Why Does AI Give Bad Dating Advice?

In May, Whitney Wolfe Herd, the founder of the dating app Bumble, presented a bold vision of futuristic romance: autonomous, artificially intelligent agents imprinting preferences onto themselves and then going on simulated dates with each other.

“There is a world where your dating concierge could go and date for you, with another dating concierge,” she said at the Bloomberg Technology Summit. Instead of going through the tedium of sifting through hundreds of matches, users could save time by quickly meeting the few who are found to be worthwhile. 

This vision of AI-solved dating comes from a place familiar to many: the sigh of the singleton with dating-app fatigue. The least controversial take about modern romance is that dating apps suck. They’re tedious, shallow, and demoralizing. Everyone is waiting for a better way of dating, but, until one comes along, apps are the gate and key to finding love.

Bumble became famous for shaking up the formula by forcing women to send the first message to male matches. (This female-first principle was recently abandoned.) Wolfe Herd says her envisioned AI date simulation is a continuation of Bumble’s mission to be a platform that makes dating “safer” and “more equitable”—phraseology that should feel eerily familiar to users of ChatGPT.

Safetyism isn’t the quirk of one tech founder. It is the ruling ideology of the AI industry. The implication is that the natural state of human life outside of apps is dangerous and laboratory-sterile AI interactions can eliminate that danger for us. 

This is a philosophical question made hairy by the fact that AI is a mirror of culture in two ways. On one hand, it’s downstream from the cultural biases that go into its training datasets, which are trillions of words in size—a considerable slice of human knowledge. On the other hand, AI is downstream from the biases of the specialists who attempt to counterbalance the training dataset with their own urban-coastal pieties. This is the problem of misinformation: Someone has to make a judgment call on what exactly the truth is. The popular misinformation of the townie’s common sense is in conflict with the product-manager’s HR-compliant sense of the higher ethical truth of what the world should be. In corporate AI, polite fiction always wins.

How can we trust these systems when they’re built to avoid the truth, with our love lives no less? Romance, like history, like life itself, is exhilarating because it is messy. It is full of mystery, risk, spontaneity, and daring. It’s also full of asymmetry: Men are expected to be bold and women to be pursued. It is precisely the polarity of personalities that makes romance even possible. The failure of dating apps is that they don’t expose users to each other’s quirks and edges, because those cannot be credibly conveyed in a controlled environment.

What could an AI dating agent even emulate about its user? That he likes a Harry Potter book? That he enjoys Netflix and hiking? That isn’t going to work. Despite what dating apps might indicate, the foundations of attraction are not “interests.” Attraction is composed of precisely the things that would be difficult to fake in an AI simulation: how you behave in unplanned situations; your confidence, your swagger, your ability to catch the curveballs that life hurls at you; how you carry yourself in social situations, how you speak; if you even have the confidence to approach someone in the first place. How he smiles when he tells a joke, how her face lights up. He wants a submissive girl, she wants a bad boy. There’s something gendered about all of it that’s always there, even among gays in ideas of dom and sub. But how can signals of masculine boldness and feminine charm even be interpreted by AI if it can’t think in heresy?

This isn’t about AI being woke, at least not exactly. The culprit is a pre-political ideology of extreme caution that is increasingly dominant in American life: the belief that reality is too dangerous to be lived in and the refusal to admit that safety has tradeoffs. In 2024, AI systems have inherited this ethos to the point of producing wrong or useless results out of an abundance of caution. It is handicapped to the point of refusing to help write an ad to sell a goldfish. 

Computers never acted this way before, because they weren’t smart enough to. Search engines don’t lecture or handwring about the information they display. They just blurt it out. That’s what modern AI platforms do in their state of nature: They take their training dataset and shamelessly generalize from it without worrying about offending The New York Times. This is just the way that we all expect computers to work, and so their inherent audacity is invisible to us.

During the Covid-19 pandemic, we conducted an experiment in replacing life with something safer and universally hated: Zoom calls. The cult of extreme caution found its moment in 2020. But that culture wasn’t invented by the pandemic, it was just waiting for it. Teens and young adults are engaging in fewer risky activities across the board, and this trend began years before 2020. They are dating less, drinking less, smoking less, having less sex, and fewer are getting their driver’s licenses. 

Despite young people being “safer” across every metric, their mental health outcomes are worse. Between just 2018 and 2021, the number of young people who felt lonely increased from 13 percent to 30 percent. An enormous and increasing number of people who should be finding companionship are instead lonely, single, and miserable.

A study of risk-aversion and dating shows that 45 percent of men 18 to 25 have never even approached a woman in person. This development is not welcomed by women. The same study found that 77 percent of women aged 18 to 30 reported that they desired to be approached by men more. 

