Why we removed profile photos from discovery

In 2024, Hinge's data science team published a finding that should have rewritten the consumer social playbook. Likes given on a written prompt were 47 percent more likely to convert to an actual first date than likes given on a photo. The company kept the photos. The market kept swiping. Both are still losing users.

Three structural facts produced that result. The first is that judgments of likability, trustworthiness, and competence from a face stabilize within one hundred milliseconds, and additional viewing time only adds confidence, not accuracy. The second is that grids of faces, presented in rapid succession, produce a documented rejection mind-set within a single session. The third is that what we evaluate when we look at a face has almost no predictive validity for whether we will actually like the person.

These three facts are not opinions. They are the load-bearing findings of three decades of social and behavioral research. They have a specific consequence for any product whose discovery feed leads with the face: the product becomes a dating product. The user's intent does not interrupt the cognitive system. The platform's stated category does not interrupt it. The only thing that interrupts it is removing the face from the discovery surface entirely.

For HAOT, that is what we did.

The hundred-millisecond decision

The cognitive system does not deliberate when it sees a face. Princeton researchers Janine Willis and Alexander Todorov ran a series of experiments in which subjects viewed faces for 100 milliseconds, 500 milliseconds, and one second, then rated those faces on attractiveness, likability, trustworthiness, competence, and aggressiveness. The 100-millisecond ratings correlated almost perfectly with the unlimited-time ratings; additional exposure produced no new information, only more confidence in the conclusion already reached at one tenth of a second^[1].

In practice, this means that any product where the user's first cue is a face has finished its identity assignment before the user can read a single word of context. Eye-tracking studies of mock dating profiles confirm the same thing at the interface level. In a 2022 study published in Communication Research, faces attracted initial fixation regardless of attractiveness, and more attractive faces received both more fixations and longer ones. Text functioned as a secondary cue that arrived after the face had already produced its impression.

The interface time scale extends only slightly beyond the perceptual one. Behavioral analysis of Tinder swipes has found that women take about three seconds to swipe right on attractive men and about seven seconds to swipe left on unattractive ones, with men averaging five to six seconds regardless. None of these durations is enough to read or reason about anything. The product has completed its decision in roughly the time a heart takes to beat four times.

The halo and the grid

What one hundred milliseconds buys is not just an attractiveness rating. It is a halo. Karen Dion's foundational 1972 paper called it "what is beautiful is good": raters shown a photo and asked nothing else about the person assigned the more attractive face higher scores on sociability, competence, occupational success, and marital happiness. The effect has been confirmed every decade since.

A 2024 replication in Royal Society Open Science used AI-beautified images and found the halo intact: attractive faces, including faces algorithmically enhanced to be attractive, were rated as more intelligent, trustworthy, sociable, and competent than the same person's unenhanced photograph^[2]. The halo is not a curiosity of mid-century psychology. It is a structural feature of how human cognition processes faces in the absence of other context.

The halo also concentrates. A 2018 study in PLOS ONE reported that inter-rater agreement on attractiveness, averaged across multiple judges, has a Cronbach's alpha of around 0.93. That is the same reliability score as a professional psychometric instrument. A photo grid does not randomize taste. It concentrates attention on a narrow consensus subset of faces that nearly all raters agree are the most attractive.

OkCupid's own internal data, published by its co-founder Christian Rudder in 2009, showed this concentration in production. The top third of women on the platform received roughly two thirds of all messages sent by men. The most attractive women received about five times more messages than average women. Women, meanwhile, rated about 80 percent of men as below average in attractiveness, and the most attractive men received about eleven times more messages than the lowest. The grid did not distribute attention. It funneled it.

Once the grid is funneling attention to that narrow subset, the user enters a documented psychological state. Researchers Tila Pronk and Jaap Denissen, in a 2020 study published in Social Psychological and Personality Science, found that online daters became 27 percent more likely to reject prospects from the first profile in a session to the last^[3]. The mechanism is what Eli Finkel calls assessment mode: with many profiles to compare against, the user starts treating each individual encounter as a screening exercise rather than as a person.

What Eli Finkel proved

Eli Finkel's 2012 paper in Psychological Science in the Public Interest remains the most cited academic critique of online dating in the field; its central argument is that browsing profiles is fundamentally different in kind, not in degree, from meeting in person^[4]. The browse mode produces a mental simulation of the person, assembled from their photo, their bio, their height, their answers, that bears almost no relationship to the actual person the user later encounters in the world.

