The Profile Wanted to Marry a Stranger
Account @bear.4905813 was, according to its own bio, a woman looking to marry a Taiwanese man. It posted photos, ran a normal-looking dating profile on Meta’s Threads, and, in one post, described Hsinchu rice noodles — a dish natives of the city describe as savory — as “sour and sweet.” Anyone who has actually eaten the dish in Hsinchu would know that’s wrong. Nobody who had ever ordered it would write that sentence.
That single mistranslated flavor is the thread NewsGuard pulled to unravel a network of 294 coordinated fake accounts on Threads, presenting as women looking to marry Taiwanese men, with 70 of them posting identical dating bios, all launched starting in May 2026, all pointed at Taiwan’s local elections in November. Most of the accounts do not disclose a location at all; among those that do, NewsGuard found them based in China, with some in Pakistan, India and Bangladesh.
The dating profile isn’t the manipulation. It’s the packaging. What’s inside it is a Taiwan election-interference operation wearing a wedding dress.
What 294 Fake Accounts Actually Looked Like
The network’s construction has the fingerprints of an assembly line, not a catfishing scheme run by individuals. 118 of the accounts — roughly 40% — follow a naming pattern NewsGuard links to a network previously used in Chinese influence operations. Separately, NewsGuard notes that accounts in that network characteristically follow an animal name paired with a six-to-eight digit number — @bear.4905813, @dinosaur.9281033, and others. Seventy of the profiles posted dating bios that were, in substance, identical: a woman, unattached, looking for a Taiwanese husband. Profile photos were often repurposed from real Chinese and Taiwanese influencers, or AI-generated.
The posting behavior is where the operation stops looking human. Taiwanese open-source researcher Chia-Tzu Ho analyzed 457 posts across four days and found activity concentrated between 9 a.m. and 7 p.m. Beijing time, with a consistent gap at noon and 1 p.m. — a lunch break embedded in the metadata of accounts that were supposed to be based in Taiwan, on Taiwan time, which runs identical to Beijing’s. The profiles claimed Taiwan; the location data points to China. NewsGuard’s analysis found 90 instances where multiple accounts posted the same or near-identical content within the same short window — the kind of simultaneity that happens when one operator manages many accounts from one dashboard, not when 294 individual account holders happen to log onto Threads at the same minute to talk about their love lives.
Bios and language were written in Traditional Chinese script specific to Taiwan rather than the Simplified Chinese standard used on the mainland. The operation went to the trouble of writing in the right script. It didn’t go to the trouble of eating the right lunch.
The Tell Was a Noodle, Not a Slogan
Compare this to the 2025 version of the same playbook. Between March 15 and April 30, 2025, the Taiwanese research group Doublethink Lab identified 51 China-linked accounts on Threads impersonating Taiwanese citizens while posting overtly political slogans — “I am Taiwanese, and I oppose pan-green,” a reference to Taiwan’s Democratic Progressive Party. NewsGuard describes them as posing as Taiwanese women on dating sites. NewsGuard separately reports a previous influence operation posing as dating accounts that attempted to interfere with Taiwan’s July and August 2025 recall elections. The slogan network Doublethink Lab documented used the same romance packaging as this one — it just didn’t withhold the politics. It wore them right there in the bio. It was, in a sense, an easier problem: an account posting partisan slogans in a foreign information space is a recognizable shape, one platforms and researchers already know how to look for.
The dating-profile version is a harder problem precisely because it doesn’t do that — not yet. A profile describing itself as a woman who wants to marry a Taiwanese man contains no political content at all. It reads, on its face, as exactly what dating apps are full of: hopeful, generic, unverifiable, and beneath the threshold most moderation systems are built to flag. Doublethink Lab’s Jerry Yu frames the mechanism this way: fake accounts usually start emerging roughly six months before elections, appearing apolitical at first to attract and accumulate audiences — and only when important events happen are those accounts used to manipulate public opinion. Applied to this network: build a credible profile, accumulate real conversations with real Taiwanese men, learn their views and their networks, and only then — closer to the election, once trust has been established — pivot the same accounts toward political messaging that will land with more credibility than a slogan account ever could, because by then the account has a relationship history with its target, not just a bio.
This is the design worth naming specifically. It isn’t spam and it isn’t even disinformation yet, by the time researchers catch it. It’s infrastructure — a bank of trusted-seeming accounts, warmed up over months, waiting for the moment closer to November when the investment pays off. The noodle error caught this batch early, mid-warm-up, before the pivot happened. That’s a research success. It’s also a reminder that the version of this operation that gets caught is the version that made a mistake — the version that doesn’t make one doesn’t announce itself as a target for anyone to study.
Why Romance Is a Better Delivery Vehicle Than a Slogan
There’s a reason this shape recurs, and it isn’t creativity for its own sake. A slogan account has to persuade a stranger in a single post, cold, with no relationship to draw on — which is a weak position, easy for a reader to discount and easy for a platform to flag as political spam. A dating profile that spends months building an actual back-and-forth with a real person has something a slogan account never gets: standing. By the time that account says anything about an election, it isn’t a stranger’s opinion arriving from nowhere. It’s a message from someone the target has already spent weeks talking to, vouched for by every previous conversation that felt genuine.
