Lab Talks #3: Inside the Science of Attraction: A Conversation with Dr. Igor Khalatian

In this edition of Lab Talks, we speak with Dr. Igor Khalatian, founder and CEO of Ideal Match Inc., the company behind Iris Dating and BestLook.ai. Igor’s work represents one of the most ambitious intersections of artificial intelligence and human behavior: using data to understand attraction. His approach combines evolutionary psychology, computer vision, and predictive modeling to uncover what truly drives mutual connection in a world dominated by algorithms and surface-level swiping.

At Dark Forest Labs, our goal with these conversations is to surface the kind of thinking that bridges science, product, and real human experience. Igor’s path, from Bell Labs researcher to serial founder and now a leading voice in AI-driven matchmaking, captures how technology can move beyond engagement metrics to measure something deeper: compatibility and chemistry.

About Dr. Igor Khalatian

Dr. Khalatian is an accomplished entrepreneur, inventor, and researcher with over two decades of experience in artificial intelligence and computer science. After earning his Ph.D. at Bell Labs in the United States, where he contributed to a landmark communications project that resulted in a U.S. patent, he went on to found several successful startups and secure thirteen patents of his own. His company LiveLOOK was acquired by Oracle in 2014, where he later served as Vice President of Development.

Today, through Ideal Match Inc., he continues to explore how machine learning can model human attraction with scientific rigor and emotional sensitivity. Beyond his professional pursuits, Igor is also a dedicated mountaineer who has climbed more than 200 peaks around the world, including Mount Everest.

The Interview

Q1. You’ve said that mutual attraction is one in a million. Can you walk us through how you arrived at that insight, both from data and from your own experience?

When I explain this, I start with a simple thought experiment we ran in our research. Ask yourself: “Out of a thousand people I pass, how often does one literally take my breath away?” Most honest answers cluster around one in a thousand. That’s just one-sided attraction, you to them.

But for a real match, it has to be mutual. Think of two roulette wheels spinning at once: one is your taste, the other is theirs. If each side is roughly one in a thousand, the joint probability is around one in a million. That’s why dating often feels like chasing a jackpot. Our large-scale data work, based on millions of labeled choices, reinforced the same bottleneck: strong one-sided attraction is rare; mutual is exponentially rarer. That’s exactly what our AI solves. It learns your personal “type” quickly, then optimizes introductions so you’re not relying on chance.

Q2. Looking back at the rise of dating apps, what do you think many of them fundamentally misunderstood about how to define success?

They are optimized for engagement loops such as volume, swipes, and time-in-app, not outcomes users actually want, which are fewer, higher-quality introductions that lead to real dates. If you define success as monetization, some apps were indeed successful. But from the user’s perspective, roughly 80 percent report dissatisfaction with the experience. Our goal is to flip the metric and measure success by mutual-attraction outcomes, not by how many times someone opens the app.

Q3. You often use the “NYC café spark” story to describe instant attraction. How did moments like that shape your decision to start building in this space?

That half-second, breathless response to a face taught me two things. The brain has a fast, personal “ideal” that fires instantly, and it’s rare. I wanted AI that could learn each person’s “type” from a few choices and then find those needles at scale so people don’t spend years swiping for a single spark.

Q4. In your research, you describe attraction as something deeply evolutionary, shaped by choices our ancestors made. What does modern AI allow us to uncover about that process that wasn’t possible before?

We can separate personal taste from popularity bias. With compact, high-signal feedback, models infer stable, individual preferences that hold over time, your unique “Attraction DNA.” That allows us to predict which faces you’ll feel drawn to, and likely vice versa, rather than chasing whatever is broadly popular this week.

Q5. You’ve compared AI-driven attraction models to Spotify’s recommendation engine. Why do you think that analogy resonates with people?

Because it’s predictive, not generative. A few thumbs-up or down quickly tune a model to your patterns, just like a music service refining a station. People already trust that feedback loop for songs; faces work in a similar way. The output isn’t “more profiles,” it’s a curated, paced feed that matches your personal taste the way a great playlist matches your ear.

Q6. When you scaled your ideas from research and patents into a consumer product used by millions, what were the biggest hurdles you had to overcome?

Three main ones: the cold start problem without fatigue, creating “taste training” that feels rewarding rather than like homework; building privacy-first data operations that learn stable preferences without hoarding unnecessary data; and managing expectations by teaching users that pacing and limited introductions are features that improve outcomes.

Q7. Early on, you’ve said that some users were confused when attraction was prioritized over filters like distance. How did you balance user expectations with your disruptive approach?

We kept “chemistry first,” but added clarity and control. Optional distance caps, explanations for why a profile was shown, and pacing so users didn’t get five high-chemistry matches at once. The principle stayed bold, but the user experience became gentler.

Q8. Across your different projects, you’ve often broken rules about how tech and human connection are supposed to work. What’s been the most surprising lesson so far?

Declared preferences collapse in the face of real chemistry. When we asked people in surveys whether they’d date someone far away, even if they found them very attractive, about 90 percent said no. But when we showed those same people an actual face they found highly attractive and asked the same question, 90 percent said yes. The takeaway is that logistics feel non-negotiable in theory, but a concrete, high-chemistry introduction changes the calculus instantly. Design for revealed preference first, then offer practical controls around it.

Q9. You’ve been both a founder and an executive after an acquisition. What were the toughest challenges you had to overcome?

Switching modes. Founder mode rewards speed and invention, executive mode demands systems and repeatability. I learned to separate exploration from scale, to protect the research loop so it stays bold, while building operational rails so the product ships predictably.

Q10. Can you share one product or business decision you got wrong, and what you learned from correcting it?

We once underestimated how much guidance the “training” step needs. Conversion dipped because people saw it as work. We rewrote the flow around unlocking “your type,” added micro-rewards, and made the process faster. The result was better engagement and stronger downstream outcomes.

Q11. You hold thirteen U.S. patents. How do you decide what’s worth protecting versus what you keep as trade secrets?

I patent the primitives, the frameworks and evaluation methods that define the space, and keep implementation optimizations as trade secrets. Patents anchor the narrative; trade secrets preserve competitive edge.

Q12. The media often describes dating apps as slot machines. How do you see AI reshaping the space, and what risks should we acknowledge?

AI shifts apps from engagement machines to outcome engines by predicting likely mutual attraction first. But we have to remain alert to new risks: popularity bias creeping back in, over-automation where bots start interacting with bots, and user overwhelm if pacing is ignored. Transparency, consent, and clear limits are essential safeguards.

Q13. As large platforms begin experimenting with AI features, where do you think they’ll stop, and where do startups have room to innovate?

Big platforms will likely stay near the surface with photo selection, moderation, and prompts because it’s low-risk at scale. The deeper opportunity for startups lies in the core loop: learning personal taste quickly, ranking for mutual attraction, and pacing delivery for measurable outcomes.

Q14. At Dark Forest Labs, our community brings together founders, engineers, and researchers working on AI in human connection. What insights do you think would help them most?

Differentiate clearly between predictive and generative systems. Know which one actually creates value for your product. And think ahead about what I call GEO, Generative Engine Optimization, structuring your data and content so that AI systems can understand and cite your work accurately.

Q15. If you were leading a Lab Talk here, what would you teach in 45 minutes, and what practical exercise would you give participants?

I’d call it “Designing for Mutual Attraction.” The exercise would be simple: in pairs, redesign an onboarding flow that learns a user’s “type” in under ninety seconds, then add one pacing rule that prevents overload. It’s a small design task that reveals a lot about human behavior and system thinking.

Q16. You’ve seen the myths and realities of building machine-learning products at scale. What are a few myths about AI in dating you’d like to debunk?

That more profiles equal better outcomes. It actually dilutes attention and lowers satisfaction. That AI should chat for users. It shouldn’t. AI should predict, not impersonate. And that there’s one universal standard of beauty. There isn’t. Taste is personal, and learnable, and that’s the point.

Q17. Finally, looking ahead, how do you see AI companions evolving over the next five years, both as a business and as part of daily life?

They’ll move from novelty to scaffolding, tools that remember, prompt, and support human relationships rather than replace them. In romance, this means fewer but higher-trust introductions, stronger verification, and a clearer line between assist and agency. The products that win will not maximize time-in-app; they will maximize the quality of outcomes.

Takeaways

Dr. Khalatian’s work reframes how we think about technology and connection. His concept of “Attraction DNA” shifts the focus from mass engagement to individualized understanding, suggesting a new foundation for how AI can mediate human relationships. His approach is rooted in measurable science yet remains deeply human, acknowledging that chemistry, trust, and timing cannot be reduced to clicks or metrics.

For founders and researchers in the Dark Forest Labs community, Igor’s perspective offers a model of how AI can move beyond novelty toward systems that optimize for meaning. It also reminds us that in the pursuit of smarter algorithms, the most important variable may still be the human one.

This conversation marks only the beginning. In the next Lab Talk, we will continue our dialogue with Igor to explore the technical layers behind predictive models of attraction and what they reveal about the evolving relationship between data, desire, and design.