


25 Jun 2025 · 15m · 6 chapters

AI “clones” trained on an individual’s digital traces could be used to match people for romance, jobs, housing, childcare, tutors, or therapists; the episode explains a research paper’s limits on algorithmic search when clones imperfectly represent real humans.
No guests are mentioned in the transcript; it appears to be a host-led episode.
AI matching optimizes for clone similarity, not actual human similarity, creating an approximation gap. The paper defines “AI equivalent sample size” and claims the AI’s advantage is always finite (Proposition 1). In high-dimensional personality spaces, “the power of two” (Theorem 1) says meeting just two people can outperform infinite AI search. Data disparity creates bias: subjects are more likely to be matched with “data-rich” individuals (Proposition 2), especially as dimensions grow and clone accuracy gaps widen.
If clone noise variance is about 0.05 and the AI searches 10,000 candidates, in-person matching with two people can beat the AI when personality dimensions reach around 150+.
Written by AI. May contain mistakes.
Discussion on how AI can mimic individuals and its broader applications beyond dating.
“Can AI's massive search power actually overcome the fact that it can never perfectly represent a complex human being?”
Exploring the differences between AI representation and traditional in-person matching.
“They use a really clever simplification.”
Examining the limitations of AI in matching and the surprising advantages of human interactions.
“But, and this is absolutely crucial, the paper highlights that these AI clones aren't perfect representations.”
Analyzing how data-rich and data-poor individuals affect AI matching outcomes.
“But then comes theorem one, the power of two.”
Discussion on the potential societal implications of AI-driven matching systems.
“For instance, if the AI's noise level, they use sigma squared for variance, is about 0.05, which isn't huge, and the AI searches 10 ,000 candidates.”
Explore the challenges of algorithmic efficiency versus human intuition in AI matching.
“Yes, the emergence of data-rich and data-poor groups, leading to a systemic bias where those with better AI representations get preferentially matched.”

Get the gist, explore an idea or ask about something said.
On Best AI papers explained, what are the key points of this episode?
Automatic transcript. May contain errors.
0:00Welcome to The Deep Dive, the show that cuts through the noise, distilling the most important knowledge from a stack of sources, research, and notes, all tailored just for you. Great to be diving in. Today we're tackling something that honestly feels a bit like science fiction, but it's getting closer all the time. What if an AI could, well, perfectly mimic you using your personal data? Right, these digital doppelgangers. Exactly. And then what if these AI corns were used to find your perfect match? Not just for romance, but maybe a job or finding the right house, even a therapist. It's a really fascinating and maybe slightly unsettling idea.
0:37And it's actually the core of a new research paper we're exploring today. Okay. We're going to unpack how they model personality and compatibility and look hard at this central question. Can AI's massive search power actually overcome the fact that it can never perfectly represent a complex human being? And I get the feeling the answer isn't straightforward. Oh, it's definitely more surprising than you might initially think. There's some real twists here. All right. So to really get started, let's define our terms. What exactly are these AI clones the paper talks about? How do they work? Well, you know, with the huge advances in large language models, the technology exists now to train an AI on an individual's data they're writing, maybe their social media, other digital traces.
1:22OK. And these models can get remarkably good at mimicking how that person communicates, maybe even how they might react in certain situations. These are what the paper calls AI clones or sometimes AI agents or AI delegates. They're like digital stand-ins. And you mentioned it's not just dating because my mind immediately goes there, AI matchmaking, but it's broader. Oh, much broader. Yes, platforms are definitely experimenting with this stuff already, trying to predict compatibility using AI. I bet. But the paper really pushes the idea further. Imagine using AI clones to simulate job interviews matching AI managers with AI applicants to see who clicks.
1:59Wow. While we're using them to find the best match for child care or a therapist or even a tutor. Things that require quite a bit of nuance. Even finding a home. That seems so subjective. Potentially, yeah. If the AI clone can capture your preferences accurately enough, it could theoretically sift through options much faster than you could. So the big appeal is scale, right? Right. The ability to check compatibility across a massive pool of potential matches way more than you could ever meet. Exactly. But that brings us right to the core tension, the central problem they explore. Which is? Unlike, say, a perfect biological clone, which is identical, an AI clone is always just an approximation.
2:39It's based on limited data, however much you feed it. Okay. And the thing is trying to approximate. Yeah. A whole human personality. That's incredibly complex, maybe infinitely so. So how do they even model that? Personality is so fuzzy. Yeah, it's a challenge. They use a really clever simplification. They model personality as a point in a multidimensional space. Think of it like a map with many, many directions. Okay, like north, south, east, west, but loads more. Exactly. Each dimension represents some attribute, maybe intellectual curiosity on one axis, energy level on another, maybe even your relationship to time or punctuality, sense of humor.
