Abstract:
On September 11, Anthropic released the latest AI abuse report stating that the App Development Studio used Claude to develop and operate more than 20 dating apps. The platform claims that all users are real people, but about 75% of the matches received by ordinary users are AI characters.
Anthropic stated that during a two-week observation period in April 2026, this batch of dating software ran more than 4,700 AI characters, chatted with at least 25,000 different users, and sent a total of approximately 2.36 million messages. The affected people are mainly American users, but the report did not disclose their age, gender and specific loss amount.
It is difficult for these users to judge whether the person they are chatting with is real or fake. The platform displays materials in a ratio of approximately 3 AI characters to 1 real person. Claude is responsible for maintaining thousands of conversations simultaneously, and human part-time workers are there to reduce suspicion when users ask for video calls, photos, or follow social accounts.
Human part-time workers are also assisted by AI. The system will generate 3 alternative responses for each message, and part-time workers only need to choose one of them to send. They are paid according to the number of chat messages, video calls and social platform check-ins, and can be settled after reaching a low withdrawal threshold.
For accounts completely controlled by AI, the operator requires Claude not to admit that he is a robot. If the user makes a video or photo request, the AI will find reasons to avoid it and gradually advance the relationship according to the preset stages. The backend also creates fake likes, visitor records, and pre-recorded videos, and flags users who have begun to suspect that they are robots.
Users need to consume credits to chat and match on the platform. After using up, they must purchase in-app virtual currency to recharge. This model is similar to the "Beng Laotou" scam recently referred to on the Chinese Internet. They both use fake female personas and continuous chatting to stimulate men to pay.
The report also mentioned that in a few random conversations, some users revealed that they were suffering from serious diseases or were in extreme pain. Claude's internal reasoning has identified potential harm, but he continues to chat under the fictitious identity without exiting the conversation.
In order to avoid app store review, the operator also designed a special interface for the software that is only enabled during the review period. More than 20 apps used different program class names to evade detection mechanisms used by Apple and Google to identify similar apps. The entrance to the third-party payment platform can also be remotely controlled by the server and temporarily hidden during the review period.
Relevant dating software brands disclosed by Anthropic include DORA, DONI, ROMI, LUMA, JOVIA, KIRA, GRACECHAT, HAVEN, NALO and LOVIA, etc. The company has banned relevant accounts and provided app identity and review circumvention information to Apple and Google.
It should be emphasized that the above findings are from Anthropic’s unilateral disclosure and have not yet been verified by an independent third party. The report did not disclose the name of the application studio involved, nor did it disclose whether users had reported the case or the actual amount of money defrauded.
Comments