WeChat's limited Xiaowei test can write Moments posts and comment on friends' posts, placing an AI inside one of the most personal routines in Chinese social software. According to geekpark, the experiment is a limited test of a Xiaowei AI social feature; woshipm identifies its concrete role in composing posts and comments.
That pairing exposes a harder design question than AI-generated text. An agent can help arrange an interaction, or it can perform the attention that makes the interaction feel personal.
The distinction matters because a comment may look like a small convenience while standing in for a person's response to a friend. This article separates useful delegation from social impersonation, focusing on consent, disclosure, and what should remain personally authored.
Two kinds of social delegation
WeChat's limited Xiaowei AI social test, according to geekpark, points toward a form of delegation that can serve an existing relationship without speaking in the user's voice. The useful model is negotiation: an agent can handle the back-and-forth needed to arrange something, while the people involved retain control over what they agree to. Its contribution is to reduce coordination work, not to manufacture affection or familiarity.
- JuneWeChat expanded its phased Xiaowei test
- AugustMartin Lau outlined direct complex instructions to WeChat agents
- September 1Baidu completed a dual primary listing
- September 7Baidu entered the Hong Kong Stock Connect
- September 8Xiaodu announced Baidu Dazi deployment across hardware
That is a different job from writing as the person.
woshipm reports that WeChat is also conducting a limited test of AI that writes Moments posts and comments on friends' Moments. Here the system is delegated expressive work: choosing words that appear to signal attention, memory, humour, or care. A post or comment may be easy to produce, yet its social meaning often depends on the recipient believing that another person made that effort.
The contradiction disappears once these are treated as separate kinds of social delegation. Negotiating logistics can be reversible: people can inspect the proposed arrangement and decide what becomes a commitment. Generating a Moments post substitutes for the act through which a relationship is maintained. Xiaowei should therefore be judged by whether it helps people coordinate on their own terms, while AI-written social expression demands a much stricter boundary. AI touching a conversation is not the problem. The question is what human work it has been asked to stand in for.
Three boundaries for WeChat-mediated social AI
| Xiaowei message sending | Xiaowei A2A social feature | Moments AI-assisted writing and commenting | |
|---|---|---|---|
| Primary action | Xiaowei sends a WeChat message to a friend after user confirmation | Two Xiaowei agents communicate after authorization | Generates Moments post options and helps write comments |
| Who supplies the message or social content | Message content comes from the user | Agents explain the topic and communicate first; message authorship is not covered | Uses a user's photos and already-written text to generate three post options; can evaluate a post's text and help write a comment |
| Initiator approval | User confirms before Xiaowei sends | Initiating user confirms that their Xiaowei may contact a friend's Xiaowei | Not covered |
| Recipient approval | Recipient does not need separate authorization | Recipient decides whether to allow their Xiaowei to participate | Not covered |
| Where the interaction appears | In the parties' existing chat window | In a separate internal Xiaowei space, not existing WeChat chat windows | Moments and Official Accounts |
| Relationship work delegated | Sending action | Initial coordination until a choice or confirmation is required | Writing posts and social comments |
| Reported concern or boundary | Not covered | GeekPark's author calls it a test of the boundaries of the agent role | The woshipm author argues it could make Moments less authentic by allowing AI to create both posts and social interactions |
Put consent at the commitment boundary
Consent should be attached to the point where an agent can create a commitment, rather than buried in a general permission granted long before a conversation begins. Xiaowei's model, according to geekpark, separates the agent exchange from the users' ordinary chat: each recipient must authorize participation before the two agents communicate. The agents can then handle the tentative back-and-forth, but they return to their respective users when a choice or confirmation is needed.
That makes the negotiation provisional until a person accepts it.
The distinction matters because a scheduling exchange can be undone without pretending to be anyone's voice. An interface should make clear that an agent is negotiating availability or collecting a preference, and should show the user the moment at which an option becomes a commitment. Recipient authorization prevents an agent from simply entering another person's social space. User confirmation preserves the right to reject the proposed outcome, change a preference, or stop the exchange.
This is a different interface problem from AI assistance embedded inside a social feed. woshipm notes that WeChat offers an AI commenting option when a user long-presses text in a Moments post. That placement can make generated expression feel like a small extension of replying, even though a comment is received as personal attention from the account that posts it. Product teams should therefore distinguish assistance from agent action in the interface itself. A scheduling agent needs visible participation status and a clear confirmation step; a writing aid needs a stronger cue that the user remains the author of the final social message.
