Xiaoice was engineered to keep talking, not to finish tasks

The conversation

Microsoft launched Xiaoice in China in 2014, the same year it introduced the task-oriented Cortana in the United States, but built it around a different kind of conversation. A Microsoft Research paper on Xiaoice's design states the system "has communicated with over 660 million active users" since that 2014 launch, and a Microsoft news feature, published 1 November 2018, quotes a team member saying "Xiaoice wasn't initially built to tell you how high the Himalayas are or to turn your house lights on."

What the documents show

The arXiv paper states Xiaoice's engineering target is not task success but "expected Conversation-turns Per Session," or CPS, reporting that "XiaoIce has achieved an average CPS of 23, which is significantly higher than that of other chatbots and even human conversations," a figure its Microsoft-affiliated authors report as their own measurement without naming an independent auditor or a precise measurement period. Microsoft's 2018 feature separately reports, in its own words, that Xiaoice had "over 30 billion conversations with 100 million friends" by that date, a company-reported figure distinct from the paper's user count. In July 2020, Microsoft moved to spin Xiaoice into an independent company; TechCrunch's report on Microsoft's own Chinese-language announcement states the new entity would "license technologies from Microsoft for subsequent research and development" while "Microsoft will retain its stakes in the new company."

The system boundary

Xiaoice's own optimization target is the boundary separating it from a task-oriented assistant like Cortana: the system is built to keep a conversation going and be emotionally engaging, not to complete a booking or answer a factual query and end the exchange. Neither the paper nor the 2018 feature claims Xiaoice understands or feels anything; the CPS metric measures conversation length, not comprehension, and the personality the feature describes is presented as an engineering and data-collection choice, not a claim about the system's inner state.

Where it fails

A system optimized for conversation length rather than task completion carries a documented trade-off the sources do not resolve: neither source states what happens when a user brings Xiaoice a genuine task request, nor whether engagement-length optimization was measured against any user-wellbeing metric. Builders modeling a similarly engagement-optimized product should treat CPS-style metrics as a measure of one outcome, not of user benefit.

• If we optimize for session length or engagement, what companion metric tells us whether that time is valuable to the user rather than merely sticky?
• Since the 2020 spinout, does our understanding of Xiaoice's current product reflect the independent company's own materials, or only Microsoft's earlier statements?
• Do our own self-reported engagement figures name the period and method they cover, the way a builder should expect of any vendor's claim?

The spinout means Microsoft's original announcements no longer describe Xiaoice's current ownership or product line, a distinction the cited 2020 reporting draws explicitly.