---
title: "From 380,000 hours to ¥10 billion: digital humans take China's livestreams"
date: 2026-10-04
category: Generative Media
site: NeuroAI
canonical: https://neuroai.site/a/na-aigc-digital-humans-livestream
language: en
---

# From 380,000 hours to ¥10 billion: digital humans take China's livestreams

> China's e-commerce platforms are deploying AI digital humans (数字人) as round-the-clock livestream hosts, blurring the line between virtual and real presenters while regulators tighten the rules.

A shopper in a midnight livestream taps "buy" without realizing the host smiling back isn't breathing. Across China's biggest shopping apps, the person pitching the product may never have been a person at all.

The rise of the AI digital human (数字人) as a livestream salesman has moved from science-fair curiosity to a line item on corporate balance sheets. What changed is not the idea — it is the believability, and the economics.

## The overnight shift that isn't staffed by humans

Live commerce in China runs on hours. Human hosts tire, unions negotiate, and the 2 a.m. slot is where sales go to die. Digital avatars don't sleep, don't ask for a cut of GMV, and don't quit.

That pitch has convinced the country's largest retailers to put virtual presenters in front of real shoppers at scale. The results, at least as reported by the platforms themselves, are no longer trivial.

## What a digital human (数字人) actually is

Strip away the marketing and a livestream digital human is three glued-together systems:

- A **large model (大模型)** that writes the script and answers viewer questions in real time.

- A voice engine that speaks the lines with a cloned or synthetic tone.

- A rendered face and body — sometimes 2D, sometimes a full 3D model — driven by the text and audio.

The best versions now match lip movement to speech, react to comments, and hold a conversation well enough that viewers can't tell within the first two minutes. The worst are still the "electronic mannequins" everyone mocked a year ago.

## The 618 stress test

China's mid-year shopping festival became the proving ground. During the event, **JD's digital-human hosts logged more than 380,000 hours of live broadcast and over 4 million viewer interactions**, according to figures the retailer shared with the financial outlet Yicai.

JD said the avatars had served more than 5,000 brands and driven GMV of **more than ¥10 billion (≈ US$1.4 billion)**. Other industry estimates put the total digital-human GMV across platforms even higher, though those numbers are projections rather than audited results.

The takeaway for brands was simple: a virtual host could cover the dead hours, repeat the pitch flawlessly, and never post a controversial off-script remark.

## Why brands say yes

The appeal is not just novelty. Three forces push adoption:

- **Cost.** A single digital human can be produced for a fraction of a human host's per-hour rate, and it scales to dozens of rooms at once.

- **Compliance.** A scripted avatar won't accidentally make a false claim or spark a scandal — the risk surface is narrower than a freewheeling human.

- **Coverage.** Off-peak and multilingual broadcasts that were never economically viable suddenly are.

According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, digital humans are shifting from marketing novelty to baseline infrastructure inside China's commerce stack, with the real question no longer whether they work but where they should not be used.

## Where platforms drew the line

Not every platform rolled out the red carpet. As digital humans flooded in, the gatekeepers got nervous about a feed full of non-human sellers.

- **Douyin** published a virtual-anchor convention back in May 2023 requiring avatars to be filed in advance, carry a clear label, and pass pre-review.

- **Kuaishou** issued an AIGC-livestream guide in June 2025 saying it would give no special traffic boost to AI-assisted content and would penalize low-quality machine-made streams.

- **WeChat Channels** revised its rules in June 2025 to classify AI virtual-image livestreams as "non-real livestreams," with penalties for selling the proxy software.

The message: you can use a digital human, but you cannot pretend it's a person, and you cannot flood the platform with cheap filler.

## The law is catching up

Under China's **deep-synthesis rules (深度合成规定)**, in force since January 10, 2023, any service that generates or significantly alters a human face, voice, or virtual scene must add a prominent label so the public isn't confused. Digital-human livestreams fall squarely inside that definition.

The newer **AI-content labeling measures (标识办法)**, effective September 1, 2025, push the requirement further down the chain — not just the generator, but the platform that distributes the content must surface a notice.

For a shopper, that means the avatar should, in theory, be tagged. In practice, enforcement on small sellers is uneven.

## Honest limitations

This article leans heavily on figures the platforms reported to Chinese media (Yicai, iResearch); none were independently audited, and conversion and GMV comparisons vary by source. We focused on China's e-commerce livestreams and did not survey adoption in the US, Europe, or Southeast Asia. Enforcement of labeling rules remains inconsistent across platforms and seller tiers. We did not measure long-term consumer trust or whether digital-human pitches actually convert better than humans outside the festival window. The deepfake-fraud and broader ethics debate is covered only in passing.

## What readers can do now

- If you run a cross-border store, pilot a **digital human (数字人)** for off-peak hours on a China-facing platform, but keep a human on trust-heavy pitches like cosmetics or health claims.

- Before posting AI-generated product videos in China, confirm they carry a visible label — the **标识办法 (labeling measures)** makes it mandatory as of September 1, 2025.

- If you clone a spokesperson's face or voice, assume you need explicit disclosure under the **深度合成 (deep synthesis)** rules; build the tag into the asset, not the caption.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-aigc-digital-humans-livestream
Free to quote with attribution and a link to the original.
