Keeper's AI Attraction Model: Brutal Honesty or Algorithmic Recklessness?

A dating app has abandoned two decades of compatibility-focused algorithmic orthodoxy in favour of something considerably more direct: AI that scores how attractive you are, then shows you people in your bracket. Keeper's approach, now live in Boston, dispenses with the personality questionnaires and values alignment that competitors have spent years marketing as the key to lasting relationships. According to founder Jordan Kozloski, the platform simply replicates what professional matchmakers have always done—prioritise mutual physical attraction first, compatibility second.

The technology marks a clear break from the algorithmic storytelling that Hinge, eHarmony, and Match have spent years refining. Those platforms built entire brand identities around the promise of deeper compatibility—'designed to be deleted', in Hinge's phrasing—positioning themselves as the antidote to the supposedly shallow mechanics of Tinder's swipe interface. Keeper's positioning suggests that narrative may have been holding the industry back.

Create a free account

Unlock unlimited access and get the weekly briefing delivered to your inbox.

The real question isn't whether that's true—it's whether scaled AI-driven attractiveness scoring can avoid replicating the biases that human judgement has spent centuries perfecting.

If Keeper's model reflects narrow beauty standards encoded in its training data, this isn't innovation—it's automation of exclusion.

The compatibility myth meets its reckoning

For years, Match Group (MTCH) properties and Bumble (BMBL) have leaned heavily on the idea that their algorithms uncover compatibility invisible to the human eye. Hinge's 'Most Compatible' feature. eHarmony's 29 dimensions of compatibility. OkCupid's question-driven matching percentages. The premise has always been that technology can identify long-term partnership potential better than users can themselves.

The evidence for that claim has been thin. Matching algorithms are largely black boxes, and the companies that deploy them rarely disclose efficacy data. What they do disclose—engagement metrics, time spent on platform, conversation starts—measures product stickiness, not relationship outcomes. Users have no way to verify whether the person served up as 'highly compatible' is materially more likely to become a long-term partner than someone they'd have found attractive on their own.

Keeper's model strips that pretence away entirely. Physical attraction is sorted first. Compatibility—if it's considered at all—comes later. That's not a bug, according to Kozloski; it's the entire point. He references the work of professional matchmakers, who allegedly use the same approach with their high-net-worth client rosters.

But that comparison falters under scrutiny. Elite matchmakers operate with small, pre-vetted pools of affluent singles, often numbering in the dozens or low hundreds. They exercise human discretion, apply contextual knowledge about their clients' lives, and operate under confidentiality. Keeper, by contrast, is a consumer-facing app designed for scale. The mechanics of one-to-one concierge matchmaking don't translate cleanly to a product that must serve thousands—or millions—of users with an automated model.

The bias problem no one wants to solve

The most significant unanswered question is whose standards of attractiveness Keeper's AI has learned. Machine learning models trained on image data reflect the biases present in their training sets. If the model learned from datasets skewed toward certain ethnicities, body types, or age ranges—and most publicly available datasets are—it will replicate those preferences at scale.

This isn't hypothetical. Facial recognition systems have repeatedly been shown to perform worse on women and people of colour. Beauty filters and photo-editing apps have faced criticism for lightening skin tones and narrowing facial features. An attractiveness-scoring AI carries the same risks, with the added complication that users are being sorted and shown—or not shown—based on those scores.

Without transparency about training data sources and demographic composition, operators and observers alike are left to assume the model reflects the preferences embedded in whatever data was available—which, historically, means it likely skews toward Eurocentric beauty standards, younger faces, and slimmer body types.

Kozloski has disclosed that Keeper collected millions of human ratings to train its proprietary attractiveness model, but questions remain about the training data sources, the demographic composition of the dataset, and the steps taken to mitigate bias.

Regulatory scrutiny may force disclosure eventually. The EU's Digital Services Act (DSA) requires transparency around algorithmic decision-making for very large online platforms, and the UK's Online Safety Act (OSA) includes provisions around harmful algorithmic amplification. Keeper isn't yet operating at a scale that would trigger those thresholds, but if the product gains traction, compliance teams will need answers.

What this signals for the rest of the market

Keeper's positioning is a test case for whether users will accept—or demand—a more explicit appearance-first model. If the app gains meaningful traction in Boston and beyond, it will validate the hypothesis that compatibility messaging was always a polite fiction, and that users prefer transparency about what dating apps actually do: sort people by physical attraction and let them take it from there.

That would be an uncomfortable admission for Match Group, which has spent the better part of a decade repositioning Tinder toward relationship formation and investing heavily in Hinge's compatibility narrative. Bumble, too, has built its brand around women making the first move and connections grounded in mutual respect—language that sits uneasily alongside an AI that ranks users by how attractive they are.

If Keeper succeeds, expect copycats. AI matchmaking startups have been attracting significant venture capital, with Keeper raising $4 million in pre-seed funding in December 2025. If it fails, the question is why: because users rejected the honesty, or because the AI's attractiveness rankings felt arbitrary, biased, or just wrong. The latter would be the more instructive outcome for the industry.

Either way, the compatibility era may be ending. Not because the algorithms were exposed as ineffective—though they likely were—but because a Y Combinator-backed app decided that pretending they mattered was no longer worth the effort.