Red Flags

How to Tell If a Dating Profile Is AI Written

CoreAllure September 16, 2026 9 min read

Quick Answer:How to tell if a dating profile is AI written? Honestly, you mostly cannot, and the confident checklists circulating are out of date. Research by Jakesch and colleagues tested exactly this, showing people AI-generated self-presentations from online dating, job applications, and Airbnb, and found identification accuracy was low. A 2025 paper in Communications of the ACM put human detection of AI content at roughly coin-toss level.

So what should you do instead? Stop trying to detect and start verifying. The tells people share are mostly artefacts of older models; detection tools are unreliable enough to be dangerous, and the only thing that reliably works is a live conversation with something unpredictable in it.

Everybody thinks they can spot it.

The slightly too-smooth phrasing. The three-item lists. The word “delve”. People share these tells confidently, and the research says they are barely better than guessing.

Worth knowing what the evidence actually shows before you accuse someone of not existing.

What the Research Says About Your Ability to Tell

This has been studied more thoroughly than most people realise, and one study is almost perfectly on point.

Jakesch, Hancock and Naaman tested how well people could identify AI-generated self-presentations in three contexts, one of which was online dating profiles. Identification accuracy was low, and the reason was interesting: people rely on heuristics about what AI writing sounds like, and those heuristics are wrong.

The broader picture is consistent. A 2025 paper in Communications of the ACM was titled, plainly, As Good as a Coin Toss: Human Detection of AI-Generated Content. A large cross-country study found people struggled to distinguish AI from human material across images, audio, and text, often performing at or below chance, and consistently rated AI-generated media as more likely to be human-made.

And it is getting harder rather than easier. The International AI Safety Report 2026 notes that in one study, after a five-minute conversation, participants misidentified text generated by GPT-4o as human-written 77 percent of the time.

Even people with training do not do much better. One review found annotators experienced with AI text performed no better than those with no exposure at all.

Why the Tells You Have Been Told Are Wrong

Because almost all of them describe a model from two or three years ago.

Em dashes, “delve”, “tapestry”, “it’s not just X, it’s Y”, suspiciously balanced three-item lists. Those were genuine fingerprints of early ChatGPT outputs. They are now the most widely known tells in existence, which means anyone using AI deliberately has been told to remove them, and current models produce them far less anyway.

There is a worse problem. Every one of those patterns also appears in writing by actual humans, particularly people who write for a living, people writing in a second language, and anyone who has ever been taught to write clearly. Using them as evidence produces false accusations, and the research on AI detection tools shows exactly that risk: a 2023 evaluation of fourteen detection tools found them broadly unreliable, and there are documented cases of students being wrongly flagged.

So the checklist approach fails twice. It misses AI text written by anyone competent, and it flags humans who happen to write tidily.

The Ones That Still Mean Something

A short list, and none of them are proof.

Nothing specific. AI-written bios tend toward categories rather than instances. Loves travel, food, and good conversation. A human writing about themselves usually names a place, a dish, a thing that happened. This is the most useful signal, and it is also just a description of bad writing.

Answers that do not touch each other. Ask something that requires them to connect two earlier things they said. Generated text and a person working from a suggestion tool both tend to answer each message in isolation.

Register that does not shift. Real people get shorter when they are busy, sloppier late at night, warmer once they like you. Consistency of tone across three weeks is unusual.

A reply arriving suspiciously fast, at length, every time. Composing a thoughtful paragraph takes minutes. Generating one does not.

Treat these as reasons to take a call, not conclusions.

AI Photos Are a Separate and Harder Problem

The text question gets all the attention, and the photograph question matters more, because a photo is what most people actually decide on.

The old advice was to look at hands, teeth, ears, and backgrounds. That advice is now largely obsolete. Research found that while people had moderate accuracy detecting AI-generated images overall, detection dropped substantially for images produced by more advanced models, and image quality has improved sharply across successive generations.

What still helps is not looking harder at the image. It is looking at the set. A profile with six photographs of the same person in genuinely different contexts, some of them unflattering, some with other people in them, some clearly old, is difficult to fabricate convincingly. A profile with three immaculate images and nothing else is worth a question.

Reverse image search remains worth thirty seconds, because a great deal of catfishing still uses stolen real photographs rather than generated ones, and that is much easier to catch.

Stop Detecting, Start Verifying

This is the whole practical answer, and it is the same conclusion that applies to voice cloning.