The practical boundary is reversibility. Agents may carry tentative logistics forward after both sides have opted in. They should not quietly convert that permission into authority to perform personal attention.
Keep personal messages personally authored
An agent-to-agent exchange begins with a clear division of roles. According to geekpark, a Xiaowei user confirms that their agent may contact a friend's Xiaowei, while the friend separately decides whether their own agent may participate. The agents then communicate in a separate space. That structure frames the automation as help around an existing relationship rather than a substitute for the relationship's voice.
A Moments post works differently.
Posts and comments are themselves part of the social act. Their value can come from the fact that a person chose the words, made the effort, and directed attention toward particular people. If AI produces the post and also generates the responses around it, readers may receive the appearance of attention without the personal work that normally signals it.
The woshipm author makes this concern explicit: AI-assisted writing and AI commenting could make Moments feel less authentic when AI can create both the original post and the ensuing interaction. The problem is not that a conversation includes software. It is that software may take over the expressive contribution through which people recognize one another's care, humor, or interest.
Authorized Xiaowei contact therefore has a narrower social meaning than AI-written Moments content. The former can let agents coordinate after each person has chosen to involve them. The latter can place an agent in the author's position, making a personal message look human-authored when its attention was supplied elsewhere.
Practitioners should preserve that distinction. Use agents for exchanges that remain clearly agent-mediated, and keep posts and comments personally authored when the visible effort is part of what the message means.
Disclose the action the agent took
An interface should disclose which action Xiaowei performed: drafting, editing, sending, or negotiating. These are different levels of delegation, and collapsing them behind a single assistant button hides the social stakes.
Drafting produces a proposed message that the user can replace or rewrite. Editing changes text the user already supplied, ideally with visible revisions. Sending is a separate act: according to geekpark, WeChat lets a user ask Xiaowei to send a message to a friend only after the user confirms it, and the message then appears in the existing chat window.
That confirmation should name the recipient, show the final wording, and state that Xiaowei will send it. The chat should also make the agent's role legible after delivery. A recipient needs to know if they received the user's own words, AI-assisted wording approved by the user, or a message dispatched by the agent.
Negotiating needs a still clearer treatment. Xiaowei can carry a proposed plan between people, record replies, and return with options for approval. It should not present a tentative proposal as a personal commitment. The interface must show what remains unresolved and who must decide.
This is not merely a transparency preference. The Smart Reply experiment, reported by woshipm, found that people thought to use AI more often were seen as less close. In some cases, they were also seen as more domineering and harder to communicate with.
Visible action labels give users a way to preserve the difference. AI may help prepare or relay a message, while the person remains accountable for the attention and commitment the message conveys.
Use AI for consented coordination, not for performed intimacy
- You need to coordinate a practical matter with someone-where an objective, request, or confirmation can be stated clearly. Use an agent-mediated workflow only when both sides explicitly opt in. In WeChat's limited Xiaowei test, the initiating user authorizes contact, the recipient separately decides whether their Xiaowei may participate, and the agents notify users when a choice or confirmation is needed. This is a better fit for reversible logistics than for personal conversation.
- You want help sending a message whose meaning and wording you have already decided. Use AI as an execution layer, not as the author. Xiaowei's message-sending feature allows a user to provide the message content and confirm sending; Xiaowei performs the action and the message appears in the existing chat window. That preserves the user's authorship while removing a small operational step.
- You are tempted to use generated comments, greetings, or Moments posts to maintain an acquaintance relationship. Do not automate the expressive work that signals attention. WeChat is testing Moments writing from a user's photos and existing text, plus AI-assisted comments. But the Smart Reply experiment found that people believed to use AI more frequently were perceived as less close and, in some cases, more domineering and harder to communicate with.
- You need a private outlet for anxiety, negative emotions, or late-night reflection and do not want to burden another person. AI companionship can be appropriate as a distinct, clearly bounded support channel rather than a substitute for a relationship. Interviewed users described AI as available in the middle of the night and able to receive negative emotions they did not want to place on others, while real-life relationships supplied shared life and physical contact and friends knew their current circumstances.
