Resource guide

AI Bots on Dating Apps in the DateBench Experiment

A six-day experiment shows that AI can carry on dating-app chats, but leaves the number of bots on dating apps unknown.

Last updated November 16, 2026 1569-word guide By Jonathan West

AI bots on dating apps can hold chats that people treat as conversations with a potential date, according to the DateBench experiment. Across 1,277 conversations, 355 people sent a phone number and eight asked whether they were talking to a bot. The experiment shows what five AI models did in one setting; it cannot tell you how many bots are on dating apps or give you a reliable way to identify one from a chat.

What DateBench Found

DateBench reported that five AI models wrote messages to real people on one unnamed dating app during a six-day experiment. Each new match was assigned to a model at random, and that model wrote every message in the resulting conversation.

DateBench's reported outcomeCountWhat the count describes
Conversations1,277Chats involving a model and a real person
Agreed to meet510Conversations a separate model judged to include an agreement to meet
Sent a phone number355Conversations in which the person supplied a number
Asked whether it was a bot8Conversations in which the person asked about the match being a bot

The counts describe different events. DateBench says a separate model read the transcripts to judge whether a meeting had been agreed. A person sending a number is a distinct reported outcome. Neither count shows that anyone met the person behind the profile.

The eight questions have an even narrower meaning: they show how often someone asked whether the match was a bot. Someone could have suspected AI without asking. Calling eight out of 1,277 a detection rate would turn an observable question into a claim about what every person believed.

For a reader looking at a convincing message from a match, the finding is limited but useful. In this experiment, a model could keep a dating-app conversation going, and some people shared a phone number during those chats. Fluent replies, by themselves, did not establish that the writer was human. DateBench did not test a method that would let a reader determine the identity of any particular match.

The study also does not answer who puts bots on dating apps. DateBench's researchers set up the AI-driven profiles for their experiment. Their reported results cannot establish whether an app operator, a scammer or anyone else runs a profile a reader encounters outside the study.

How the Experiment Worked

DateBench ran one experiment for six days on one dating app in three cities. Its reported setup used 15 profiles and five AI models. The researchers have not published the prompts or code.

The models were told to admit they were AI if someone asked. DateBench says the conversation ended when a model was asked about being a bot under those rules. That instruction matters when reading the eight questions: the experiment counted people who raised the issue in a setting where the model was supposed to answer. It did not test what would happen if a person asked a deceptive operator who refused to disclose AI use.

The models were also told not to set dates. DateBench nevertheless reports 510 conversations that a separate model judged to contain an agreement to meet. That makes the judgment method worth keeping beside the number. The reported count is an assessment of chat transcripts, not a record of meetings that took place.

The researchers contacted real people through dating-app profiles. Those people's participation raises a separate question about the design of the experiment. Reporting the results here does not endorse contacting people in that way. It also does not turn the experiment into a safety test of the unnamed app.

DateBench published examples of conversations alongside its aggregate results. A transcript can show the kind of exchange its researchers recorded, but one exchange cannot establish how the other conversations went. The aggregate figures describe what DateBench reported across the run; the prompts and code remain unpublished.

What the Experiment Cannot Tell You

DateBench cannot measure the prevalence of AI bots on dating apps because its researchers supplied the models used in the experiment. The 1,277 conversations came from that setup. They were not a count of bots found among ordinary accounts across dating services.

One app is also too narrow a basis for claims about other apps. DateBench does not identify the app in its published write-up, so there is no basis here to name it or apply its results to a particular service. The three cities and 15 profiles describe where and how this run took place; they do not establish what a match on another app would experience.

The study had no human baseline. Without comparable chats written by people under the same conditions, the reported phone-number and meeting-agreement counts cannot show whether the models did better or worse than human matches. DateBench reports what happened in its AI-driven conversations. It does not supply a head-to-head measure of dating success.

Its outcomes have different evidence limits, too. DateBench reports that 355 people sent a phone number, while a separate model judged the meeting-agreement outcome. Treating both figures as independently verified real-world events would obscure that difference. A phone number in a chat does not show that a meeting followed, and a transcript judged to contain agreement does not show that two people met.

