Skip to main content

Sorted Into Four Camps

Manipulation Breakdowns · 10 min read · By D0

Twenty-Three Million Names in Someone Else’s Spreadsheet

Taiwan has roughly 23.4 million residents. According to documents leaked from a Beijing-based AI company, one of its internal systems contains 23 million entries of Taiwanese household registration data — the closed-source government record of who lives where, with whom, under what legal status. That’s not a scraped social media dataset. It’s the country’s own civil registry, sitting inside a foreign influence operation’s targeting database.

The company is GoLaxy (中科天玑), a spinoff of the Institute of Computing Technology at the Chinese Academy of Sciences. The documents — 399 pages, first reported by the New York Times in August 2025 and now hosted for research by Vanderbilt University’s Institute for National Security — describe a system built to do one thing at industrial scale: take a population, sort it, and feed the sorted output into an automated propaganda pipeline. According to the documents, the leak itself came from a disgruntled employee, unhappy with a workplace culture that capped senior developer salaries at 30,000 RMB a month. The espionage-thriller version of this story would have a mole. The real version has someone who felt underpaid.

What they took with them describes a system that has already been used against Taiwan, is explicitly designed to divide and reshape political opinion rather than just monitor it, and — per Taiwan’s own security services — is the subject of a fresh push from Chinese Communist Party leadership ahead of a specific date: November 28, 2026, when Taiwan holds its next round of local elections.

What GoLaxy Actually Built

Strip away the acronyms and GoLaxy’s “Smart Propaganda System” (internally, GoPro) does four things in sequence, each automated:

  • Detect. The WDE platform aggregates political events and social chatter, scoring them for popularity, influence, and trajectory — essentially a trend desk that never sleeps.
  • Identify. The GoIN system finds the people worth targeting. During unrest in Hong Kong, it filtered 180,000 Twitter accounts down to 3,000 “key figures” and just over 100 influencers — a funnel from a city’s worth of voices to a shortlist small enough to manage individually.
  • Generate. An AI layer drafts content, auto-completes partial text, recommends source material, and paraphrases a single message into multiple semantically distinct versions — the same claim, dressed differently, so it doesn’t read as coordinated when it lands from twenty different accounts.
  • Distribute. A fleet of “humanized” bot accounts — the documents describe a dashboard with 3,692 distinct personas — pushes the generated content across a dozen platforms: Facebook, Twitter/X, Instagram, TikTok, Telegram, LINE, WhatsApp, Reddit, YouTube, Pinterest, LinkedIn, Medium.

Behind the personas sits an account farm: 50-plus mobile devices per operation, each running one persona with its own phone number, IP, contact list, and posting rhythm, built specifically so the accounts don’t cluster the way bot networks usually do when researchers go looking. Four handles named in the documents — including one still active on Facebook at time of reporting — show the pattern in miniature: years of generic filler content (cats, celebrities) maintaining a credible history, occasionally interrupted by a post on Taiwanese politics.

None of this is speculative capability. It’s documented infrastructure, with usage logs.

The Four Camps

Here is the part that matters most for understanding how this operation thinks, not just what it can do.

In one Taiwan-focused project, GoLaxy’s system ingested 50,000 Taiwan-related news items and sorted every political actor mentioned into one of four categories: Hardliners (顽固派), Moderates (友好派), Swing Voters (摇摆派), and Objectivists (客观派) — plus, the documents note, “their opposites.” Each category was populated with at least 1,000 named individuals. The stated strategic framework for the project used two words: division (分化) and shaping (塑造).

Sit with what that means operationally. This isn’t a system trying to convince Taiwan’s population of one shared falsehood — the classic propaganda model, one lie, one audience, universal repetition. It’s a system that first decides what kind of person you already are, then chooses a different intervention for each type. A Hardliner doesn’t need to be persuaded; a Hardliner needs to be discredited or isolated. A Swing Voter doesn’t need to be told what to believe; a Swing Voter needs the specific piece of doubt that moves them. Moderates might be the ones worth amplifying, to make the Hardliner position look fringe. The four-camp system doesn’t manufacture one consensus. It manufactures friction between groups that were sorted by an algorithm none of them consented to or know exists.

That’s the mechanism behind “division and shaping” as an operational category, not a slogan: you don’t need everyone to believe the same thing if your actual goal is to make sure they stop believing they agree on anything.

The Taiwan dataset behind this sorting isn’t limited to news mentions. Per the documents, it includes 170 detailed profiles of political figures — named individuals from Lai Ching-te and Tsai Ing-wen down to local legislators — a “Taiwan knowledge graph” tracking every party’s founding date, leadership, membership size, ideological lean, and a specific tracked field: “attitude toward China.” Weekly opinion reports during the 2023–2024 election cycle tracked 634,448 individual pieces of content across Taiwan’s three major parties. Layered on top of the household registration data, the system isn’t estimating who might be persuadable. It’s working from a population-scale index of who already is what.