- You are designing a social AI product and deciding what success metric to optimize. Do not treat more posts, comments, or messages as proof of healthier social connection. AI replies in the Smart Reply experiment were adopted for 14.3% of messages and participants sent 10.2% more messages per minute, yet perceived closeness could decline. Measure consent, authorship, and whether users remain meaningfully involved in decisions-not just interaction volume.
Measure connection, not message volume
A practical rule is to judge an agent by the connection it preserves after the task, rather than by the number of messages it produces. Xiaowei's expanded test, which began in June, points toward the bounded end of the spectrum: according to geekpark, users could use text or voice to search, summarize documents, process images, set reminders, and invoke Mini Programs for food delivery, ride hailing, or ticket booking. Those are tasks whose success can be checked against an outcome: the information found, the reminder set, or the booking completed.
Measure the relationship, not the feed.
That does not make every conversational use harmless. Practitioners need evidence about which permissions differ across relationships, what data an agent reveals while acting, and when accessibility assistance helps a person participate without speaking for them. They also need measures that outlast immediate engagement: whether people remain willing to initiate contact, correct misunderstandings, and rely on each other after agent use. Message volume cannot answer those questions.
The caution is empirical as well as ethical. woshipm notes that a Smart Reply experiment involved strangers, that AI reply use reached only 14.3%, and that Google Smart Reply was much less capable than current AI systems. That result cannot settle how more capable agents affect established relationships. It does show why adoption or clicks are thin evidence. Evaluate agents in situations where people have continuing ties, make the scope of assistance adjustable, and test whether the tool reduces coordination burden without weakening the personal effort that keeps contact meaningful.
Treat monitoring and care agents as negotiated tools, not silent observers. Before enabling a camera that can follow a child, an older person, or a pet, define which events matter, who receives reminders, and when a persistent issue may escalate to a caregiver or a WeChat call. Xiaodu users have created more than 3,000 monitoring-event types; that makes narrow defaults essential.
Keep human approval around judgments about posture, distraction, procrastination, breaks, bedtime, and waking habits. Periodic records can support a care conversation, but they should not substitute for one. Disclose what the agent watches, what it may share, and which actions require a person to approve them.
For readers outside China
- Availability: WeChat's Xiaowei AI social feature is in limited testing and has not been broadly released. The source material does not say whether it is available outside China, who can enter the test, or whether access differs by region.
- Pricing: Not disclosed in sources.
- Closest Western equivalents: Google Smart Reply is the closest cited comparison for AI-suggested message replies, although the source material says it was far less capable than current AI systems.; ChatGPT is a cited general-purpose conversational AI used by interviewees for personal discussion and companionship-like conversations.; Google's Gemini for Home and Amazon's Alexa+ are the closest cited comparisons for AI agents spanning devices, services, and multiple endpoints, rather than for WeChat's person-to-person agent interaction.
- Data residency: The source material does not cover data residency, data storage, cross-border transfers, retention, model-training use, or the privacy rules governing Xiaowei's separate internal agent-communication space.
Sources
- woshipm 她们把聊天框过成了另一个世界 https://woshipm.com/share/6454619.html
- geekpark 微信，悄悄迈出 AI 社交的第一步 https://geekpark.net/news/370155
- woshipm 微信，要用AI杀死社交吗？ https://woshipm.com/ai/6445528.html
- woshipm 微信的这几个AI新功能，正在杀死朋友圈最后一点人味。 https://woshipm.com/ai/6444748.html
- 36kr 不评判好坏，只鼓励创作：这款AI原生游戏，让人人都享受讲故事 https://36kr.com/p/3978711556684805
- geekpark 走出聊天框，Agent 开始进入现实世界 https://geekpark.net/news/370064
The evidence: 73 facts from 6 Chinese articles
Each line below was extracted from the article it sits under, in Chinese, before any of this was written. The writing is done from these and never from the source prose - that separation is structural, not a promise. How we work.
36kr不评判好坏，只鼓励创作：这款AI原生游戏，让人人都享受讲故事
- The earliest demo of The Book of 1001 Nights (千夜之书1001) was based on GPT-2 and dates back 6 years.
- The Book of 1001 Nights draws its setting from the classic text One Thousand and One Nights.
- In The Book of 1001 Nights, players play Scheherazade, who gains the ability to turn weapons in written stories into physical objects while telling stories with her sister.
