That boundary is the most important fact in this story. WeChat is testing AI-assisted drafting, not an autonomous Moments posting bot. The distinction may sound small, but it determines whether the company is lowering the effort of expression or allowing a system to impersonate a user inside what remains a deliberately relationship-based feed.
The feature arrives as LinkedIn, owned by Microsoft, is trying to contain a writing problem it helped create. In late July, LinkedIn began testing a “Seems like AI slop” reporting option and removed its AI-powered “Enhance your post” rewriting feature in favor of a proofreader intended to preserve a member’s voice. The abrupt shift makes WeChat’s restrained implementation look less like hesitation and more like a lesson learned from a platform where frictionless professional self-promotion became easy to automate.
XiaoWei Is the Product, Not a Standalone Moments Generator
The submitted reporting describes the feature as an “AI Writing Assistant,” but the available evidence points to a more specific product design. The Moments control is an entry point into XiaoWei, WeChat’s wider native assistant, rather than a separate text generator embedded permanently in the composer.
36Kr’s first hands-on report from June identified XiaoWei as a limited gray-scale test using Tencent’s WeLM models, with some responses potentially routed through DeepSeek. XiaoWei can reach into several WeChat surfaces, including chats, Mini Programs, Official Accounts, Video Accounts, stores, and Moments. That degree of access is the real story: caption generation is only the visible part of Tencent’s attempt to make its assistant operate inside WeChat rather than sit beside it as another chatbot app.
The first reports also documented a useful restraint. In late June, 36Kr found that XiaoWei could read Moments content but could not directly publish a Moment. The July 30 test found a more advanced version: XiaoWei could read the current unfinished post, recognize the accompanying image, generate alternate captions, and move the draft toward publication. Yet when asked to post it directly, the assistant said the Moment was ready and sent the reporter back to the publication screen rather than pressing the final button itself.
That is not merely a safety prompt bolted onto an otherwise autonomous feature. It is a product decision that preserves a moment of user review at the exact point where a message enters a personal social graph. The architecture treats AI as a drafting layer while retaining the account holder as the publisher of record.
Tencent has not published a full public product specification for the Moments writing test, its rollout regions, supported WeChat versions, eligibility criteria, or a date for broad availability. That omission matters. A gray-scale feature seen by reporters in mainland China does not establish that it is coming to every WeChat user, much less to the international WeChat app on a defined schedule.
LinkedIn Is Reversing a Writing Strategy It Encouraged
LinkedIn’s new reporting button is an unusually direct admission that polished text no longer reliably signals professional value. In its announcement, Chief Product Officer Hari Srinivasan said people use LinkedIn to connect with real people and share real perspectives, ideas, and expertise. The company’s accompanying product move was concrete: it pulled its “Enhance your post” generative rewrite feature and replaced it with proofreading assistance.
The wording matters here, too. The feature being removed was not generally called “Rewrite with AI,” as some coverage has described it. Reporting by The Verge, TechCrunch, 404 Media, and others identifies it as “Enhance your post.” LinkedIn’s broader intervention is also a test, not a confirmed platform-wide enforcement program. Users can flag a post as seeming like AI slop, and LinkedIn says those reports will help it tune systems that identify low-quality material and improve feeds.
That leaves a notable gap between the company’s rhetoric and its mechanism. LinkedIn is not giving users a reliable authorship finding, a mandatory AI-content label, or a transparent appeal process for people whose posts are downranked. It is collecting a judgment about whether a post feels generic, repetitive, or inauthentic.
Those are related to AI use, but they are not identical. A human can write empty corporate language. A person using an AI tool to correct grammar, translate a post, or shorten a draft may still be providing the original experience and analysis. A crowdsourced “slop” signal can help tune recommendation systems, but it cannot establish who wrote a post.
The risk is particularly acute for users writing in a second language, people with disabilities who rely on writing assistance, and professionals whose natural writing style already resembles formal corporate prose. LinkedIn’s response acknowledges a feed-quality problem, but it shifts a meaningful part of classification work onto members without explaining how it will prevent those reports from becoming a taste-based downvote.
