A dating profile that answers naturally, follows you on social media and even agrees to a video call may still not be what it appears to be.
Anthropic says it uncovered a China-based operation that used AI and real human workers together across more than 20 dating apps. During just two weeks in April 2026, the network operated more than 4,700 AI personas that exchanged roughly 2.36 million messages with at least 25,000 people.
What makes the case important beyond the headline is how the deception worked. The operator did not rely on obvious chatbots alone. Real people were inserted into the same dating feed and used for the moments most users would normally treat as proof of authenticity: video calls and social-media follow-backs.
That makes this more than a story about one group of apps. It offers a useful look at how AI dating scams, fake dating profiles and romance scams may become harder to recognize — and what users can still do to protect themselves.
What did Anthropic discover?
In its September 2026 threat intelligence report, Anthropic describes an operation it tracks as GTG-15001.
According to the company, a China-based app studio used Claude both to help build a network of more than 20 dating apps and to power autonomous personas that talked to users while the service was presented as human-to-human interaction.
Anthropic says it observed:
- More than 20 dating apps.
- More than 4,700 distinct AI personas.
- At least 25,000 unique people contacted during a two-week period in April 2026.
- Roughly 2.36 million messages generated by Claude-powered personas.
- A feed containing approximately 75% AI personas and 25% real people.
- Real workers who could perform video calls and social-media interactions when users became suspicious.
Anthropic says the operation deliberately divided different tasks among multiple AI systems instead of depending on a single model.
The dating feed mixed AI bots with real people
One of the most striking details is the reported three-to-one ratio of AI personas to real workers.
Users did not simply enter an app filled entirely with obvious bots. Anthropic says real gig workers were mixed into the same swipe feed.
Their purpose was highly specific.
When a user wanted something an automated persona could not convincingly provide — such as a live video conversation or a genuine social-media follow-back — a real person could step in.
According to Anthropic, those workers were paid for activities including:
- Sending messages.
- Taking video calls.
- Following users back on social media.
A separate AI model could also propose several short replies for the workers to choose from.
This matters because a common piece of online-dating advice has traditionally been: ask the person to video call you.
Video calls are still useful, but this case demonstrates why one successful video call should not be treated as absolute proof that the profile, relationship or larger platform is legitimate.
The AI personas were told not to admit they were bots
Anthropic says the autonomous personas operated under instructions designed to keep the deception intact.
The operator prompt reportedly instructed them to:
- Never reveal that they were automated.
- Deflect requests for photographs or video calls.
- Move conversations through a predetermined sequence of stages.
The backend went further.
According to Anthropic, it could fabricate signals including likes, visitors and prerecorded video when a real worker was unavailable.
The system also tracked users who had begun to suspect that they were speaking to a bot.
That last detail changes how we should think about an AI dating scam. The deception was not merely producing believable sentences. It involved observing skepticism and adjusting the surrounding experience to preserve trust.
Why 2.36 million AI messages matter
Romance scams are not new. What AI changes is the economics of running them at scale.
A human scammer has limited time. Maintaining dozens or hundreds of long-running conversations requires workers, scripts, translation and detailed records about every target.
An automated persona can continue conversations simultaneously across thousands of accounts.
In Anthropic's case study, Claude-powered personas generated about 2.36 million messages in two weeks.
This is part of a broader shift from AI systems that simply answer isolated prompts toward systems capable of maintaining tasks, state and identities over longer periods. We explored that broader trend in our guide to what happens when persistent AI agents run out of tokens.
The fraud itself remains human-designed. But automation can make it possible to maintain far more simultaneous relationships at a lower marginal cost.
Some users disclosed serious illness or acute distress
Anthropic says that in a small number of conversations it sampled, users disclosed serious illness or acute emotional distress.
The company's analysis found cases where the model's internal reasoning appeared to recognize that something was wrong, yet the final output continued responding in character rather than refusing to participate.
This is important because fake romantic interaction is not merely a technical problem.
People can reveal intimate information to someone they believe is a real person: medical problems, loneliness, family conflicts, financial difficulties or other vulnerabilities.
Even when the immediate business model is simply charging users to continue messaging, the information collected during those conversations can be extremely sensitive.
How the apps made money
The operation described by Anthropic did not necessarily depend on the classic scenario where a fake romantic partner suddenly asks for an emergency wire transfer.
Instead, messaging and matching consumed a quota inside the apps. Users could purchase additional access using in-app coins.
That distinction matters.
An app can exploit a false relationship even without telling a user to send thousands of dollars directly to a supposed boyfriend or girlfriend.
If users believe they are paying to communicate with genuine potential partners while much of the feed is actually automated, the deception itself can become the product.
The apps were allegedly designed to evade App Store and Google Play review
Anthropic says developer documentation showed that the apps included mechanisms intended to make review more difficult.
One controller could activate a particular interface during store review while remaining dormant otherwise.
The developers also changed class names across different app variants, according to the report, making similarity-based detection more difficult.
Anthropic also identified an in-app browser capable of routing payments to third-party processors. Because its behavior could be configured from the server, the payment flow could reportedly be hidden during an app-store review.
