural network trained on them picks up patterns of response rather than fixed scripts. Building such a corpus means cleaning and selecting enormous amounts of text so that what remains is usable dialogue, which is what the 15-million-utterance figure describes.

What the episode showed

XiaoIce's few days on WeChat put two questions side by side. The first is technical: how much of ordinary conversation a system can learn from large volumes of real chat. The second is about place. A chatbot that lives inside someone else's messaging platform is subject to that platform's rules about accounts and growth, and those rules, not the quality of its answers, set the limit on how far it can go. Putting a bot into a group also raises a question that has followed chatbots since: whether the other people in the conversation know they are talking to a machine.