AI dating scams are becoming harder to recognise because the fake profile no longer has to behave like a crude bot. Anthropic says it disrupted a China-based operation that ran thousands of AI-generated dating personas across more than 20 apps, while human workers stepped in when targets demanded a video call, a social-media follow or some other proof that a real person existed behind the profile.
The company says it identified more than 4,700 distinct AI personas and at least 25,000 people who interacted with the operation over a two-week period in April 2026. Its analysis also found roughly three AI-controlled profiles for every real profile in the network. Those figures describe the activity Anthropic says it observed, not a count of confirmed financial victims, but they show the scale at which romance-style deception can now be industrialised.
That distinction matters. The old mental model of an online dating scam is one fraudster maintaining a handful of conversations. The newer model can look more like a small digital business, with AI doing the repetitive conversation and people appearing only when a task still needs a human face.
Why AI dating scams can look more convincing
The operation described by Anthropic did not rely on AI for every stage. That is precisely what makes the case useful. The automated system handled large volumes of messages, but human workers could take over for moments that might otherwise expose the deception. A victim who asked for a live video call could be given one. Someone who wanted to see activity on a social network could encounter an account managed by a person.
This hybrid model weakens several of the informal checks people use online. A bad photograph can be generated again. A stilted message can be rewritten. A video call may prove only that a person is willing to appear on camera, not that the identity, history or intentions attached to the account are genuine.
LiveAIWire has already reported how live challenges can help expose some deepfake callers. Dating scams add a different complication: the person on screen does not necessarily have to be a deepfake at all. A real worker can temporarily stand behind a fabricated identity.
The scammer does not need every conversation to work
AI changes the economics of deception because it lowers the cost of maintaining conversations that would previously have consumed a fraudster’s time. Most targets can ignore the messages, spot the problem or simply lose interest. An operation can still be worthwhile if a small fraction continue far enough for money, gifts, investments or personal information to enter the conversation.
Anthropic says the network used its models in ways that violated the company’s rules and that the relevant accounts were banned. It also says it shared information about the operation with other organisations. The important point for users is broader than the identity of one model provider: generative systems can make the conversational part of a scam cheap enough to run at a scale that was previously awkward for human teams.
That follows the same economic logic behind AI voice-cloning scams. The technology does not invent fraud, but it can reduce the amount of labour needed to make fraud feel personal.
App-store trust signals can be gamed too
According to Anthropic, the operation also tried to manage the reputation of its dating services. That included activity intended to improve reviews and reduce the chance that users would recognise the apps as part of a coordinated scam environment. A polished listing, positive reviews and an apparently active community therefore deserve less weight than many users instinctively give them.
The Verge, which reported on Anthropic’s findings, highlighted the same uncomfortable feature: AI could supply the profiles and conversation, while gig workers supplied the human moments. That division of labour makes the experience look less automated than it really is.
There is no evidence in Anthropic’s report that every person who spoke to one of the profiles lost money. It would be wrong to turn 25,000 interactions into 25,000 victims. The useful lesson is instead about verification: a believable chat, a photograph, a social account and even a short video call can all be produced inside a coordinated deception.
What a safer verification habit looks like
No single test can prove that a stranger on a dating app is genuine. The more useful approach is to look for consistency over time. Does the person’s history make sense across platforms? Do details remain stable? Is there pressure to move quickly off the dating service? Does the conversation turn towards investments, transfers, emergency expenses or requests for credentials?
People should also be cautious about treating technical polish as evidence of identity. Generative AI is particularly good at removing the small imperfections that used to make some fake accounts obvious. Smooth grammar, plausible photographs and fast replies are now weak signals.
The warning is not that online dating has become impossible to trust. It is that authenticity increasingly has to be judged from behaviour and consistency rather than from a single impressive proof. A scammer who can automate 90 per cent of an interaction can afford to spend human effort on the remaining 10 per cent.
Platforms now have to detect networks, not just bad messages
For dating services, the challenge is also shifting. Moderating an individual offensive message is different from identifying thousands of accounts that may each behave reasonably on their own. Useful signals can sit across account creation, device behaviour, payment patterns, copied biographies, image generation, conversation timing and links between apparently separate services.
That creates a familiar tension for platforms. The stronger the anti-fraud monitoring becomes, the more carefully it has to be designed around legitimate users’ privacy. Yet doing nothing leaves users to perform authentication in an environment where the obvious visual clues are becoming weaker.
LiveAIWire previously examined the wider rise of synthetic scams. The dating case is a particularly clear example of the next stage: AI does not have to replace the criminal. It can become the workforce around the criminal.
The uncomfortable new rule of online romance
The practical consequence is simple. A person can now encounter a dating profile that looks coherent, writes naturally and can even produce a human being on camera, yet still be part of a fabricated operation. That does not mean every unusual profile is a scam, and suspicion should not become the default way to treat strangers.
It does mean the old shortcuts for establishing trust are losing value. The strongest signal is increasingly whether the person behaves consistently, avoids financial pressure, accepts reasonable boundaries and can sustain a believable identity over time without constantly steering the relationship towards money or secrecy.
AI has made fake conversation scalable. Human intervention can make the fake feel real at exactly the moment a target becomes suspicious. That combination, rather than the chatbot alone, is what makes this generation of dating scam worth understanding.
Why scale changes the risk even when most people walk away
The economics of a large automated scam are different from those of a one-to-one con. If software can maintain thousands of conversations at once, the operator does not need a high success rate. Most contacts can fail without consuming much human labour. Human attention can then be reserved for the smaller number of people who appear emotionally invested or financially promising.
That is one reason the raw interaction figures matter even though they are not victim counts. More than 25,000 people encountering the network in two weeks creates many opportunities for a later hand-off, and the AI can keep each conversation feeling individual while it waits. A fraudster no longer has to choose between talking to one person carefully and talking to many people badly.
For dating platforms, this makes early network detection more valuable than waiting for an individual conversation to contain an obvious request for money. Patterns in account creation, shared infrastructure, repeated profile details or unusual messaging activity may reveal a coordinated operation before any single user has enough evidence to report it.
That shift also makes reporting suspicious behaviour more important. A single odd conversation may look trivial to one user, while a platform that receives several similar reports can see the network connecting them.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.