In 2017, Samantha Joel, Paul Eastwick, and Finkel pushed the critique further. They ran machine learning on more than one hundred self-report variables from participants in speed-dating studies and asked whether any combination of those variables could predict who would actually be attracted to whom on the night; the answer was no^[5]. Stable individual preferences, including stated preferences for specific physical features, explained essentially none of the variance in real chemistry. An algorithm with full access to everything a person says they want in a partner could not do better than chance at predicting whether they would feel that pull when they met someone in person.

For a dating product, this finding is damaging. For a friendship product, it is more so, because the entire premise of leading with a face is harder to defend in the first place. Photos in a dating app at least gesture toward a biological signal that has been selected for over evolutionary time. Photos in a friendship app have no such signal to draw on. They simply import the assessment mode of dating into a category that does not require it.

The dating industry is publicly conceding

The most powerful argument against photo-led discovery is that the operators of photo-led products are publicly conceding it does not work. The Hinge finding that opened this piece was published in a January 2025 newsroom release by Hinge Labs.

This is the operator of a photo-included dating product reporting that text-led likes outperform photo-led likes on the single conversion event that matters: the in-person meeting. The number was not a marketing slogan. It was a structured comparison of two engagement pathways inside the same app. Hinge kept the photos because removing them would be unrecoverable from the user base they have already trained. But the data is unambiguous about which discovery pathway produces real-world conversion.

At the parent company level, the broader market is making the same admission. Match Group's Q3 2024 earnings reported Tinder monthly active users down 9 percent year over year and payers down 4 percent, with a la carte revenue down 13 percent^[6]. Match Group's share price is down approximately 70 percent since its spinout from IAC in July 2020. The market is pricing the structural decline of photo-led discovery.

Hinge, the one growing app in the Match portfolio, runs a prompt-first interface, one profile at a time, no grid. Its revenue grew 23 percent year over year in Q1 2025. Its design philosophy is published under the slogan "designed to be deleted." The economic outcome is not subtle.

And on the friendship side specifically, the most instructive failure is Bumble BFF, because it ran the dating playbook on a friendship use case and the playbook did not transfer. BFF was a swipe-and-match product. A grid of faces, a binary judgment, a single one-to-one match at the end of it. Morgan Stanley analysts estimated its revenue at around one million dollars in 2023, set against a parent company generating hundreds of millions. The following February, Bumble's CEO Lidiane Jones told analysts the company had "been slow to realize this broader vision with Bumble BFF thus far," and announced a move away from the swipe-and-match paradigm for friendship.

The lesson is precise, and it is close to the opposite of the one a casual reader might take from it. Bumble BFF did not fail because friendship is not a business. It failed because photo-led, swipe-to-match discovery is a dating mechanic, and a dating mechanic does not produce friendships. The architecture was imported wholesale from a category it did not fit. That is a design failure, not a verdict on the category. The demand for new friendship in adulthood is real and large. What has not worked is routing that demand through an interface built for romantic selection.

What replaces the photo grid

For HAOT, the conclusion follows from the data. Discovery is the activity itself, not the person. When a user scrolls the feed, what they see is the plan: the time, the location, the host's reputation score, the current group size, the type of activity, the language it will be conducted in. The face does not appear. It does not need to.

After RSVP confirmation, photos appear inside the activity chat, because at that point the user has committed to meeting a specific group of people for a specific shared experience. The face arrives after context. It does not lead context. This is the same logical inversion that Hinge has been moving toward at the prompt level, and that S'More, a small dating app launched in 2020, executed in its purest form by blurring all photos and unblurring them progressively only as users exchanged messages.

The honest counterargument is that some users will leave the product because the discovery feed does not give them the visual cue they have been trained to expect from a decade of social products. This is correct. The users we lose are the users who would have used a friendship product to filter people by appearance. They were never the target.

The other honest counterargument is that we have gone further than any major comparable. Hinge kept the photos and reduced the grid. S'More blurred the photos and unblurred over time. We removed the photos from discovery altogether. The aggressiveness is the point. We wrote about this earlier in the journal, in the piece on why anti-dating is architecture and not policy. Removing the photo grid from discovery is the third of four structural layers that make HAOT mechanically incapable of becoming a dating product. The other three are the absence of one-to-one messaging before a shared activity, the three-person group minimum, and real-time intent classification on chat.

The market has the data. We just chose to read it. The other structural gaps this category shares, the ones that survive even after the photo grid is removed, are the subject of the next piece in this journal.