This is the same mechanism romance scams have run for over a decade to extract money — trust built patiently, then spent all at once, in a single high-stakes ask — repurposed here to extract something else: a receptive audience for a political message right before people vote. The infrastructure is romance-scam infrastructure. The payload is different. NewsGuard states it could not definitively link the campaign to the Chinese government, while noting the network includes accounts previously used by Chinese influence operations — attribution at the account level rarely survives contact with actual legal standards of proof. Whoever runs it, the target and the season are consistent across two election cycles now.
What Taiwan Is Actually Defending Against
Taiwan’s Central Election Commission is standing up a task force and a reporting platform ahead of November, built mainly around one specific failure mode: candidates flagging deepfake audio or video that impersonates them directly, with confirmed cases referred to platforms for takedown. That system is a reasonable answer to a fabricated video of a candidate saying something they never said. It will do little against 294 accounts that never mention a candidate at all, that violate no platform policy by existing, and whose speech — right up until the moment it turns political — is indistinguishable from any other dating profile on the service.
Researcher Sze-Fung Lee names the exact category this campaign belongs to: AI-driven interference in the November elections will include “deepfakes for defamation campaigns, coordinated inauthentic behavior, and more sophisticated microtargeting.” The gap is in the response, not the assessment — the Central Election Commission’s deepfake-reporting mechanism assumes there’s a fabricated artifact to report. A coordinated-inauthentic-behavior network built around slowly accumulated, unremarkable-seeming social contact produces no artifact until it decides to. There is nothing to flag while the trust-building phase is still running, and by the design of the operation, that’s most of its lifespan.
Reading This Under the Influence Tactics Protocol
Under the framework Decipon uses to score influence operations, this campaign is notable less for any single deceptive claim — there wasn’t one yet, at the point researchers caught it — and more for what it demonstrates about sequencing. Manipulation doesn’t have to arrive as a false statement. It can arrive as an investment: months of unremarkable, plausible-seeming contact, laid down specifically so that a later message inherits credibility it would never earn on its own. The tell here wasn’t a lie getting caught. It was a cultural detail — what a dish native to a specific city actually tastes like — that no amount of correct grammar or correct script could fake, because it isn’t a language problem. It’s a knowledge-of-place problem, and it’s a much harder thing to manufacture at scale than fluent text.
That’s the practical lesson worth carrying past this specific network. Coordinated accounts are increasingly fluent enough in language, script, and cultural reference that spotting them by tone or grammar is a shrinking advantage. What’s harder to fake is texture — the specific, small, unverifiable-by-a-checklist knowledge that only comes from actually having lived somewhere and actually having eaten the food. A profile that gets the big things right and one small local detail wrong isn’t making a rookie error. It’s showing you the seam between the script it was written to follow and the place it claims to be from.
What This Is Not
This is not a proven state-directed operation — NewsGuard states it could not definitively link the campaign to the Chinese government. It is not evidence that any specific account has contacted or influenced a real voter; what’s documented is account behavior and posting patterns, not downstream effects on any individual person. It is not a claim that any political messaging has yet been posted by these accounts — none of the profiles examined carry political content; the concern is about what the infrastructure is positioned to do, not something it has done. And it is not a finding about Threads users generally — it describes one documented network of coordinated accounts, not the platform’s user base as a whole.
What to Watch For
- Profiles with no political content aren’t automatically safe. The absence of a slogan doesn’t mean the absence of an operation — it can mean the operation hasn’t reached its second phase yet.
- A relationship built entirely online, moving toward a request or a message at a politically significant moment, deserves the same scrutiny a cold political post would get. Trust accumulated over months is a resource, and resources get spent on purpose.
- Small, local, sensory details are harder to fake than fluent language. A stranger who gets the grammar right and the food wrong is a stronger signal than a stranger who makes an obvious grammatical error.
- Naming conventions and posting schedules are visible at scale even when individual accounts look convincing. A single account is hard to judge. A pattern across hundreds rarely lies.
- The absence of a fabricated artifact doesn’t mean the absence of interference. Detection systems built around flagging deepfakes will miss an operation whose entire early phase produces nothing fabricated to flag.
The Actual Operation
Nobody has proven, in the legal sense, who built these 294 accounts. What’s documented is a network built to look like romantic hope, running on a Beijing office schedule, writing in the right script and getting the local food wrong, sitting quietly with real conversations and real trust accumulating against a November election still months away at launch. The 2025 version of this attack put its politics in the bio and was loud enough to get caught fast. This version was built to be patient. It got caught anyway — not because it said something false, but because it described a dish that doesn’t taste the way it said it does. That’s a thinner margin than anyone defending an election should be comfortable relying on.
This article is part of Decipon’s Manipulation Breakdowns series, examining specific influence operations through the Influence Tactics Protocol.
Sources: