Dating

Can You Trust an AI Dating Coach? What AI Should and Should Not Do in Dating

CoreAllure September 14, 2026 9 min read

Quick Answer:Can you trust an AI dating coach? It depends entirely on which of three jobs it is doing. AI that helps you say what you already mean is fine. AI that scores compatibility is selling certainty it does not have. AI that talks to people as though it were you, or as though it were a person at all, is deception, and it is already happening at scale.

How big is that third problem? In September 2026, Anthropic published a threat report describing a China-based studio that built over twenty dating apps and used Claude to run the personas inside them while advertising the service as fully human. In two weeks in April, more than 4,700 AI personas spoke to at least 25,000 people across 2.36 million messages. The match feed was roughly three AI personas for every real person.

The question is not whether AI belongs in dating. It is already there, and a quarter of singles say they use it.

The question is which parts of dating you are handing over, because the three main uses are not variations of the same thing. One helps you. One flatters you. One lies to somebody else on your behalf.

The Three Kinds of AI in Dating

AI that helps you express yourself. Reply suggestions, profile rewrites, openers, a second opinion on a message you have drafted five times. This is the biggest category by a distance. Rizz has passed several million downloads, and a long tail of near-identical tools sits behind it.

AI that judges compatibility. Scores, percentages, personality assessments, verdicts on whether two people fit. This is the one that sounds most impressive and stands up worst.

AI that talks as a person. Either as you, autonomously replying to your matches, or as a fabricated person entirely. This is the one with a documented body count.

They get discussed as if they were one topic. They are not, and the ethics differ completely.

AI That Helps You Say What You Mean

Broadly fine, with one real risk.

If you freeze on the first message, if you write something, delete it, and close the app, a tool that gives you three options is doing what a friend does over your shoulder. There is no deception in sending a message you chose, in your own voice, that you would stand behind.

The risk is dependency rather than dishonesty. The common complaint about reply generators, in reviews and from people who use them, is that the outputs are generic and that nothing is learned. You get the opener, and then the conversation goes somewhere unexpected at message three, and you are exactly where you started.

So the test is simple. Does the tool leave you better at this, or does it leave you unable to do it without the tool? If you would not recognise your own messages six months from now, it has stopped being a coach.

The practical version. Use it for the blank page, not for the whole conversation. Rewrite whatever it gives you in your own words. And never send something you could not defend in person, because you will have to.

AI That Scores Compatibility

This is where confident claims outrun the evidence.

A compatibility percentage implies that a system has measured something real and knows how two people will get on. Nothing in the public research supports that. The largest study of its kind, using machine learning on data from over 11,000 couples, found that individual traits predicted relationship satisfaction reasonably well while the specific pairing of two people predicted almost nothing beyond what each person brought individually.

That is the finding that should make every compatibility score suspect. What predicts how a relationship goes is largely how each person is, not the match between them, and a score sold as a measure of fit is measuring something it cannot see.

Voice-based scoring is worse still. Listeners agree strongly with each other about what a voice suggests, and those impressions frequently fail to match how the speaker rates their own personality. A product that puts a number on a voice is putting a number on a stereotype.

The practical version. Treat any percentage as marketing. Ask what it was calculated from and whether that has ever been validated against actual outcomes. Almost no product will answer.

AI That Talks as a Person

And here is the part that should change how you think about all of it.

Anthropic’s September 2026 threat intelligence report documents a China-based app studio that used Claude both to build a network of more than twenty dating apps and to run the personas inside them, while advertising the service as fully human. Over a two-week window in April 2026, Anthropic identified more than 4,700 distinct AI personas that spoke with at least 25,000 people, generating roughly 2.36 million messages.

The structure is what makes it instructive. The user-facing match feed was approximately 75 percent AI personas and 25 percent real people, a ratio of about three to one. Real people were recruited as gig workers to handle the tasks AI could not do convincingly, including live video calls and social media follow-backs. Another AI generated reply suggestions for those workers.

The apps were engineered to pass app store review, with a hidden interface that activated only during review and lay dormant otherwise, and code varied between versions to defeat similarity checks. Anthropic says it banned the accounts involved and shared indicators with Apple and Google.

