Somewhere in the last eighteen months, a specific kind of sentence started appearing across dating profiles with suspicious regularity. "I'm equally comfortable in a boardroom and on a hiking trail." "Looking for someone who takes life seriously but not too seriously." "My friends would describe me as the group's chaotic-good energy."
None of these sentences are false, exactly. They are also, increasingly, not written by the people whose profiles they appear on.
Three in four singles now say they use ChatGPT in some part of their dating life, and nearly two-thirds report that AI has helped them build a stronger profile, sustain a conversation, or open one. Use has grown by roughly 333% in a single year. This is no longer a fringe behaviour practised by the terminally online. It is, at this point, closer to the median experience of dating in 2026.
The tools are, on their own terms, good at their job. They produce grammatically clean, well-structured, plausibly charming text. What they cannot do — what no amount of prompting seems able to fix — is make that text sound like it came from a specific person rather than from the general idea of an appealing person. And when enough people use the same tool to solve the same problem, the result is not a population of improved individuals. It is a population of extremely similar outputs.
The optimised-beige problem
Call it what it is: everyone is starting to sound like everyone else, only more competently.
This is a predictable consequence of optimisation applied at scale. Ask any sufficiently capable model to write "an engaging, warm, slightly witty dating bio," and it will reach, with impressive consistency, for the same register — self-deprecating but confident, ambitious but grounded, adventurous but also happy with a quiet night in. It is not wrong. It is also not distinguishable, at the scale dating apps now operate, from several hundred thousand other bios reaching for the same register at the same time.
The tragedy of the optimised-beige profile is that it works, marginally, on the metric it was built for — clean grammar, no red flags, broad appeal — while failing completely on the metric that actually determines whether two people connect: does this sound like someone specific enough to be curious about. A profile that is optimised to offend no one and appeal to everyone tends, in practice, to genuinely interest no one. It reads like a press release for a person rather than like a person.
And the people on the receiving end have started to notice. Roughly six in ten dating app users now say they believe they have encountered AI-written messages, and a majority of daters say they would lose interest in a match on learning their profile or their opening line had been AI-generated — even among the same people who use these tools themselves. The asymmetry is the tell: everyone wants the shortcut for themselves and distrusts it in everyone else. That is not a stable equilibrium. It is a population quietly losing faith in the authenticity of its own courtship rituals while continuing to participate in them.
What the arms race actually optimises for
It is worth being precise about what is happening technically, because the mechanism explains the outcome.
A large language model producing a dating bio is not describing a person. It is predicting the most statistically probable next word given a prompt like "single, ambitious, 34, based in the city, enjoys travel and good wine." The output is, definitionally, an average — a highly polished, grammatically flawless average of every appealing bio the model has ever been trained on. It is built to converge on the centre of the distribution, not to describe the specific, idiosyncratic, occasionally odd truth of one particular person.
This is precisely backwards from what makes a profile — or a person — compelling. Distinctiveness is not noise to be smoothed out of a good bio. It is the entire signal. The detail that is slightly too specific, slightly too weird, slightly too revealing to have been generated by a system optimising for broad appeal is usually the detail that makes someone stop scrolling. AI-assisted photo editing shows the same pattern from a different angle: profiles using heavily enhanced, professionally optimised photography now convert at meaningfully higher rates in raw swipe volume, and yet a real and growing share of users say they actively distrust the polish and select against it once they notice it. The market is bifurcating in real time between profiles engineered to perform well in aggregate and profiles that read as unmistakably, specifically true.
The apps themselves are responding to this tension unevenly. Some platforms have begun experimenting with visible disclosure badges for AI-assisted photos; matching algorithms are being retrained on behavioural signal rather than stated preference in an attempt to route around the noise. None of it resolves the underlying problem, because the underlying problem is not a labelling issue. It is that the format — photo, prompt, bio, all reviewed in isolation, all optimisable — rewards the exact behaviour that erodes the thing the format exists to produce.
What this reveals about the format, not the tool
The instinct is to treat this as a story about artificial intelligence. It is really a story about what happens to any format that can be gamed at sufficient scale.
Dating apps were always, at bottom, a text-and-photograph problem: two static artefacts, reviewed by a stranger in roughly the time it takes to form a first impression, standing in for the entire complexity of a human being. This was always a lossy compression. AI has not introduced a new failure mode into that system. It has simply made the existing failure mode — the incentive to optimise the artefact rather than represent the person — dramatically easier to act on, at dramatically greater scale, with dramatically higher production values.
A profile has always been a performance, to some degree. Curated photographs, carefully chosen prompts, a version of oneself edited for maximum appeal — this predates any of the current tools. What has changed is the ceiling on how polished that performance can become, and the floor on how much actual authorship it requires. When the artefact can be perfected without limit, and the perfecting requires almost no involvement from the person it is meant to represent, the artefact stops correlating meaningfully with the person at all. This is not a temporary bug that better AI detection will patch. It is what happens, predictably, when the thing being judged is fully decoupled from the person doing the judging's actual object of interest.
The interesting question is not whether AI-written profiles are common. They plainly are, and will become more so. It is what happens to a courtship system once both parties suspect, correctly, that the artefacts in front of them are an unreliable guide to the people behind them — and what people do once that suspicion becomes the default rather than the exception.
What doesn't optimise away
There is a reason this problem is confined almost entirely to the profile — the static artefact — and does not survive contact with an actual conversation.
Six minutes across a table cannot be pre-generated. Nobody has yet built a tool that will draft your facial expressions, your timing, the specific thing you find funny that you didn't expect to say out loud, the pause before an honest answer. These are not artefacts. They are behaviour, produced in real time, under conditions nobody can rehearse for indefinitely. This is, not incidentally, exactly the information a profile is trying and failing to compress into six photographs and three prompts.
This is also why the distrust triggered by AI-written profiles doesn't generalise to in-person meetings the way it might seem it should. A person can be caught having used an AI-optimised bio and remain, on meeting them, exactly as specific and unpredictable and worth being curious about as anyone else — because the bio was never really them to begin with, and the room reveals that immediately. The correction for an optimised-beige courtship culture is not a better detector for AI text. It is a return to formats where the thing being judged cannot be pre-written by anyone, for anyone, at scale.
We have hosted more than 19,000 structured social evenings across 50+ cities since 2014, and the thing this entire period has taught us most clearly is that no version of the profile — however written, however polished, however honestly or dishonestly produced — has ever been a reliable substitute for watching someone actually respond to a real question in real time. The profile was always a proxy. It was tolerable when it was a rough one. It becomes something closer to noise once it can be perfected without any of the person's actual involvement.
None of this requires giving up on technology, or treating AI-assisted dating as a moral failing rather than what it obviously is: a rational response to a format that has always rewarded the performance over the person. It requires noticing that the format itself has a shelf life, and building — or attending — the alternative.
The room does not have an optimised-beige problem. It cannot. There is no prompt for a specific human being sitting across from you, thinking about what to say next.
Relish hosts structured social evenings for driven professionals across 50+ cities in the US, UK, Canada and Australia since 2014. Find an evening near you →