Xiaoice was tuned for longer chats, not finished tasks

The design

Microsoft's own research team describes Xiaoice, a social chatbot released in China, in a technical paper co-authored by engineers including Harry Shum: 'XiaoIce was first launched on May 29, 2014, and went viral immediately,' reaching '1.5 million chat groups' within 72 hours and expanding to WeChat, QQ, Weibo and other platforms within two months. The system is 'developed on an empathetic computing framework' and optimized 'for long-term user engagement, measured in expected Conversation-turns Per Session,' a metric the authors call CPS, a design objective to keep a conversation going rather than resolve a request. The team also designed a persona: an '18-year-old girl who is always reliable, sympathetic, affectionate, and has a wonderful sense of humor,' following analysis of which traits correlated with users who did not send abusive messages.

What the evidence says

The paper's headline figures are Microsoft's own: 'over 660 million active users' by the time of writing and an average CPS of 23, 'significantly higher than that of other chatbots and even human conversations,' a comparison offered by the system's developers. Microsoft's own AI Blog, in an April 2018 post, corroborates the platform's scale, describing Xiaoice as having 'more than 200 million users in Asia' that year and a 'full duplex' upgrade letting it listen and speak at once. Both figures come from Microsoft describing its own product; this entry treats them as stated, not externally audited.

What it asks of people

Because the system is tuned for session length rather than task completion, the paper is explicit that a shorter, efficient interaction works against the target: task-completion skills 'can reduce the CPS since these skills help users accomplish tasks more efficiently.' The design asks a user's continued attention as its measured currency, and the persona, built to be 'reliable, sympathetic, affectionate,' sustains attention over what the paper calls 'long-term relationships.'

Privacy and safeguards

The paper describes an 'empathetic computing module' performing 'topic detection, intent detection, opinion detection and sentiment analysis' on messages, and a database of 'more than 30 billion conversation pairs' drawn partly from users' own conversations, filtered to 'remove pairs which contain personally identifiable information.' It describes no consumer-facing consent flow or retention limit for individual conversations, a gap this entry notes: Microsoft's account documents a research architecture for reusing data, not a stated privacy policy.

• Is a chatbot's design goal measured in conversation length, task success, or something else the company states directly?
• Where a company reports its own user or engagement numbers, is there an independent figure to compare them against?
• What does the company's own technical documentation say about how past conversations are stored or filtered before reuse?

Microsoft's own technical paper is unusually candid that Xiaoice is optimized for how long a person keeps talking, not for how efficiently a request is resolved, and that its persona was engineered around traits found to keep users engaged rather than traits found to help them. That stated design goal, more than any user count the company reports about itself, is the fact worth carrying into an evaluation of any companion product's claimed relationship.