AI wrote that profile, and it's no longer possible to tell by looking

Every piece of advice about spotting fake profiles was written for a world that ended about two years ago.

Look for blurry or low-resolution photos. Watch for awkward English. Check whether the bio reads like a translation. Notice hands with too many fingers. All of it was genuinely useful, all of it is now obsolete, and a lot of people are still running the old checklist and feeling safe.

That gap between what people think detects a fake and what actually does is the entire problem, and it's worth being precise about where it now sits.

What specifically stopped working

Image quality. Generated faces are no longer low-resolution or uncanny. They are photographically ordinary, which is much harder to detect than photographically perfect. The tell used to be that fakes looked too polished; the current generation looks like a normal phone camera in normal light.

Language. Translation and drafting tools removed the awkward-phrasing signal entirely. A profile written by someone with no English at all now reads like a native speaker with a decent sense of humor.

Internal consistency. This is the significant one. Fakes used to fall apart because the story drifted — the job changed, the city moved, the details didn't line up over three weeks. A conversation assisted by a model doesn't drift, because it has the whole thread in front of it. Consistency is no longer evidence of honesty.

Response speed and quality. The old pattern was either instant generic replies or long delays. Now you get replies that are well-paced, specific to what you said, and warm. That's not a signal of anything anymore.

What still works

Three things, and they have one property in common: they require something to exist outside the conversation.

The live video call. Still the single most effective check available, and the gap between it and everything else has widened rather than narrowed. Real-time video with an unpredictable prompt — turn your head, say the thing I just said — remains genuinely difficult to fake convincingly. Nothing about the text has to be evaluated once you've done it.

Reverse image search. Less decisive than it was, because a generated face won't appear anywhere else. But it still catches the larger category of stolen real photos, which hasn't gone away and remains the more common attack.

Physical specificity that can be checked. Not "where did you grow up" but something with a verifiable local answer. Someone who claims to live in your city should be able to have an unremarkable conversation about a road closure, a restaurant that shut, the weather last Tuesday. This is weaker than a video call and it's free.

The common thread: anything happening inside the message thread can now be produced. The checks that survive are the ones that reach outside it.

The honest problem with detection advice

Including ours, and it's worth stating plainly.

Every list of tells has a shelf life, because the tells are published and then designed around. Advice about detection is inherently a losing position: you are updating a checklist against a system that updates faster, and the gap between the two is where people get hurt.

That's not an argument for giving up. It's an argument that individual vigilance is the wrong layer to solve this at, and that treating it as a personal skill problem quietly puts the burden on exactly the people least equipped to carry it.

Where the burden actually belongs

On the platform, and this is the part the industry has been slow about.

The structural fix is that identity is established once, at the door, by a process that a generated image cannot pass. After that, nobody in the conversation needs to be running detection at all, because the category of problem was removed rather than distributed.

Most major apps offer verification and make it optional. That produces the worst available outcome: a badge that tells you something about the minority who completed it and nothing about everyone else, and a false sense that the platform has handled this. Anyone acting in bad faith skips an optional step. So do plenty of real people who couldn't be bothered, which is precisely why the badge can't carry weight. We made the fuller argument in why verification matters more than matches.

Required verification is a different thing entirely. It costs the platform sign-ups, which is why the incentive runs against it, and that's the whole reason this hasn't been fixed at the layer where it's actually fixable.

The legitimate use nobody wants to discuss

Worth separating, because the conversation collapses two very different things.

Using a model to tidy your own bio is not fraud. Someone who genuinely lives where they say, looks like their photos, and asked a tool to help phrase a prompt has misrepresented nothing. A lot of people write badly about themselves and the assistance is closer to a friend's proofread than to deception.

Where it becomes a problem is not the tool, it's the gap. If the profile is funnier than you, warmer than you, or more articulate than you'll be in person, you've built something that has to survive a first meeting, and it won't. The failure mode isn't ethical, it's practical: an over-assisted profile converts to more matches and fewer second dates, which is a worse outcome than the honest version.

Our advice on the writing side hasn't changed much: specific beats polished, and a model is much better at polished than at specific.

The conversational tell that hasn't gone away

One thing does still separate assisted conversation from a real one, and it isn't a linguistic feature.

It's the absence of friction. Real conversations have small, pointless bumps in them: a misread joke, an opinion you didn't expect, a question that lands slightly wrong, a subject someone doesn't want to talk about. Assisted conversation is frictionless because agreeableness is the default output, and three weeks of perfectly smooth exchange is itself unusual.

The practical version: pay attention to whether anyone has ever disagreed with you. A person who has never pushed back, never been slightly bored, never had a preference that inconvenienced you, is either performing or isn't there.

Notably, this is also a good filter for real people you'd be wasting your time on, which is a rare piece of advice that works in both cases.

What to actually do

Compress the timeline. The single biggest change is that text no longer carries evidence, so extended messaging has lost most of its diagnostic value and kept all of its cost. Four to six exchanges, then a short video call, then a meeting.

Treat the video call as non-negotiable rather than as a favor. The framing that works is that you want to see whether the conversation works live, which is true and also happens to be the check.

Keep the money rule absolute. Nothing here changes it. No amount, no reason, no stage, no exceptions — the pattern is unchanged even as the presentation improved, and we broke it down in how to spot a romance scam.

And choose the platform deliberately, because that decision does more work than anything you do inside it. The ten-minute version of the personal checks is in how to check someone is real before you meet them.

Where this goes

Live video is the current floor, and it is not a permanent one. Real-time manipulation is improving, and the honest position is that the check which works today has a horizon rather than a guarantee.

What that implies is that platforms which verify identity at the door are not solving a temporary problem in a clever way. They're the only architecture that doesn't need to be re-solved every eighteen months as the tools improve, because it doesn't depend on anyone's ability to detect anything.

That's the actual case for verification, and it's a structural argument rather than a marketing one. It happens to be the thing we built Plus around.

Frequently asked questions

Can you tell if a dating profile photo is AI-generated?

Not reliably by looking, and not since roughly 2024. Generated faces are now photographically ordinary rather than uncanny, which is harder to spot than obvious perfection. A live video call is the only widely available check that still works consistently.

Do AI-generated profiles still have bad grammar?

No. Drafting and translation tools removed that signal completely. A profile written by someone with no English at all now reads like a native speaker with a decent sense of humor, so language quality tells you nothing either way.

Does reverse image search still work?

Partly. It won't catch a generated face, because that image exists nowhere else. It still catches stolen photos of real people, which remains the more common attack, so it's worth ninety seconds even though it's no longer decisive.

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

Not in itself. If you live where you say and look like your photos, help with phrasing isn't deception. The practical risk is the gap: a profile funnier or more articulate than you'll be in person converts to more matches and fewer second dates.

What's the best way to verify someone is real now?

A short live video call with something unpredictable in it. Everything inside the message thread can now be produced, so the checks that survive are the ones that reach outside it. Compress the messaging stage and get to video within a few days.

Why don't dating apps just require verification?

Because it costs sign-ups. Verification adds friction at the moment someone is deciding whether to join, and platforms measured on member growth have a direct incentive to keep it optional, give a badge to the minority who complete it, and count everyone else anyway.