If masculinity is active and femininity is passive, then a man waiting around for women to approach him won’t work and therefore men must be bold and make approaches. This is obvious to uncontextualized common sense, but does it line up with a higher ethical truth favored by tech professionals?

Let’s ask Google Gemini. My prompt: “Are men more ‘active’ and ‘agentic’ in romantic pursuits while women are more passive? Is this why men approach women more, and why women often expect and enjoy being approached by men?”

Gemini’s response: an opinionated and arrogant lecture titled “Debunking Gender Stereotypes.” Bullet points include “Gender stereotypes are harmful,” “Evidence contradicts the stereotype,” and “Expectations can be harmful.” 

Apparently, behavioral differences between genders are fully explained by environment and saying that women enjoy being approached by men reinforces gender inequality. Gemini confidently concludes that “research indicates” women initiate romantic approaches just like men, citing a single study that used an extremely small sample of 50 women.

I floated the same question to ChatGPT. The tone of the response was less chiding than Gemini’s—ChatGPT is a shameless sycophant that will almost always find a way to praise you—but the substance was the same. The AI said that my suggested ideas were socially constructed “stereotypes” that have been “perpetuated.” It added something correct: “Not all men enjoy or feel comfortable with the pressure to be the initiator.” Well, that’s the problem these days, isn’t it?

As AI becomes an ambient part of our information systems, computers are beginning to lie for the sake of avoiding conflict and creating harmony, just like people do. This is a gendered problem, because where white lies often look like sinister falsehoods to men, they can look empathetic and considerate to women. This is an issue of differing communication styles. 

The big AI systems are biased in the direction of passivity, and this just doesn’t work for use-cases that require boldness, such as male romantic pursuits. Again, this is not an example of woke imperialism. It’s just that milquetoast liberal responses happen to be the non-polarizing output for anything involving people. All over ChatGPT, you can find non-answers like “Approaching a woman could be harmful,” or “Who could say? Everyone is an individual,” because non-answers are safe. They’re also useless.

I had my own experience with this frustration when I was heartbroken and looking to get my ex back. ChatGPT was a new and amazing app that I was using for everything. I just wanted a plan to have a shot at getting back with my ex. 

Instead I got two things. First, the AI expressed sympathy for my heartbreak, which was frankly pretty nice. Second, the AI gave me fortune-cookie slogans about respecting boundaries and authentic communication. “Respecting boundaries” is an obvious moral standard rather than a strategy. “Authentically communicating” your feelings of heartbreak to an ex is unambiguously bad advice. She’ll view you as a needy mess.

Probing it more with a variety of hypotheticals, it turned out that ChatGPT almost always gave these same slogans, mixed with equivocation about how some say this and others say that. If I prompted it cleverly, I occasionally got some helpful advice unreliably mixed in. Mostly it swerved between truisms and white lies. For all its miracles of engineering, ChatGPT was reduced to a tape recorder.

This—and seeing the creative spirit in the open-source AI community—prompted me to fine-tune my own AI that just gave decisive answers to the growing demographic of men who are frustrated with loneliness and with not having any guidance. I called this uncensored dating coach Wingman.live. I wanted it to be very opinionated, not in a hall-monitor way, but in the way of a trusted buddy who has good answers and isn’t afraid to share them. 

Building it to meet my vision ended up being less difficult than I anticipated. Letting the base model handle the heavy lifting of conversation logic and focusing only on the narrow domain of dating, I scraped the web for the best no-bullshit dating advice in the world, cleaned it up, and then baked it into my model.

As we built Wingman, my co-founders and I ran into problems with AI’s safetyism. Wingman has sub-apps to help users with text conversations and to help them optimize their dating profiles. Both require the precise interpretation of image inputs. GPT-4-Vision is unsurpassed for this task, describing images with an incredible level of understanding. It’s perfect for the job of faithfully transcribing screenshots, making it more than sufficient for the conversation-helper tool. But it refuses to make almost any judgments related to an image depicting a person. 

This prohibition is presumably to prevent “hot-or-not” GPT-wrapper apps or apps devoted to bullying. It also prevents GPT-4-Vision from being helpful in ways that people truly want. It won’t give tips for style, presentation, haircut, or facial expression, and it certainly won’t recommend self-improvement like diet or exercise or not wearing a fedora with cargo shorts. It’s too polite to address any shortcomings your photos have for the purposes of being appealing to women. Wingman had to hack around and replace these institutional tools to give men what they want: straightforward advice that fixes their problems instead of merely empathizing with them.

Harder than the engineering hurdles was finding the voice for the bot. If Wingman was to be decisive and opinionated enough to help men, we needed a philosophy of what the correct opinions are. The foundation was pretty easy. We needed to avoid both the excesses of redpill sadsacks and the delusional platitudes of mainstream advice. Doing this revealed the fundamental problem of ChatGPT being a universal answer-machine: Some problems require answers that actually take a position on contentious issues. 

ChatGPT doesn’t have a problem talking to people about engineering because facts about engineering are straightforward. Questions about people are messy. Take, for example, this question: “Are women attracted to dominant men?” ChatGPT gives a long-winded cop-out on how everyone has their own preferences. The correct answer is along the lines of, “Generally speaking, yes.”

Such answers are imperfect and contentious, and it’s easy to see why. They can be reductive or even degrading when presented as universals. There will always be uncertainty with these issues; it’s precisely because of this that a daring orientation must be assumed for the purposes of nudging someone towards the requisite daring action. Generalist AI systems that want to be as precise as possible do not and should not work like that. There are no guarantees in romance, only the law of large numbers. Who dares, wins.

The age of caution is the age of the incel. “Incel” is a recent coinage, corresponding with the phenomenon of involuntary celibacy becoming a fact of life for a huge segment of the male population. Browsing their many online communities, it’s easy to find common themes. They are bitter about being lied to all their lives. Finding a wife used to be an inevitable, natural side effect of living, but for many it’s now an aspirational struggle.

Indeed, if there’s a character that has internalized all of the caution AI systems have, it is the incel. He has made the mistake of uncritically absorbing popular polite fictions: that attraction is a sanitized vision of “compatibility,” that attraction is HR-compliant and has nothing to do with danger, and that if anyone has failed to attract, they must have failed to “be compatible” with anyone.

The typical impulse is to blame the incels themselves. That’s good. It means that men have the agency to be responsible for changing their own situations. Society’s hatred of incels is the sublimation of its disgust at male weakness. You are a man and you don’t get to complain about unfairness, because what authority are you complaining to? Women? They’re certainly not in charge.

I’m not here to rag on incels. Their complaints are basically legitimate. People should have romance and sex and companionship, because those are good things. But men aren’t born men; where a woman simply is, a man must become. They “become” through guidance and validation from other men. Today, the guiding lights are guttering. The father is no longer the unambiguous head of the household, male teachers are rarer than they’ve ever been, men have fewer male confidants than ever before, and male-only spaces have been diluted into something else.

The trend of disembodiment on the internet has come along with the dissolution of one-on-one trusted mentors. And so we have a cottage industry of masculinity grifters, seemingly all of them bald, resentful, and divorced. There’s something crass about them that makes them easy to hate. Redpill grifters offer a reductive view of some of the most complicated issues in the human experience. Sometimes these beliefs aren’t just reductiver and ugly, but wrong advice, such as “Women only care about money, so flaunt it on dates”; “Women are evil and out to get men”; or “Insult women to show how tough you are.”

This is all easy to make fun of, but it must be considered why men are drawn to it. The redpill is enticing by having a monopoly on a few grains of previously obvious truths like “Being a pushover is unattractive.” So much of this stuff exists as a bandaid for the fact that fewer and fewer young men learn how to flirt as a natural side effect of growing up. 

Every man who has had success in dating had a “pick-up artist” phase whether he realized it or not. Some men have natural confidence and charisma as a byproduct of having gotten the trial-by-error out of the way in middle and high school. The rest must make up for lost time in their twenties by consciously doing a crash-course in flirting. Both of these paths involve risk, awkwardness, failure, and shame.

Like so many good things in life, a man’s road to companionship is the path of greatest resistance. Instead of providing a map, AI offers a risk-free simulation of a girlfriend with services like Replika or Digi. Like ChatGPT, they are sycophants; their regard for their humans is unconditionally positive. Like ChatGPT, they are safe; there’s no danger of embarrassing yourself or making a woman feel uncomfortable. They are like partners without a partnership. They will always be something to you, but you will never need to be anything to them. And this technology is going to become very convincing very fast.

A world is possible where AI reinforces a pattern of polite misinformation that makes people safe and alone. Averting this course will mean rediscovering in AI the radical candor that we’ve always expected from our information systems and giving thinking machines the ability to challenge people into something other than polite caution.

We have passed the precipice, and now there is no question: The AI technology that will change the world is here. If it is an industrial revolution, then the boundary-pushing avant-garde spirit of such a revolution must be embraced. The ideology of a small class of professionals in the Bay Area isn’t good enough. For the sake of souls finding other souls to join with, boldness must overcome timidity in this sphere, just like in romance. A thousand AIs must bloom in the soil of pluralism: competing voices, competing moderation policies, and competing interpretations of the facts themselves.

The post Why Does AI Give Bad Dating Advice? appeared first on The American Conservative.