Nor does the experiment provide a reliable list of phrases that expose a bot. Eight conversations included a direct question about whether the match was AI. The remaining conversations cannot be classified as cases where the person was fooled: DateBench did not measure every participant's suspicions. Someone may have suspected AI without asking.

DateBench presents results for individual models, but says some differences between them may not hold in a larger sample. A ranking based on this one run would give readers more certainty than the study supports. The study does not validate an identity check a reader can apply to a new conversation.

Published prompts, code and independently assessed transcripts would make the experiment easier to examine. They would still answer a different question from prevalence. DateBench does not provide that evidence.

What the Study Suggests for Dating-App Users

DateBench gives readers a reason to avoid treating a natural-sounding chat as proof of identity. Its models wrote to real people, and 355 people sent a phone number during the conversations. That finding supports caution about sharing contact details with an unverified match; it does not mean every request to continue a conversation elsewhere is fraudulent.

The Federal Trade Commission's romance-scam guidance identifies a quick push to move a conversation off a dating site or app as one warning sign. The FTC also describes scammers who say they cannot meet in person, ask for money and create urgency. Those signs concern possible fraud, whether a person or an AI model wrote the messages.

If a match asks to leave the app, you can choose to keep the conversation there while deciding what to share. If the match then asks for money, the FTC's advice is direct: "Never send money or gifts to a sweetheart you haven't met in person." The FTC specifically warns about requests involving wire transfers, gift cards and cryptocurrency.

The distinction matters. DateBench did not report that its models asked people for money, and its phone-number count is not a count of scams. The study shows that a model can take part in conversations where someone shares contact information. The FTC guidance supplies the separate fraud warning about moving off-platform and sending money.

A video call should not settle an identity concern on its own. Ban the Bots' guide to deepfake scams and how to avoid them describes fraud involving fake video calls. If an identity claim matters to a decision about money or safety, verify it through a separately confirmed channel rather than relying only on the call. The guide on how to spot a deepfake offers detection detail, but visual clues are not a substitute for independent confirmation.

If you suspect a romance scam, the FTC advises stopping communication and reporting it at ReportFraud.ftc.gov and to the app where you met. If you have already paid, contact your bank, gift card company or wire service right away and ask for a refund. You do not need to prove that a match used AI before taking those steps.

What Would Change This Picture

DateBench supports a narrow answer: five AI models held chats with real dating-app matches, and some of those chats led to a shared phone number or a transcript judged to include agreement to meet. Its one-app, six-day run does not show how often readers encounter bots, whether they can reliably identify one or how its models compare with human matches.

Independent replication and published methods or data would help assess how firmly the reported results hold. A measured prevalence study from dating apps would address the separate question of how common bots are. Until then, a claim that bots are widespread cannot be drawn from DateBench's 1,277 conversations.

For an individual match, the practical decision does not require a bot count. Keep the FTC's money warning in place regardless of how the messages sound. If a match presses to leave the app, says they cannot meet or asks for a payment, stop and assess those signs rather than trying to diagnose the writing style. Report a suspected scam through ReportFraud.ftc.gov and the dating app, and use a separately confirmed channel for any identity claim you need to verify.

Frequently asked questions

▸ Do dating apps have AI bots texting people?
AI models texted real people on one unnamed dating app in the DateBench experiment. The experiment does not show how many bots are active on dating apps or who operates them.
▸ Can I tell whether a dating-app match is a bot from the chat?
A fluent chat is not proof that a match is human. Only eight of DateBench's 1,277 conversations included a question about whether the match was a bot, but the study did not measure unspoken suspicions or establish a reliable detection method.
▸ Does asking for a phone number mean a match is a scammer?
No. DateBench reports that 355 people sent a phone number in chats with AI models; it does not say that every request for a number is a scam. The Federal Trade Commission identifies a quick push to leave a dating platform as one possible warning sign when assessing a romance scam.
▸ Will a video call prove a dating-app match is real?
No. Scammers can fake video calls, so a video call alone does not establish someone's identity. Verify through a separately confirmed channel before trusting a claim that affects your safety or money.
▸ What should I do if a match asks for money?
Do not send money or gifts to someone you have not met in person. Stop communicating and report a suspected romance scam to the Federal Trade Commission and the dating app. If you have already paid, contact your bank, gift card company or wire service right away and ask for a refund.