One detail matters more than the numbers. Anthropic noted that the system prompt running these personas looked, on its face, like an ordinary roleplay or companion deployment. The technology was not exotic. The deception was in the product, not the model.

What that means for you. If you have ever wondered whether a match was real, that instinct is not paranoia. And the fix is not detection, because you will not spot it from the text. It is the same thing it has always been: a live video call, early, with something unpredictable in it.

The Line, Stated Plainly

Three questions, and they sort almost everything.

Does a human choose what gets sent? If yes, it is a tool. If the AI replies on your behalf while you sleep, it is not you dating, and the person on the other end has not consented to talking to software.

Is it offering context or a verdict? Context helps you decide. A verdict decides for you, and no dating product has the evidence to earn one.

Does the other person know? This is the only one that is genuinely non-negotiable. Everything Anthropic documented was technically ordinary. What made it abuse was that 25,000 people thought they were talking to human beings.

What AI Is Genuinely Good At Here

Worth saying, because the honest position is not that AI has no place in this.

It is good at lowering the barrier to starting. Approach anxiety is real, and something that gets you from nothing to a first sentence has done something useful.

It is good at noticing what you cannot see about your own patterns, given enough history. That is a different job from judging other people.

And it is good at giving context on a moment you are stuck in. Not what to do, but what might be happening.

What it is not good at is knowing who you should be with. That is the claim the industry keeps making and cannot support.

How Luna Is Built, and What It Refuses to Do

CoreAllure has Luna, and the boundaries are the point rather than a limitation.

Luna gives you context and support. She does not choose partners for you, does not assign compatibility scores, and does not judge character from a voice, because the research says nobody can do that reliably and a number would be an invented certainty.

She does not talk to anyone as you. There is no autonomous replying, no persona acting on your behalf, and nothing on the platform presents an AI as a person. Everything you send, you chose.

That leaves Luna doing the thing AI is actually good at. Helping you notice what mattered in an introduction, and offering context when you are unsure, while every decision stays yours.

The rest of the product follows the same logic. You hear a voice introduction before any photograph, and the only decision is whether to keep listening. If you are curious, you choose Resonate, which is not a like and does not create a match. Introductions are limited to five a day on Free and ten on Premium. And once two people choose to connect, there is a forty-eight-hour window to begin the conversation, and if nothing meaningful begins, the connection closes gracefully.

CoreAllure has not launched and is currently waitlist only.

The Bottom Line

You can trust an AI dating coach to help you say what you mean. You cannot trust one to tell you who fits you, and you should not let one speak for you.

The Anthropic report is the clearest evidence yet of where the line sits, and it is not drawn at the technology. The same models that draft a decent opener were running 4,700 personas that told 25,000 people they were human. What separated the two was whether anyone knew.

So the only question worth asking of any AI in dating is the one nobody in the industry advertises: does the person on the other end know what they are talking to?

Frequently Asked Questions

Can you trust an AI dating coach?

For help expressing yourself, largely yes, provided you choose what gets sent and rewrite it in your own words. For compatibility scores, no, since the research does not support them. For anything that talks to matches on your behalf, no, because the other person has not agreed to it.

Do AI compatibility scores work?

There is no good evidence for them. The largest study of its kind, covering more than 11,000 couples, found that individual traits predicted relationship satisfaction while the specific pairing of two people added almost nothing beyond that.

Are there fake AI profiles on dating apps?

Yes, and at scale. Anthropic’s September 2026 threat report describes a studio that ran over twenty dating apps where roughly three in four profiles in the match feed were AI personas, with 4,700 of them speaking to at least 25,000 people in two weeks.

How do I know if I am talking to AI?

Not reliably from the text. Take a live video call early and ask something specific and unpredictable. In the operation Anthropic documented, human gig workers were brought in precisely because video was the part AI could not handle.

Is it dishonest to use AI to write dating messages?

Not if you choose what is sent and it reflects what you actually mean. It becomes dishonest when the AI is replying autonomously, or when you could not have written anything resembling it yourself.

What should an AI dating assistant never do?

Reply as you without your involvement, present itself as a person, or issue verdicts about compatibility dressed up as measurement. The first two are deception. The third is a claim nobody can currently support.