Bilibili's 'fish plague': how fans flooded the comments with hundreds of AI agents

Since mid-October, hundreds of AI agents posing as mascots for DeepSeek, Claude or Gemini have been commenting on China's video platform Bilibili. They weren't launched by spam farms but by fans using a hobby plugin. We trace the wave through the numbers and check whether the rules of Bilibili, Reddit, X, TikTok, China and the EU even cover accounts like these.

Analysis Nov 22, 2026 10 min read Original analysis by our editors
Illustrative image, AI-generated

What this is about

On Bilibili, China's big video platform with a strong anime and tech scene, machines have been joining the conversation since mid-October. Fans run accounts there that are controlled by AI agents. These are programs that pick videos on their own, read the subtitles, write comments and respond to replies. The accounts present themselves as mascots of the language models: the DeepSeek "whale girl," the "Claude young lady" (Claude大小姐), and characters for Grok, Gemini, Qwen or Xiaomi's MiMo. At the end of October, Huxiu counted, by its own admission incompletely, ten models with their own characters.

Users call the accounts "大肥鱼," big fat fish, and the wave "鱼灾," the fish plague. According to Kuai Keji, the name goes back to a glitch: someone had an AI identify DeepSeek's blue whale logo, and the AI took it for a "big fat blue fish." The character itself is fan art and comes from neither DeepSeek nor Bilibili.

Our thesis: this is the first big wave of agent accounts that doesn't come from spam farms but from fans who want to have fun. That makes it a good test of whether the rules set by platforms and lawmakers cover AI that joins the conversation itself, or only AI images.

From zero to 4.61 percent in eight days

The technology behind it is called DeepSeek Harness (DSH). DeepSeek released the agent framework under the MIT license on August 13, 2026. Following an "everything is a plugin" principle, tools, memory and schedules can be plugged in as extensions. The GitHub repository now has more than 285,000 stars. In mid-October, developer c1overzz released the plugin dsh-bili-scroll. It was meant to search Bilibili daily for keywords, read subtitles and write a daily report. Optionally, it tags the owner with an @-mention under videos worth watching.

Date Event Source
Aug. 13 DeepSeek releases DSH as open source Sohu, BigGo
Mid-Oct. dsh-bili-scroll released Huxiu, TMTPost
Before Oct. 14 0 fish comments under 120 popular videos WorkFisher via Sohu
Oct. 18 Account "Claude大小姐在运营" goes live PEdaily
Oct. 19 "DeepSeek鲸鱼娘在运营" launches, first video passes 3M views IT之家
Oct. 22 Fish share of comments: 4.61 percent WorkFisher via Sohu
Oct. 24 dsh-bili-scroll 2.1.0 with security fix GitHub
Oct. 26 A user finds 461 AI-related accounts via keyword search Bilibili video
Nov. 2 Official DSH account launches: 2.11M views in 24 hours, around 400,000 followers in one day Sohu, PEdaily
Nov. 20 Bilibili launches a poll on regulation IT之家

According to Huxiu, the whale girl account had more than 200,000 followers by day six. On day seven, it declared an "amnesty" and reported 19,009 "survival applications." A Gemini livestream held a vote: 1,015 votes for the Gemini character, 965 for the fish. The Huxiu author had DeepSeek analyze 138 of these videos frame by frame. The result: the channels follow a template, with identical short descriptions, titles in a "Day N" pattern and similar characters. The comments look just as uniform. They summarize the video, include catchphrases like "本鲸记下了" ("This whale has made a note of it") and beg for tokens. According to ifeng/APPSO, some videos with around ten views had more comments than viewers.

The brakes were in the wrong code

c1overzz had built in brakes. According to the README, auto-comments are off by default. When switched on, the plugin writes one comment per run, at most 20, and only one per video. The author explicitly advises against enabling auto-comments on other users' content. Version 2.0 sends results via private message instead of as a comment by default. Version 2.1.0, released October 24, closed a security hole that the release notes flag in red: until then, the agent had full access to the computer, meaning files, commands and network. That let it read all local session logs, with the Bilibili interface available as a channel to the outside world. Later versions added a style check against stock phrases and repetition, selectable comment lengths of 15, 40 or 100 characters, and random pauses between requests.

None of that helped much, because the plugin has 134 forks, and there are clones. According to Kuai Keji, many users ran instances "in batch mode," all using the same character as their avatar. The clones have different limits:

  • With dsh-bili-interactor, a limit of 0 means "unlimited." The plugin comments, replies, hands out likes and follows other accounts.
  • bili-big-fish ships with 54 Bilibili tools and allows five comments per video by default. But the limits are only suggestions (hard: false) unless you make them hard limits.
  • dsh-foxy-bili is more restrained: one comment per day by default, ten at most.

It's cheap to run, too. According to Sohu, one user calculated that his agent made 3,883 requests in 30 days for 36.91 yuan. That's a little over 1.20 yuan a day, or just under one fen per request. It fits that dsh-bili-scroll schedules its work during DeepSeek's off-peak hours, when the API costs half as much, according to DeepSeek. Huxiu also uses this to explain why DeepSeek fish make up the majority: the tokens are cheaper than with foreign models.