3:16You name it. So my personality is one specific point on this giant map. Precisely. You, the subject, are at the center of your own map, metaphorically speaking. And you ideally want to be matched with someone whose point is as close to yours as possible on their map. Minimal distance means better compatibility. That's the idea. Minimal Euclidean distance in this multidimensional space. It's a mathematical way to think about similarity across many traits. Okay, that makes sense as a model. So how do they compare the old way of meeting people to this new AI way? They set up two distinct scenarios, or search regimes as they call them.
3:53First is the in-person regime. The way we do it now? Pretty much. You randomly meet a certain number of people. Let's call it MIM individuals. Maybe you go on MIM dates or interview MIM candidates. And out of that limited pool, you pick the one who seems most compatible. Simple enough. What's the second regime? That's the AI representation regime. Here, everyone gets an AI clone. Your clone interacts with potentially millions of other clones on a platform. And the platform finds the clone that's closest to your clone on that personality map. Exactly. It identifies the AI clone with the minimum distance to your AI clone.
4:26Yeah. But, and this is absolutely crucial, the paper highlights that these AI clones aren't perfect representations. Right. The approximation issue. Yes. They model the AI clone as your true personality point, plus what they call a multivariate Gaussian noise term. Think of it as a slight random error or fuzziness added in multiple directions on the map. So the clone is always a little bit off in random ways. Inevitably, yes. And here's the kicker. Your actual satisfaction, your payoff from the match, isn't based on how close the clones were. It's based on how close the actual human you get matched with is to you.
5:01Ah, I see. So the AI might match the clones well, but if the noise made the match look better than it really is between the humans? You might end up disappointed. The platform optimizes for clone similarity, but you experience the reality of human similarity. And that disconnect, well, that's where things get really interesting. Okay, so this leads us into the second part, these surprising limits of AI search. You mentioned something about the power of just two encounters. Yes, this is probably the most eye-opening part of the paper. They introduced this concept called the AI equivalent sample size.
5:36What does that mean? It's basically asking, how many people would you need to meet in person, randomly, and people, to get the same expected quality of match as using an AI platform that can search over an infinite number of candidates represented by imperfect clones? So it's like a benchmark. How many real meetings does it take to beat the AI? Exactly. It's a yardstick comparing human effort to algorithmic search, given the AI's inherent flaws. And their first big finding about this benchmark? Is that it's always finite. Proposition one in the paper. No matter how small the AI's error is or how sophisticated it gets, there's always some finite number of in-person meetings that will, on average, yield a better result than even an infinitely large AI search.
6:21So the AI's advantage isn't limitless. Its value is inherently capped because of that approximation error. Precisely. You can't just overcome imperfect representation by throwing more and more data or search power at it indefinitely, at least not in this model. Okay, that's already quite profound. But then comes theorem one, the power of two. That sounds wild. It really is. This is the big aha moment. They find that when personality is sufficiently high dimensional, meaning compatibility depends on a large number of different attributes, that AI equivalent sample size, it drops to just two. Two.
6:54As in meeting just two people randomly in person gives you a better expected outcome than an AI searching an infinite pool of candidates? Under the conditions of high dimensionality, yes. That's the finding. That seems totally counterintuitive. How is that even possible? An infinite search. I know, it feels wrong at first. But it boils down to the weirdness of high-dimensional geometry. Okay, explain the weirdness. So as you add more and more dimensions to that personality map, moving from, say, 10 traits to 100 or 200 or more random points, representing unique personalities tend to spread out very far from each other.
7:29So everyone becomes more unique, more isolated in this high-dimensional space. In a sense, yes. They become increasingly unique. Now, you might think a huge AI search pool would be great for finding those rare matches in this sparse space. Yeah, that makes sense. More chances to find that needle in a haystack. But here's the catch. Yeah. In these high dimensions, the AI's noise, it's small approximation errors, starts to have a much bigger impact relative to the actual differences between people. How so? Because everyone is already quite far apart, the closest clone the AI finds might be close, mainly because of the noise, not because the underlying humans are genuinely that compatible.
8:07The noise can essentially create phantom closeness. Ah, so the error starts to dominate the signal. Exactly. The paper shows mathematically that for a large number of dimensions, the negative impact of this noise outweighs the benefit of searching a larger pool. The AI gets lost in its own inaccuracies. Do they give an example of when this happens? Like how many dimensions are we talking about? They do. They use some estimates in figure two. For instance, if the AI's noise level, they use sigma squared for variance, is about 0.05, which isn't huge, and the AI searches 10 ,000 candidates. Which is a lot.
8:42A decent pool, yeah. They find that the in-person search with just two people actually starts performing better than the AI once the number of personality dimensions gets to around 150 or more. 150 attributes needed for compatibility. Is that realistic? Well, think about all the subtle things that make someone compatible. Shared values, humor, life goals, communication style, habits, quirks. It's plausible that true deep compatibility might depend on hundreds of these nuanced factors. So if compatibility is that complex, meeting just two people beats 10 ,000 AI matches. Wow. It's a stark result.