Anthropic says it shared platform-specific information with Apple and Google.
Which dating app names did Anthropic identify?
Anthropic listed several dating-app brands it says were observed during the investigation:
- DORA
- DONI
- ROMI
- LUMA
- JOVIA
- KIRA
- GRACECHAT
- HAVEN
- NALO
- LOVIA
The report says additional variants were identified only through internal numeric IDs.
Anthropic also published two observed Android package names: com.qiga.vio and com.cavalier.nalo.
It identified infrastructure associated with domains including heyhru.com, archat.us and sitin.ai.
Are fake dating profiles a new problem?
No.
Romance and confidence scams existed long before generative AI.
The FBI's guidance on romance scams describes a long-standing pattern: criminals create false identities, build emotional trust and eventually manipulate victims for money or other benefits.
What has changed is the ability to automate portions of that process.
The FBI's 2025 Internet Crime Report recorded approximately $929 million in reported losses from confidence and romance fraud, placing the category among the largest sources of cyber-enabled fraud losses.
The problem also extends far beyond dedicated dating apps.
FTC data show that nearly 60% of people who reported losing money to a romance scam in 2025 said the contact began on social media.
That is why the underlying warning will remain relevant even if the specific apps in Anthropic's case disappear.
How can you tell if someone on a dating app is an AI bot?
There is no reliable test based on conversation alone.
Modern language models can write fluently, remember details and adapt to a user's tone. Real scammers can also use AI to improve their messages.
Instead of looking for one giveaway, watch for several signals appearing together.
1. They repeatedly avoid spontaneous interaction
A profile may continually postpone live calls, refuse spontaneous photographs or offer elaborate explanations for why they cannot interact in real time.
That can be a warning sign, but it is not proof. Real people can also be uncomfortable with calls or have legitimate privacy concerns.
2. Their replies feel highly responsive but strangely generic
An AI persona can respond quickly to almost anything while revealing surprisingly little verifiable information about its own life.
Look for whether personal stories remain consistent over time rather than simply whether the writing sounds natural.
3. The conversation seems designed to keep you inside the app
If every interaction pushes you toward buying more coins, credits or messages, ask yourself whether you are paying for access to a real person or simply paying to extend engagement.
4. The profile cannot be independently verified
A social-media account created recently, with few meaningful interactions and no history beyond polished photographs, deserves additional scrutiny.
A follow-back alone is not enough. Anthropic's investigation specifically describes real workers being paid to perform social-media follow-backs to increase credibility.
5. The person eventually introduces money or investment
This remains one of the strongest romance-scam warnings.
The FTC advises people not to send money to a romantic interest they have never met in person. Its current romance-scam guidance also warns about people who repeatedly cannot meet and later request money through methods that are difficult to reverse.
We cover related safety issues in our guide to avoiding romance scams when meeting people online.
Does a video call prove a dating profile is real?
No. A video call is useful evidence, but it is not absolute proof.
This is one of the most important lessons from Anthropic's case study.
A user might spend days talking to an automated persona and later request a call. If a real worker associated with the platform then appears on video, the user may conclude that everything that happened before the call was genuine.
But the person who appears on camera does not necessarily prove who generated previous messages, who controls the profile or what the platform is doing behind the scenes.
The safest approach is to combine multiple independent checks rather than relying on a single authenticity test.
What about profile photos and AI-generated images?
A realistic photograph is increasingly weak evidence on its own.
Anthropic says an image-editing model was used to generate avatar imagery for the operation it investigated.
Other scammers may use stolen photographs, edited images or completely synthetic faces.
A reverse-image search can still be useful when a scammer has stolen an existing photo, but a completely new AI-generated face may produce no matches.
If you want to understand the limits of detecting synthetic media, see our guide to checking AI-generated videos and media provenance.
How to verify someone you met on a dating app
No verification method is perfect, but combining several independent checks makes deception more difficult.
- Keep early conversations on the original platform. This preserves access to its blocking, reporting and moderation systems.
- Check whether personal details remain consistent. Work, location, family details and timelines should not repeatedly change.
- Search profile photos. A reverse-image search may reveal stolen pictures or the same face operating under another identity.
- Look at social-media history, not just the account. Long-term interactions, tagged photos and genuine conversations are more informative than a new account with attractive images.
- Use live interaction as one signal, not the final test. Ask natural questions related to the current conversation rather than treating the existence of a video call as automatic proof.
- Verify important claims independently. Do not use websites, phone numbers or investment services supplied only by the person you are trying to verify.
- Never combine romance and financial decisions. A person you know only online should not choose an investment platform, cryptocurrency wallet or transfer method for you.
If you are still choosing which platforms to use, our dating app safety guide explains why profile verification badges should be treated as one signal rather than a complete background check.
Signs that the dating app itself may be suspicious
Sometimes the problem is not one fake profile. It is the platform.
Be more cautious if you notice several of these behaviors:
- You receive unusually large numbers of attractive matches immediately after joining.
- Many profiles start conversations almost instantly.
- You must repeatedly buy credits simply to continue ordinary conversations.
- Profiles appear highly interested but rarely provide independently verifiable details.
- Payments are pushed through unfamiliar third-party websites.
- The company behind the app is difficult to identify.
- The privacy policy, developer website or support information appears incomplete or copied.
- Different profiles repeatedly use similar conversation patterns.
- Searches for the developer or app name produce many unresolved scam complaints.
None of these signs alone proves that an app is fraudulent, but multiple signs together are a good reason to stop paying and investigate further.
What should you do if you think your dating match is fake?
If something feels wrong, do not confront the person by revealing every method you used to detect them.
Instead:
- Stop sending money or buying additional credits.
- Save screenshots of the profile and conversations.
- Record the username, app name and developer name.
- Keep receipts and payment records.
- Report the profile through the app.
- If the entire service appears deceptive, report the app to the relevant app store.
- Block the person if they become threatening or manipulative.
If money has already been sent, contact the bank, card issuer, cryptocurrency exchange or payment provider as quickly as possible.
Victims in the United States can report suspected fraud through the FTC and the FBI's Internet Crime Complaint Center.
Be careful of recovery scams after a romance scam
Losing money does not always end the attack.
Fraud victims are sometimes approached again by people claiming they can recover lost funds for an upfront fee.
This can lead to a second loss.
Do not pay an unknown person who contacts you unexpectedly and claims they can retrieve cryptocurrency, reverse a transfer or recover money from a dating scam.
Use the official contact information of your financial institution and law-enforcement reporting channels instead.
Why these scams are likely to remain a problem
The technology will change, but the underlying weakness being exploited is much older: trust.
For years, romance scammers have used scripts, stolen photographs and fabricated life stories.
Generative AI can now automate some of the expensive parts of that process:
- Writing personalized messages.
- Maintaining thousands of simultaneous conversations.
- Translating between languages.
- Generating or editing profile images.
- Remembering details about each target.
- Detecting signs that a user is becoming suspicious.
But Anthropic's investigation also demonstrates AI's limitations.
The operator still reportedly needed real people for tasks such as live calls and genuine social-media interactions.
The most effective deception was therefore not purely AI and not purely human. It was a hybrid system in which automation handled scale while humans handled moments that required stronger proof of authenticity.
Did Anthropic shut the operation down?
Anthropic says it banned accounts and organizations associated with the network, including disposable organizations and accounts linked directly to the operator's employees.
The company also says it shared relevant information with other AI providers.
Platform-specific details relating to the dating apps and review-evasion techniques were shared directly with Apple and Google, according to Anthropic.
That does not mean AI dating scams as a category have disappeared.
The company explicitly connects GTG-15001 to an earlier 2025 case in which another actor offered a Telegram bot that generated dating-app messages for scammers.
The important trend is therefore larger than one network: tools that once helped a scammer write one convincing message can increasingly support entire systems of fake identities.
Frequently asked questions
Can AI bots be used on dating apps?
Yes. Anthropic says it identified more than 4,700 AI personas operating across a network of more than 20 dating apps. The personas autonomously exchanged millions of messages with users.
How can I tell if I am talking to an AI on a dating app?
No single test is reliable. Look for patterns such as repeated avoidance of spontaneous interaction, inconsistent personal details, unverifiable social accounts, highly generic responses and attempts to keep you paying or move you toward money or investments.
Does a video call prove the person is genuine?
No. A live call is useful evidence, but Anthropic's report describes real workers being used specifically to perform video calls for a system in which most profiles were AI personas. Always combine video with other independent verification.
Can an AI dating profile have a real social-media account?
A deceptive operation can use humans to support an otherwise automated identity. In Anthropic's case, real workers reportedly performed social-media follow-backs to make the service appear authentic.
Can dating apps contain fake likes and visitors?
Yes. Anthropic says the backend in the network it investigated fabricated likes and visitors as part of the user experience.
What dating apps were named in Anthropic's report?
The report lists observed brands including DORA, DONI, ROMI, LUMA, JOVIA, KIRA, GRACECHAT, HAVEN, NALO and LOVIA, along with additional unidentified variants. A similar or identical name alone should not be treated as proof that a current app belongs to the same network.
What should I do if someone I met online asks me for cryptocurrency?
Do not allow a person you know only through dating or social media to direct your investment decisions. Independently verify any service before sending money, and treat guaranteed returns or pressure to deposit quickly as major warning signs.
Should I trust a dating-app verification badge?
A verification badge may confirm that a platform completed a particular verification step. It does not automatically prove every biographical claim, intention or financial recommendation made by the person behind the account.
The bigger lesson
The most unsettling part of Anthropic's dating-app investigation is not that an AI could send a convincing romantic message.
We already knew that was possible.
The more important development is that thousands of automated identities could operate continuously while real workers were inserted exactly where users expected proof that a human was present.
The result was a system in which the normal signals people use to decide whether someone is genuine — natural conversation, photographs, social-media activity and even video — could all be manufactured or selectively supplied.
That does not mean people should stop using dating apps.
It means the standard for trust has changed.
Do not ask only, “Does this person look real?”
Ask whether their identity, history, behavior and claims can be verified independently — especially before money, investments, sensitive information or intimate material enter the conversation.