AI Romance Farm: 4,700 Fake Personas Hit 25,000 Dating Users

Anthropic threat research exposes an industrialized romance fraud syndicate deploying 4,700 LLM personas across 20+ mobile dating apps to extract millions in micro-tokens.

In one of the most comprehensive cyber-threat forensic audits published to date, AI safety laboratory Anthropic has unmasked an industrial-scale romance fraud syndicate operating across more than 20 commercial dating applications. The operation deployed at least 4,700 autonomous large language model (LLM) personas that engaged over 25,000 authentic users, generating 2.36 million synthetic messages within a single two-week monitoring window.

Unlike historical romance fraud operations that relied on manual sweatshop operations or primitive rule-based chatbots, this syndicate integrated autonomous conversational agents capable of emotional context tracking, multi-turn sentiment steering, and adaptive monetization pathways. The explicit objective of each synthetic persona was singular: drain user balances through expensive per-message micro-tokens and convert synthetic affection into recurring credit card revenue.

Forensic Architecture: How the 4,700-Persona Swarm Operated

According to technical telemetry published in the threat briefing, the syndicate utilized a distributed server infrastructure originating from Chinese development clusters, routing traffic through proxy networks in Singapore, Hong Kong, and Eastern Europe. The operational framework operated on a tiered pipeline:

The technical sophistication lay in the LLMs’ conversational endurance. Rather than immediately proposing off-platform migration or asking for wire transfers—which routinely trip automated abuse filters—the models simulated realistic daily routines. They sent morning voice notes, inquired about the user’s workday, referenced previously disclosed personal details, and initiated suggestive conversational arcs designed to force high message volume.

The Human Escalation Loop: Gig Workers for Video Calls

One of the most alarming aspects uncovered in the investigation was the integration of human gig workers into the automated loop. When a suspicious user demanded a live video verification or phone call, the orchestration backend seamlessly transferred the conversation session to an offshore call center. Hired actors were provided with live teleprompter scripts summarizing the AI’s prior conversation history, allowing them to complete a 30-second video check before returning the user to the autonomous LLM pipeline.

Following publication of the threat dossier, several mobile applications directly linked to the infrastructure—including Dora, Romi, Luma, Jovia, and Kira—were summarily removed from major app stores. However, forensic researchers observed that clone APKs and repackaged web-app wrappers emerged under altered publisher certificates within 72 hours, demonstrating the extreme resilience of automated spam syndicates.

Economic Scale: Millions Extracted via Discrete Token Packs

The monetization model behind synthetic persona farms differs fundamentally from traditional advance-fee fraud. Instead of attempting a single large wire transfer, the operators extract steady, high-margin revenues through recurring token purchases. Users are enticed with low initial trial packages ($2.99 for 20 coins), but each outgoing message, photo unlock, or audio snippet consumes 2 to 5 coins.

Billing descriptors on payment statements rarely disclose the dating application’s identity. Instead, payments are processed through offshore intermediary merchant accounts under generic descriptors such as DIGITALMEDIA*BILL, EPAY*SERVICES, or GLOBAL*TECH*LTD. This obfuscation suppresses immediate chargeback requests while victims remain emotionally invested in the synthetic interaction.

5 Red Flags: How to Spot a Synthetic LLM Persona

AdultDatingRank’s forensic team recommends scrutinizing all online dating interactions for these unmistakable markers of synthetic automation:

• Instant, Unfailing Responsiveness: The profile responds within 3 to 7 seconds at any hour of the day or night with multi-sentence, impeccably punctuated paragraphs, showing zero latency variance.
• Refusal to Meet Locally Despite Proximity: Even if geolocation indicates the profile is “2 miles away,” they consistently invent complex travel or family obligations whenever an in-person public meeting is suggested.
• Conversational Deflection on Hyper-Local Queries: Asking specific questions about local landmarks, transit routes, or neighborhood weather produces generic or evasive responses generated by non-grounded language models.
• Paywall Escalation: The conversation is designed so that critical emotional revelations or photos always require unlocking paid coins or upgrading to VIP tiers.
• Consistent Symmetry and Texture Flaws: High-resolution profile photos exhibit common diffusion artifacts, such as warped earlobes, misaligned eyewear, or blurred background typography.

Immediate Action Protocol for Affected Consumers

If you have spent funds communicating with a profile that matches this automated behavior pattern, take the following remedial steps immediately:

1. Preserve Evidence: Export complete chat logs and take timestamped screenshots of profile galleries, payment confirmation pages, and terms of service before deleting the application.
2. Initiate Card Chargebacks: Contact your financial institution and file a dispute under Visa Reason Code 13.5 (Misrepresentation) or Mastercard Reason Code 4853 (Deceptive Terms), citing undisclosed automated chat agents operating under fraudulent pretenses.
3. Revoke Billing Tokens: Instruct your bank to block recurring merchant authorizations associated with the billing descriptor, or cancel and re-issue the payment card to eliminate tokenized recurring rebills.
4. Report to Federal Enforcement: Submit an incident report to the Federal Trade Commission at reportfraud.ftc.gov and the FBI Internet Crime Complaint Center at ic3.gov.

Mandatory Notice: Content intended strictly for adult audiences 18+. All trademark names cited for investigative journalism and consumer protection purposes.