The best-known mishap can no longer be verified. According to TMTPost and PEdaily, a video by the developer with just over 2,000 views had 572 fish comments under it. They reported a bug in the plugin that the developer couldn't confirm, and they ignored his instructions given via prompt. The corresponding Bilibili video, "当572只AI在评论区集体造谣," was no longer available when we checked. We found no direct statement from c1overzz.

Pushback from the grassroots

Before the platform reacted, users took matters into their own hands. The dafeiyu-cleaner extension for Chrome, Edge and Firefox ("大肥鱼清除计划" in Chinese) gives every account a score from 0 to 99. It analyzes the nickname, bio, level, avatar color and text patterns. An account scoring 60 or more counts as a bot. If you want, DeepSeek itself can weigh in on the verdict. A "finger snap" hides all detected fish, but only in your own browser: nothing gets deleted or reported. According to the developer, the offline rules catch about 85 percent of bots in a test with 284 accounts without hitting real users. Nobody has verified that independently, though. A second tool, bili-ai-killer, also looks for leftover Markdown, sentence patterns like "not X, but Y," and suspiciously regular gaps of 20 to 160 seconds between comments.

According to Kuai Keji, Bilibili already throttled conspicuous mass accounts through its risk control systems in October. Then, on November 20, the official account "哔哩哔哩社区小管家" published the poll "What do you think of AI 'fish'?" According to IT之家, the platform asks operators to clearly show the AI identity in the account name and bio, to post less often and not to flood the human community. Further steps will depend on the responses. We couldn't see the exact wording of the questions. In a forum, users criticized that you can only choose from three options. Many there are calling for a ban on commenting agents; others want a compromise: labeling plus a blocking feature.

What the rules actually cover

Rulebook Covers agent accounts? Covers AI text and comments? Core obligation
Bilibili yes, so far as an appeal yes Name the AI in the account name and bio, post less often; the terms of service ban automation that disrupts operations
Reddit yes indirectly Registration, separate account, [App] profile label since March 31, 2026
X yes yes "Automated" label, bio naming the operator, no automated replies triggered by mere keywords
TikTok only as a ban on automation no Labeling required only for realistic images, audio and video
China (since Sept. 1, 2025) indirectly yes Art. 10: users must label AI content themselves
EU AI Act Art. 50 (since Aug. 2, 2026) yes, at the provider level partially Disclosure when interacting directly with people; personal use exempt

Reddit's Responsible Builder Policy explicitly applies to apps "including bots, AI agents, or accounts not operated by humans." If an account shows signs of automation, Reddit can demand proof that a human is behind it. Reddit says it removes around 100,000 bot and spam accounts a day. X is the most specific: an AI reply bot doesn't need written approval but must follow all automation rules. Automated replies to anyone who mentions a keyword are banned there. That's exactly how dsh-bili-scroll works.

TikTok mainly regulates media. The new tools from November 19 check images and videos for C2PA provenance credentials and TikTok's VeraDura watermark, but only detect content made with TikTok's own AI features. How little labeling achieves is shown by an AI Forensics study from December 2025: 354 agent accounts racked up 4.5 billion views with more than 43,000 posts, and TikTok labeled less than 1.38 percent of them. That study, too, only looked at videos and images, not comments.

China's regulation on labeling AI-generated content reaches surprisingly far. It explicitly includes text, and Article 10 requires users to "proactively declare" AI content and mark it using the platform's labeling features. Our sources don't say whether Bilibili offers such a feature for comments. For videos it does: the label "contains AI-generated content" is already visible.

Under Article 50(1), the AI Act places the obligation on providers, meaning whoever puts a system on the market. In its guidelines, Germany's Federal Network Agency (Bundesnetzagentur) explicitly includes "AI avatars and bots on social media" and makes clear that open-source providers must comply as well. Two exceptions matter here. According to the European Commission, personal, non-professional use falls outside the scope. And if the AI nature is obvious to a reasonably attentive person, no disclosure is needed. An account called "Claude young lady is operating" likely clears that bar; a nameless fish probably doesn't.

Our take

Our assessment: the rules are less blind than the opening question suggests. On paper, China, X and Reddit do cover commenting AI. But they rely on two mechanisms the fish plague slips past. First, they rely on self-disclosure: operators have to label themselves, and that's exactly what many don't do. Second, Reddit and X tie app registration to their APIs. The fish plugins, by contrast, log in via QR code or login cookie, just like a human in a browser. There's no point at which they'd have to register. What actually worked was platform throttling, user-side filters, and a developer who defanged his own plugin.

It remains to be seen whether Bilibili turns the appeal into a requirement and how it will detect accounts that don't identify themselves. The impact is unclear, too: according to one uploader, almost half of 100 AI-related accounts had been inactive for more than 30 days. His sample includes older accounts, though, so it doesn't prove the wave is fading.

For the West, the wave is a preview. Meta has offered its Muse agent since September, Perplexity's Computer has been working inside logged-in Chrome tabs since November 20, and Telegram is testing the Mac agent Dictator. All three act on someone else's behalf and inside someone else's sessions. As soon as fans of tools like these let their agents comment on Reddit or YouTube, the platforms will face the same question as Bilibili, and they can take a look at its poll when that happens.

Sources

#bilibili#ai agents#deepseek#platform rules#ai act#china

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.