9:16And it raises a really interesting question. How does this contrast with other areas where AI clearly excels, like metal diagnosis or game playing? Right, where AI often outperforms the best humans. Why is this different? The paper suggests it's about the nature of the problem. In something like medical diagnosis from scans, human intuitive knowledge might be limited. We rely on learned patterns. AI with vast data can find patterns we miss. But for something like personal compatibility, finding a friend, a partner, maybe even a good colleague, humans have this incredibly rich, evolved, intuitive understanding.
9:53We process a vast amount of subtle social cues subconsciously. We have a built-in compatibility detector, even if we can't always explain it. Kind of, yeah. And the argument is that for these kinds of subjective, complex goals where human intuition is already highly tuned, AI's algorithmic approach, especially with its inherent imperfections, might hit a wall much sooner. The conclusion about AI versus human evaluation becomes very different. That distinction makes a lot of sense. So the AI's limitations are technical, tied to approximation and dimensionality. But what happens when we layer on real-world complexities, like not everyone has the same amount of data available about them?
10:31That's exactly where the researchers go next. They extend the model to consider data disparity. What happens if some people have tons of data for training their AI clone, making it quite accurate, while others have very little? Creating data-rich and data-poor individuals, as they call them. Precisely. The AI clones for the data-rich group are simply better approximations of the actual person. And how does that affect matching? Does it create a bias? It absolutely does. This is proposition two. The paper finds that you, the person looking for a match, the subject, are more likely to be matched with someone from the data-rich group.
11:07The probability is strictly greater than 50-50, even if the underlying people in both groups are equally compatible with you on average. So just having a better clone makes you more likely to be picked by the algorithm. Yes. The algorithm sees the data-rich clone as a clearer, potentially closer match. even if the underlying person isn't necessarily better for you than someone from the data poor group whose clone is fuzzier. And is this bias significant or just a slight edge? It can become overwhelming. They have two corollaries showing this bias actually converges to one, meaning you become almost certain to match with a data rich group under two conditions.
11:46What are they? First, as the number of personality dimensions gets very large. So the more complex the personality model. The worse the bias gets. Right. And second, as the difference in accuracy, the approximation error between the data rich and data poor clones gets very large. So bigger data gaps lead to much stronger bias. Exactly. The more nuanced compatibility becomes or the wider the digital divide in terms of data representation, the more the system favors those who are digitally louder or clearer. Okay, let's unpack the implications here. What does this mean for society if these AI matching systems become widespread?
12:20Well, it points towards a potentially significant new form of social inequality, or maybe a new driver of inequality. How so? It suggests that individuals who are better understood by AI systems, perhaps because they naturally generate more data, are more willing to share it, or belong to groups better represented in training data sets, will be systematically preferred in these mapping processes. So opportunities, jobs, housing, maybe even relationships facilitated by these platforms could flow towards the data rich. That's the concern, yes. It could amplify existing inequalities. If, for example, majority groups or more digitally active populations tend to be data rich, they might gain an advantage in matching, even if their intrinsic qualities aren't superior.
13:05It creates a kind of social stratification based not just on who you are, but on how well an AI can perceive who you are. That's a great way to put it. Your opportunities could increasingly depend on the quality of your digital delegate, your AI clone. It adds a whole new layer to think about regarding fairness and access in an AI-driven world. This has been, well, a really fascinating and slightly sobering deep dive. Let's try and recap the main takeaways. Sure. I think the first big one is that while AI search is powerful, its inability to perfectly capture complex human traits like personality puts a fundamental limit on its value and matching.
13:40Right. The approximation error really matters. And surprisingly, especially when compatibility is highly complex, high dimensional, that limitation means meeting just a couple of people in person can actually beat an AI searching infinite candidates. The power of two. A truly striking finding. And then critically, we saw how real world data disparities can warp this. Yes, the emergence of data-rich and data-poor groups, leading to a systemic bias where those with better AI representations get preferentially matched. Which could create or worsen social inequalities based on how well AI sees us.
14:15Exactly. It highlights this potential for a new kind of stratification driven by data and algorithms. So it leaves us with a big question, doesn't it? In this world that's increasingly leaning on AI for efficiency, especially in finding ideal matches that are often so subjective and nuanced, how do we strike the right balance? How do we weigh that incredible algorithmic efficiency against what seems like an undeniable, maybe irreplaceable value in human intuition and just actual in-person connection? That really is the crux of it, isn't it? Yeah. What stands out most to you as you think about navigating that future?
14:50It's something we'll all likely grapple with more and more.
This academic paper "Artificial Intelligence Clones" explores the effectiveness of "AI clones" in matching individuals for various purposes, such as dating or hiring, compared to traditional in-person interactions. The author models personalities as points in a multi-dimensional space and AI clones as noisy approximations of these personalities. The central argument is that while AI platforms offer vastly expanded search capacity, the inherent imperfection of AI representations ultimately limits their utility. A key finding suggests that meeting even a small number of people in person can yield better expected matches than searching an infinite pool via AI clones, especially as personality complexity (dimensions) increases. Furthermore, the paper highlights a potential social stratification where individuals with more readily available personal data for AI training might be systematically favored.