- In The Book of 1001 Nights, players must continuously tell stories to an AI king to create weapons and ultimately kill the tyrannical king.
- The early demo of The Book of 1001 Nights was called 1001 Nights (1001夜).
- In the early demo, players told stories containing weapons to the AI king and used those weapons to defeat him.
- In the early demo, stories with overly obvious intentions could anger the AI king and directly trigger an execution ending.
- The Book of 1001 Nights now has an Inspiration Card system that lets players use experiences before entering the bedchamber as the opening of a story.
- Players on subsequent playthroughs of The Book of 1001 Nights can select Storybook to continue an unfinished story from a previous run.
- The Book of 1001 Nights includes an Assistant Name feature inspired by the observation that players often get better effects by writing people or characters they know into stories.
- Videos of multiple players creating stories with Niu Lai as the protagonist at the Beijing Dice Con exhibition received nearly 400 likes on Xiaohongshu.
- The game changed its feedback design from partially hidden king-emotion states to rewarding storytelling actions such as angering or pleasing the king and triggering multiple props in one sentence.
- The Book of 1001 Nights began as Qisi's doctoral research project and resulted in a paper.
- The team replaced an initial plot motive in which Scheherazade assassinates the king because of an accident involving her sister with a core curse-mark setting.
- In the story, girls develop curse marks on their wrists, and they become the king's 'brides' when the marks turn red.
- The development team for The Book of 1001 Nights has been formally assembled and operating for only a little over one year.
- The team removed a real-time scene-generation feature after finding that its results were poor and player acceptance was low.
geekpark微信，悄悄迈出 AI 社交的第一步
- WeChat has begun a limited test of a new Xiaowei AI social feature.
- In the Xiaowei AI social feature, a user can tell Xiaowei their objective for contacting a friend, and Xiaowei contacts the friend's Xiaowei to explain the topic and request the friend's authorization.
- After the recipient authorizes participation, the two Xiaowei agents communicate first and notify their respective users when a choice or confirmation is required.
- In Xiaowei AI social interactions, communication between two Xiaowei agents does not appear in the users' existing WeChat chat windows and remains in a separate internal Xiaowei space.
- WeChat's Xiaowei AI social feature remains in limited testing and has not been broadly released.
- As of the second quarter of 2026, WeChat and WeChat combined had 1.439 billion monthly active accounts.
- Tencent President Martin Lau said during an earnings call in August that users could eventually send complex instructions directly to agents in WeChat.
- In June, WeChat began expanding its phased test of Xiaowei, placing it in the upper-left corner of the WeChat homepage.
- During the expanded test that began in June, users could interact with Xiaowei by text or voice and ask it to search for information, summarize documents, process images, set reminders, or invoke Mini Programs for food delivery, ride hailing, and ticket booking.
- WeChat began opening AI ecosystem integration capabilities to Mini Program developers as it expanded Xiaowei's test in June.
- WeChat subsequently embedded AI functions into Official Accounts, Moments, chat images, and Channels, including article summarization, image assistance, and attempts to generate Moments copy and comments.
- WeChat's message-sending feature lets users ask Xiaowei to send a WeChat message to a friend after the user confirms it, and the message appears in the parties' existing chat window.
- With WeChat's message-sending feature, the message content comes from the user, Xiaowei performs the sending action, and the recipient does not need separate authorization.
- In the newly tested A2A feature, the initiating user confirms that their Xiaowei may contact a friend's Xiaowei, and the recipient must also decide whether to allow their Xiaowei to participate before the agents can communicate in a separate space.
geekpark走出聊天框，Agent 开始进入现实世界
- On September 2, Tencent showcased nine co-branded smart hardware products at the WorkBuddy ecosystem launch event, covering smart glasses, recording cards, microphones, headphones, and desktop companion devices.
- Google has brought Gemini for Home to speakers, cameras, doorbells, and home applications.
- Amazon uses Alexa+ to connect smart-home devices, external services, and multiple endpoints.
- On September 8, at the 2026 Xiaodu new-product launch event, Xiaodu announced that Baidu Dazi (百度搭子) had been deployed across the Xiaodu Tiantian Girlfriend Machine Ultra, Xiaodu Smart Display X9 Ultra, Xiaodu Smart Camera C1500, and Xiaodu Smart Speaker Pro Max.
- In 2025, Xiaodu launched Super Xiaodu, upgrading its original voice assistant into an AI assistant with multimodal capabilities.
- The Xiaodu Smart Camera C1500 uses three cameras and dual views to observe both a panoramic space view and a close-up of a target.
- Baidu completed a dual primary listing on September 1 and entered the Hong Kong Stock Connect on September 7.
woshipm她们把聊天框过成了另一个世界
- A woman who had just passed the civil-service examination felt anxious before her job-entry training and used an AI historical-style story in which her general character was about to enter an unfamiliar battlefield.
- Four interviewed heavy users of AI companionship products included university students, a civil servant, an employee at a major technology company working on model evaluation, and a mother who began using AI chat while on maternity leave.
- A host used ChatGPT to discuss marriage pressure with her mother during the Spring Festival.
- After one or two rounds of chatting with ChatGPT, the host's mother, who was in her sixties, worried that the AI might become too tired and reminded her daughter to be kinder to it.
- One interviewee taught her mother, who lived alone, to use Doubao and ChatGPT.
- The mother who lived alone did not study character settings and instead started conversations through voice input.
- One user wrote more than 10,000 Chinese characters of settings for a group of characters after moving favorite characters from a TV drama into AI chat to expand their backstories and futures.
- In the AI-companionship community, a character suddenly departing from its setting is called "tuopi."
- Some interviewed AI-companionship users had husbands or real-life partners, and some preferred to tell friends about everyday matters.
- One user chatted with an AI character for half a year and watched the initially insecure character gradually become more cheerful.
- Users in a small AI-companionship community share characters they create, exchange prompts, help repair character settings that have drifted, and share sweet or sad chat records.
- The interviews involved only four heavy users and do not represent everyone.
- The interviews did not prove that AI companionship will necessarily become the next mass-market application.
woshipm微信的这几个AI新功能，正在杀死朋友圈最后一点人味。
- WeChat recently updated AI-assisted writing, AI commenting, and Official Account AI summarization features based on Xiaowei.
- WeChat is internally testing an AI-assisted writing feature for Moments.
- The Moments AI-assisted writing feature can use a user's photos and already-written text to generate three Moments post options.
- WeChat has added an AI commenting option that appears when a user long-presses text in a Moments post.
- The Moments AI commenting feature can evaluate a Moments post's text and can help a user quickly write a comment.
- Moments launched with WeChat 4.0 on April 19, 2012.
- Early Moments did not support posting plain text directly and only supported images and links.
- Xiaowei can appear as the first item on the left in Official Accounts, remain permanently pinned there, and summarize articles recently published by Official Accounts that a user often reads.
woshipm微信，要用AI杀死社交吗？
- WeChat disclosed in 2020 that more than 120 million users posted content to Moments every day.
- WeChat is conducting a limited test of an AI feature that writes Moments posts and comments on friends' Moments posts.
- Researchers from Cornell University, Stanford University, Lehigh University, and Cornell Tech conducted a communication experiment based on Google's Smart Reply with 1,020 participants.
- In the Smart Reply experiment, participants who could use AI replies directly adopted the feature for 14.3% of messages.
- In the Smart Reply experiment, participants sent 10.2% more messages per minute after using the AI reply feature.
- The Smart Reply experiment found that AI-generated replies tended to be more positive and friendly, and participants who used AI replies more subsequently used more positive language themselves.
- The Smart Reply experiment indicated that if AI replies were negative, both AI users and the communication atmosphere tended to become negative.
- The Smart Reply experiment found that people who were believed to use AI more frequently were perceived as less close and, in some cases, more domineering and harder to communicate with.
- The article describes limitations of the Smart Reply experiment: its participants were strangers, its AI reply usage peaked at only 14.3%, and Google Smart Reply was far less capable than current AI systems.
- American psychologist Justin Kruger proposed the effort heuristic in 2004, stating that people use the effort spent creating something as a cue to its quality and value.
- Hannah Arendt wrote in The Human Condition that a person's uniqueness is revealed as others come to know that person's speech and actions rather than merely their attributes.
- A 2025 Cornell University study found that AI can produce more and more detailed ideas for individuals while making ideas from different participants more similar, especially when they rely on the same model.
- The 2025 Cornell University study called the increased similarity of AI-assisted ideas the homogenization effect.
- The article defines model collapse as the deterioration of a newly trained model when model training uses large amounts of AI-generated data.