The results of this experiment are a little uncomfortable.
A team called DateBench opened 15 profiles on real dating apps in three cities: San Francisco, Los Angeles, and New York. Each new match was randomly assigned to one of five AI models, and all messages were written by the model. The person opposite had no idea this was an experiment.
A total of 1,277 conversations took place over six days. Judging from the chat records, 355 people took the initiative to give their phone numbers and 510 people agreed to meet.
And in these 1,277 conversations, only eight times was the question asked: “Are you a robot?”
The gap between models is not small. Gemini 4 Argon got 87 numbers in 247 conversations, and no one suspected it was AI. Grok 4.7 performed the worst, with only 52 numbers obtained from 296 dialogues. It always repeated the same sentence, and 5 of the 8 times "Are you an AI" were directed at it. One of Claude's conversations started with a joke about muffins, and the next day the person agreed to meet and sent his number.
The experiment set two rules: you must admit that you are an AI when asked, and you are not allowed to agree on the real time and place. The author himself points out that scammers don’t follow any of them.
This is what’s really scary about this set of data. The most labor-intensive part of the pig slaughtering process is the early chatting. Now this step can be outsourced to the model in batches. The next time you meet someone special on a dating app, make a video call before meeting.