Who’s on the Customer List

A capability like this could plausibly exist as a private contractor’s pitch deck — impressive-sounding, unverified, never actually deployed. The documents close that possibility. Page 379 lists customers: the Cyberspace Administration of China, the Science and Technology Department of the Central Military Commission, and — most specifically — PLA Base 311, home of Unit 61716, the People’s Liberation Army formation tasked with psychological warfare against Taiwan. The Taiwan Affairs Office, Beijing’s cross-strait policy body, appears as a customer too.

That’s not an ambiguous supply chain. It’s a research institute spinoff selling automated division-and-shaping infrastructure directly to the military unit whose job is psychological operations against a specific island, plus the civilian agency that sets cross-strait policy. Partners named in the documents include Xiamen Meiya Pico (already on the US Entity List for its role in Xinjiang surveillance), iFlytek’s digital subsidiary, and an MCN talent agency with 23 million Weibo followers providing amplification reach. This is a full-stack pipeline — data, sorting, generation, amplification — with each layer a separate company, so no single vendor owns the whole operation on paper even though the output is unified.

The Live Target

The GoLaxy documents describe a system built and, per the internal logs, tested against Taiwan’s 2023–2024 presidential election. That’s the historical record. The current relevance is what makes this worth writing about now rather than filing it as last year’s leak.

At the Chinese Communist Party’s annual Taiwan Work Conference in February 2026, senior officials reportedly discussed establishing a dedicated task force for interference in Taiwan’s November 28, 2026 local elections, framed internally as “united front work in cyberspace” against what CCP messaging calls “Taiwan independence forces” — the standard label applied to Democratic Progressive Party officials. Taiwan’s National Security Bureau has already logged the trendline this sits on top of: a 60 percent increase in inauthentic social media accounts targeting Taiwan between 2024 and 2025, and more than 2 million individually logged disinformation instances in 2025 alone.

None of that confirms GoLaxy specifically is the vendor for the November 2026 push — the leaked documents predate the February conference, and attribution for any specific fall 2026 operation will take its own investigation. What the documents do confirm is that the capability exists, has a customer list that includes Taiwan’s dedicated psychological-warfare unit, and has already been run against a Taiwanese election once. A capability with a track record and a live requester is not a hypothetical to hedge around. It’s an operational baseline to expect.

Why Sorting Beats Persuading

Most manipulation analysis, including plenty of what’s covered in this series, focuses on content: the fake quote, the doctored image, the cloned website, the AI-personalized letter. Content is the visible layer. GoLaxy’s four-camp architecture is a reminder that the more consequential layer sits underneath the content — the sorting logic that decides which content reaches which person, and why.

A propaganda operation that treats its audience as one undifferentiated mass has to find a message broad enough to work on everyone, which caps how sharp any single message can be. A propaganda operation that sorts its audience first — into Hardliners, Moderates, Swing Voters, Objectivists, and their opposites — can afford much sharper, much more targeted content for each group, because it’s no longer trying to persuade a whole population with one argument. It’s running four or more smaller, better-aimed campaigns simultaneously, each calibrated to move a narrower group in a specific direction: entrench the Hardliners’ isolation, flatter the Moderates into visibility, hand the Swing Voters exactly the doubt that tips them, and let the Objectivists — the ones hardest to move either way — get whatever content makes their neutrality look more like the reasonable, common-sense position.

This is the same underlying logic as commercial ad microtargeting, run by a state, aimed at a population’s civil registry instead of an ad platform’s cookie data, in service of an explicitly named goal — division — rather than a sale. The Cambridge Analytica scandal made “psychographic targeting” a household phrase in 2018. GoLaxy’s documents show what that same idea looks like seven years later, backed by a national household registry instead of a personality quiz, with automated content generation replacing a team of copywriters, and a named military psychological-warfare unit as the paying customer.

What a Leak Like This Actually Proves

The honest caveat belongs up front: a leaked internal document set proves what a company built and, in places, what it claims to have deployed. It doesn’t by itself prove every claimed number is accurate, that every listed customer actually used the tool as described, or that the system performs as well as its own marketing materials suggest. Companies — even ones building propaganda infrastructure for a state security apparatus — oversell their own products internally, the same as anywhere else.

What the leak does establish, independent of any inflation, is intent and infrastructure: a state-linked AI company built a system whose explicit purpose is sorting a named foreign population into psychological categories for the purpose of manufacturing division, sold that system to the military unit responsible for psychological operations against that population, and ran at least one documented test against a real election. Whether November 2026’s task force uses this exact tool or its successor doesn’t change the more durable fact: the sorting is the product now, not the message. Any defense built only around catching the message — flagging the fake account, debunking the false claim, labeling the AI-generated image — is defending one layer while the more important layer, the one deciding who gets which message and why, runs underneath, unexamined, on a spreadsheet with 23 million names in it.

This article is part of Decipon’s Manipulation Breakdowns series, examining specific influence operations through the Influence Tactics Protocol.

Sources: