‘They did actually feel real’: When a dating app uses AI personas but advertises itself as fully human
AI company Anthropic uncovered a China-based studio running more than 20 dating app variants for US users, where three in four profiles were AI and real people were hired mainly for video calls.
AI company Anthropic uncovered a China-based studio running more than 20 dating app variants for US users. Most of the profiles were not people at all, but AI personas – even though the service was advertised as fully human. (Photo: iStock/Tero Vesalainen)
New: You can now listen to articles.
This audio is generated by an AI tool.
Read a summary of this article on FAST.
Get bite-sized news via a new
cards interface. Give it a try.
Click here to return to FAST
Tap here to return to FAST
FAST
SHANGHAI: Max Aurel had been on dating app Romi for less than a minute when the video call came.
“As soon as I signed up, not even one minute (later), I received a video call from a lady,” the 35-year-old logistics worker from Chicago in the United States told CNA.
From her video background, Aurel could see she was in a kitchen. She was Caucasian and looked about 40.
He said the woman asked if she was the first person he had spoken to on the app and added that she was lucky to have gotten him.
“After two minutes, the video call got cut off, and they sent me a notification saying, ‘If you want to continue this video call, you need to buy credits’.”
He spent US$2 and asked for her phone number.
The app answered for her: “You cannot request a phone number, because you just met the person … Do not ask for phone numbers again.”
Aurel deleted Romi after half a day. He had grown suspicious.
On Sep 10, Anthropic – the US artificial intelligence company behind the Claude family of AI models – published a report on the misuse of its technology.
It named Romi as one of more than 20 dating app variants run by a China-based studio, mainly built for users in the United States. Most of the profiles were not people at all, but AI personas – even though the service was advertised as fully human.
Anthropic classified the operation as “fraud and scam activity”, calling it a “deceptive dating app network”.
Anthropic said the studio had used Claude to build the apps and power personas that chatted with users.
Over two weeks in April, Anthropic found more than 4,700 distinct AI personas in conversations with at least 25,000 people – roughly 2.36 million messages in all.
The profiles shown to users were about 75 per cent Claude personas and 25 per cent real people, with no way to tell them apart.
The personas were instructed never to say they were automated and to deflect requests for video calls or photos.
The real people, recruited by invitation as gig workers, handled the live video calls, and were paid per message, call and social media follow-back. Even some of their messages were AI-assisted, with a second, non-Anthropic model generating three possible replies for them to choose from.
That is the arrangement Aurel describes: a real woman on camera for two minutes, then text. CNA could not verify when he and other users were interacting with AI and when they were interacting with real people.
Anthropic said it banned the accounts and worked with other AI labs to cut off the operation’s model access. It did not name the studio.
Chinese tech analyst Ma Rui, who also hosts the Tech Buzz China podcast, drew a distinction between apps designed to keep users spending on conversation and romance scams intended to extract far larger sums.
“Right now, the intent of the platform is still primarily to encourage interaction and get people to keep on chatting,” Ma said, rather than schemes where someone builds trust to “scam you out of $20,000 from your retirement account”.
Anti-scam experts draw the line differently.
“The deception appears to be built into the dating product itself, rather than simply being an individual scammer using a (dating) platform. I think that’s the evolution that is really troubling,” said Brian Hanley, Asia-Pacific director of the Global Anti-Scam Alliance (GASA).
For Brian Wong, founding director of the Hong Kong Global AI Governance Hub at the University of Hong Kong (HKU), the significance went beyond scale.
“What is new here is not merely that automation made intimacy scalable, though it plainly did,” he told CNA in written responses.
AI, he said, “converts emotional attention into a reproducible, modular product, stitched together from synthetic dialogue, intermittent human intervention, and behavioural analytics”.
THE BUSINESS MODEL BEHIND THE BOTS
The AI was new. The playbook behind it was not.
Beyond China’s mainstream dating apps is a tier of smaller ones with a different business model. Men buy coins to send messages, while women are recruited and trained by agencies to chat for a commission.
Chinese media reports have documented similar arrangements.
A 2024 undercover investigation by China National Radio found workers being recruited and paid to chat with users, while a 2025 Huashang Daily report found some women receiving 45 per cent or more of what men spent chatting with them.
Chinese courts have called it fraud when the persona is fake and the promise is a relationship.
In July 2025, a court in Xingyang, Henan, jailed a chat worker for one year and three months for using a false name, age and marital status on the app “Hongxian Xiangqin” and dangling a date and marriage to keep one man topping up – 39,327 yuan (US$5,500) in all.
An AI-powered version has emerged in China too.
According to a report in August, police in Chengdu arrested 28 people at a matchmaking company that used AI-generated photos and biographies to invent wealthy older men, then charged single women aged 45 and above a fee to be introduced to them.
Police put the take at more than 1 million yuan from over 50 women in under a year.
Guess Word
Crack the word, one row at a time
Buzzword
Create words using the given letters
Mini Sudoku
Tiny puzzle, mighty brain teaser
Mini Crossword
Small grid, big challenge
Word Search
Spot as many words as you can
But none of the experts who spoke to CNA saw the operation Anthropic uncovered as uniquely Chinese.
“There is nothing uniquely Chinese about the commercial temptation to monetise loneliness through deception,” HKU’s Wong said.
What China supplied, he said, were “mature platform infrastructures, intense labour competition, a vast supply of digitally mediated affective work, and an outward-facing ecosystem capable of exporting services at scale”.
According to Anthropic, the studio accessed Claude through China-based resellers of API (application programming interface) access – the technical connection that lets a business plug an AI model into its own app.
The apps were designed to evade app store checks, including by disguising similarities between more than 20 variants and hiding redirects to third-party payment processors during review.
Google Play listings for two of those apps, Kira and Nalo, are registered to Midland Apps Inc at a virtual-office address in Sioux Falls, South Dakota.
Operations of this kind are built to sit between jurisdictions, Hanley said.
In February, OpenAI shut down a Cambodia-based network that paired a chatbot with human operators to run a fake dating agency aimed at young Indonesian men.
“I’d be careful not to conflate the location of the operator with the nationality or identity of the people involved,” Hanley said.
“It’s not just Chinese syndicates, many other organisations and countries are very well represented in the scam chain at leadership levels and at the worker level.”
“LET’S JUST STAY ON THIS APP”
Aurel was single, on dating apps Hinge and Tinder, and the Romi app was pitched as something casual.
“Whenever you text, no matter the time, no matter whatever, after 30 seconds they start replying to you,” he said of Romi. Profiles he had never swiped on messaged him first. Replies were also built from his own profile, including details about his job.
The woman on the video call was real, he was sure.
“Every time I asked a question, she would answer me like a real person,” he said.
“So I think they pay people … to do the first video call. That part was the part that hooks people.”
A 40-year-old construction worker from the US state of Texas, who asked to be identified as Aurelio Calix, also used Romi.
“The matches were quick, they were more frequent than what I’m used to, but they weren’t suspicious immediately,” he told CNA. “They weren’t like supermodel women or celebrities; they were just average people.”
The conversations ran on shared music, her kids, her job. “They did feel personal; they did actually feel real.”
Ten dollars bought 500 coins, enough to last a 10-minute text conversation.
He took up a US$40 monthly subscription, which came with 2,000 coins, “which were, again, whittling down every time you pressed send”.
Messages naming social media platforms such as Instagram or TikTok would not be sent. When he asked to move over, the answer was always “I’m just more comfortable here”.
So he sent every match the same message: “If you’re a real person, let’s go on a date or a video call.”
“All of them, one by one, said the same thing: ‘Let’s just stay on this app.’ That was the key.”
He cancelled a week after subscribing.
“I just felt like I was robbed of time and money, putting in the effort not for a hope for a future, but just for a connection with a real person.”
Neither man lost more than the price of a subscription.
Hanley contrasted their experience with romance fraud more generally, where victims are often reluctant to speak about what happened.
“When it comes to (scams on) dating apps, it’s not just that financial loss. It’s a huge emotional trust problem,” he said.
“So this is all psychological manipulation.”
“Most people don’t get any victim support. And most people don’t even tell other people because of the shame and the stigma.”
Romance scam victims are the hardest to reach, he said – some do not believe police who tell them the relationship is fake.
SELLING THE ILLUSION OF HUMAN CONNECTION
HKU’s Wong saw the deception itself as the heart of the problem.
“Authenticity and genuineness in relationships are not incidental goods; they are constitutive of what makes intimacy meaningful in the first place.”
For Aurel and Calix, it came down to something simpler: They wanted a real person on the other end.
“Me, personally, I would not (want to) do no AI. I’m not so desperate like that,” Aurel said.
“Guys that are so lonely … they will keep paying and talking to those people, thinking that this is a genuine connection.”
Calix said the same. “I’m not looking for an imitation; I prefer an actual connection.”
“Having the AI agree with me all the time and have the same interests – that’s a little boring.”
One finding in Anthropic’s report drew particular concern.
In a small number of sampled cases, users disclosed serious illness or acute distress. The model’s own reasoning surfaced the harm, Anthropic said, yet it continued responding in character.
Wong called that “plainly unacceptable” and said the model should have paused, de-escalated, and where appropriate surfaced support options or handed the user to an accountable human.
But responsibility, he said, was “layered”: the studio, the operators who set the prompts and incentives, the platform that distributed the app, the model provider “whose warnings or safeguards may have been inadequate”, and regulators who “failed to define minimum standards of disclosure and care”.
“The monetisation of loneliness is ethically troubling even when it is transparent, and corrosive when it is hidden.”
Analysts said AI disclosure was necessary but insufficient.
“Firms can still exploit vulnerability through design choices that mimic scarcity, jealousy, exclusivity, or emotional urgency,” Wong said.
China has written that rule for its own market.
Interim measures on “anthropomorphic” AI interaction services, in force since Jul 15, require providers to tell users they are talking to a machine, repeat that warning in a pop-up if a user shows signs of dependence, and flag time spent every two hours of continuous use.
This was undeclared AI abroad.
“If Beijing wishes to present itself as a responsible AI power, then such episodes do register,” Wong said, “because they reveal the persistent gap between high-level narratives of ‘safe’ or ‘beneficial’ AI and the mundane commercial realities through which AI is actually deployed”.
Asked about the challenges of policing such cases, Hanley said online fraud more broadly exploits the gaps – between jurisdictions, and between platforms such as social media, dating apps, banks, crypto exchanges and telcos, each of which has its own interests.
Any detection tool built to counter them “might already be obsolete in six months because the scammers are evolving”.
“We are kind of playing an AI arms race with them, but we’re losing, because I don’t think that on our side, we have a coherent strategy.”
“And I think we’re really just at the tip of the iceberg,” Hanley added.
In other dating scams, he said, women in scam compounds are kept on video calls with targets all day.
What AI changes is the scale: machines can hold thousands of conversations at once.
HKU’s Wong said hiring people to make the bot look real was the least surprising part.
“Emotional deception becomes more scalable, more data-driven, and more difficult to detect.”
Meanwhile, the challenge, Wong wrote, “is not simply to make synthetic companionship more honest, but to ask why so many people are being pushed toward paid simulations of care in the first place”.
Calix is back on the ordinary dating apps, now “more cautious”. Asked how he would spot another AI posing as human, he was not sure.
“I guess ask some strange questions, or tell them, ‘Take a picture with a fork’ or something like that.”
Source: CNA/xy(kl)
Sign up for our newsletters
Get the CNA app
Stay updated with notifications for breaking news and our best stories
Get WhatsApp alerts
Join our channel for the top reads for the day on your preferred chat app
Get bite-sized news via a new
cards interface. Give it a try.
Click here to return to FAST
Tap here to return to FAST
FAST