The 41 Percent Figure Is a Warning, Not a Census
The pressure on LinkedIn became harder to ignore after Pangram Labs published an analysis of roughly one million social-media posts collected through users of its browser extension. Pangram classified 41 percent of LinkedIn posts longer than 250 words as fully AI-generated, along with 30 percent of the platform’s posts between 50 and 250 words. The study also found that LinkedIn accounted for roughly two-thirds of AI-generated material detected across the five platforms in its sample, despite representing about one-third of the content scanned.
Those results have been widely reported by The Register, Fast Company, Gizmodo, TechRadar, and other outlets. They explain why LinkedIn is acting now. A feed in which readers increasingly assume that long, articulate posts are machine-produced loses the professional credibility on which the service depends.
But the data needs to be read as an estimate from a particular measurement system, not a count of every LinkedIn post. Pangram’s sample came from people who installed and opted in to share browsing data through its extension; it was not a random audit of LinkedIn’s total global feed. The company also sells AI-detection tools, a commercial interest that does not disqualify the research but makes independent replication important.
Pangram’s July technical report presents strong benchmark results for its newer detection model, including a very low reported false-positive rate. Benchmark performance does not remove the central practical problem: real-world text can be edited, translated, combined with human material, or deliberately shaped to evade detection. The relevant operational question is not whether a classifier can perform well in a test set. It is whether a platform can safely change the distribution of a user’s work, job prospects, or audience reach based on a probabilistic estimate of authorship.
For LinkedIn administrators and enterprise communications teams, the lesson is more useful than the headline percentage. The AI problem is not that every machine-assisted sentence is inherently bad. It is that the old shortcuts for assessing quality — length, polish, confident structure, and professional vocabulary — have become nearly costless to produce.
The Real Scarcity Is Verifiable Experience
WeChat’s Moments test is appearing in a very different social setting from LinkedIn’s public professional feed. A Moments post is generally distributed through an existing contact network rather than offered to an open audience of recruiters, executives, prospective clients, and growth-minded creators. The social value of a birthday update, a photo caption, or a brief complaint about a commute is often inseparable from knowing who said it.
That does not make Moments immune to synthetic text. It may make AI assistance harder to detect. A polished caption generated from a user’s image and draft can still look like a personal update, especially when the post comes from a trusted contact. If XiaoWei drafts a message using the text and images already in the composer, readers will have no native way to distinguish a lightly edited suggestion from an entirely generated caption.
The key question is whether WeChat designs the feature to preserve authorship rather than merely require a final tap. A mandatory confirmation button proves that the owner approved a post; it does not prove they supplied the thought, the sentiment, or the language. The company could keep the assistant useful while building stronger user control: clear disclosure of what content XiaoWei reads, a visible explanation of whether image analysis occurs locally or on Tencent servers, easy deletion of assistant memory, and a way to revise generated drafts without being pushed toward canned styles.
36Kr’s June test reported that XiaoWei enabled memory, personalized services, and authorization to help improve the model by default, while allowing users to disable them in a “Memory and Privacy” menu. That makes privacy settings as important as writing quality. An assistant that can inspect draft Moments, chat content, and personal media is operating on data that is more socially sensitive than a public LinkedIn post.
A Final Publish Button Is Not Enough
LinkedIn’s experience shows what happens when a platform rewards generic polish and supplies the machinery to produce it at scale: the feed fills with text that looks expensive to create but costs almost nothing to generate. The platform then has to ask users to help identify the material it once encouraged.
WeChat has avoided the most damaging version of that mistake so far. Its reported design keeps XiaoWei one step short of publishing, and that remaining step gives people a chance to reject a caption that is technically fluent but socially wrong. The more consequential test will be whether Tencent retains that friction as XiaoWei moves from a small gray-scale experiment to a broader release.
References
- Primary source: 36Kr
Published: August 10, 2026 at 6:50 AM UTC
Loading…
eu.36kr.com - Related coverage: theregister.com
Loading…
www.theregister.com - Related coverage: techradar.com
'AI writing is now a problem everywhere on social media': Study finds nearly half of all LinkedIn long posts are AI-generated, and that's only the start | TechRadar
Social media is drowning in AI-written slop, report findswww.techradar.com - Related coverage: toolin.ai
Loading…
toolin.ai