You are not going to win an arms race against a generative model by reading more carefully. What you can do is ask for something a generator cannot supply.

Take a live video call, early. Not a recording. Live, with a question you did not warn them about. Something about the room they are in, or what they did an hour ago. In the AI persona operation Anthropic documented in September 2026, human gig workers were brought in specifically because video was the part the AI could not handle convincingly.

Ask something that requires continuity. Reference something they said last week and ask them to expand on it. Systems that reply message by message struggle with this. So do people using a suggestion tool without reading the thread.

Notice the shape of the conversation, not the prose. Does it deepen? Does it go anywhere neither of you planned? A generated exchange tends to stay pleasant, symmetrical, and strangely flat.

And watch what it moves towards. The reason this matters at all is fraud, and fraud has a direction. Money, investments, crypto, urgency, an emergency. FTC data published in April 2026 found nearly 60 percent of people who lost money to a romance scam in 2025 said it began on a social media platform rather than a dating app. The destination matters far more than the prose style.

Does It Even Matter If a Bio Is AI Written?

Worth asking, because the answer is not always yes.

Someone who used a tool to tidy a bio they wrote themselves has done roughly what a friend proofreading it would do. A quarter of singles report using AI somewhere in their dating, and most of that is drafting help rather than deception.

Where it matters is when there is no person behind it. A profile that is entirely generated, replying autonomously, representing nobody. That is what Anthropic’s threat report described at scale: more than 4,700 AI personas speaking with at least 25,000 people across roughly 2.36 million messages in a fortnight, on apps where the match feed was around three-quarters AI.

So the useful question is not whether a sentence was generated. It is whether there is somebody on the other end. Those are different questions, and only the second one can be answered.

Where CoreAllure Fits

CoreAllure does not claim to detect any of this, and would be overreaching if it did.

What it changes is what you encounter first. You hear someone’s voice introduction before any photograph is shown, which means the opening signal is not a written bio and not an image. A voice is harder to fabricate casually than a paragraph, and considerably harder to sustain across a conversation than a generated reply.

That is not verification, and the site says so. Voice cloning exists, and hearing someone proves nothing about who they are. What it does is move the first encounter to a signal that takes real effort to fake, and give you enough time to reach your own judgement.

Luna gives context on what you heard. She does not score compatibility, does not judge character from a voice, and does not reply to anyone on your behalf. Nothing on the platform presents an AI as a person, which is the specific line the Anthropic report shows being crossed elsewhere.

Introductions are limited to ten a day on Free and twenty-five 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 cannot reliably tell whether a dating profile is AI written, and the studies that tested exactly this found people barely above chance, including on dating profiles specifically.

The tells being shared are largely obsolete, detection tools are unreliable enough to wrongly accuse real people, and the images are getting harder to spot rather than easier.

So stop reading for fingerprints. Take a live call early, ask something unpredictable, and watch where the conversation is trying to go. That works whether the person is real, and matters most when they are not.

verifying someone with a live video call

Frequently Asked Questions

How to tell if a dating profile is AI written?

Largely, you cannot. Research testing exactly this, including on online dating profiles, found low identification accuracy, and a 2025 paper in Communications of the ACM described human detection of AI content as roughly coin-toss level. The reliable approach is a live video call rather than reading for tells.

Do AI detection tools work on dating profiles?

Not dependably. An evaluation of fourteen detection tools found them broadly unreliable, and there are documented cases of human writing being wrongly flagged. A detector score is not evidence.

What are the signs of an AI-written bio?

The only one worth much is vagueness. Generated bios tend to name categories rather than instances, so travel and good food rather than a specific place or dish. Em dashes and certain words are now widely known tells, which means anyone using AI deliberately has removed them.

Can you spot AI-generated profile photos?

Less easily than you could. Research found detection accuracy drops substantially for images from more advanced models. Looking at the whole set helps more than scrutinising one image, and reverse image search is still worth doing because much catfishing uses stolen real photos.

Is it wrong to use AI to write a dating profile?

Not if you wrote the substance and used it to tidy the wording, which is what most people are doing. It becomes a problem when there is nobody behind the profile at all, or when AI is replying on someone’s behalf without the other person knowing.

How do I actually verify someone is real?

A live video call, early, with a question they could not have prepared for. In the AI persona network Anthropic documented in 2026, human workers were recruited specifically because live video was the part the system could not fake.