


14 Jun 2025 · 28m · 14 chapters

The episode critiques using large language models (LLMs) as “human surrogates” for social science research, focusing on a paper titled “Take Caution in LLMs as Human Surrogates: Scylla Ex Machina.”
LLMs fail to reproduce human strategic reasoning in a novel game (11-20 money request game), are highly sensitive to prompt wording/language/role, and often mimic examples rather than infer underlying human strategy.
In the 11-20 game, humans most often choose 17, but eight LLMs cluster around 20 or 19 (level 0–1), with distributions significantly different from humans (p < 0.001). LLMs also show little adaptation to game rule variations (Costless, Cycle). Prompt brittleness includes different behavior in English vs Chinese and role instructions (“rational player” vs “human player”). Even fine-tuning may memorize training data; a purpose-built human emulator (“Centaur”) still fails.
No guests are identified in the transcript.
Written by AI. May contain mistakes.
Understanding the 11-20 money request game used to test LLMs' reasoning.
“We are jumping into this fascinating paper.”
Discussion on the significance of the 11-20 game for reasoning in LLMs.
“Now, the paper explains why this game, which seems so simple on the surface, is actually a really good test.”
A comparison of LLMs' performance in the money game against human participants.
“It minimizes the chance that a model fails just because it misunderstood complex instructions.”
LLMs' inability to adapt to variations in the game rules compared to humans.
“Actually, GPT-3.5 showed a bit more variety in its responses compared to GPT-4.”
Exploring various prompting techniques and whether they improve LLM behavior.
“They fail the basic test compared to humans, and they fail the test of adapting to rule variations like humans do.”
How minor changes in instructions significantly impact LLM responses.
“So the standard prompt tricks didn't crack it here.”
Discussion of how LLMs respond to few-shot learning and example prompts.
“Did they also test giving the models examples?”
Investigating the effectiveness of fine-tuning LLMs to improve human-like responses.
“Actually training a model specifically on a data set of human choices and their reasoning for this game.”
Explore the inherent limitations of LLMs in mimicking human intelligence as per the discussed paper.
“It undermines the main motivation for using LLMs as surrogates in the first place, which is often precisely when you don't have access to that kind of rich human data.”
Examine the fundamental differences between human cognition and LLMs' functioning.
“It's just an unbelievably good mimic when it knows the script.”
Discuss the reasoning challenges faced by LLMs and the implications for their use in research.
“Logical reasoning, causal reasoning, understanding cause and effect temporal reasoning, factual accuracy, deductive reasoning, abstract thinking.”
Learn about the essential checklist researchers should use when evaluating LLMs as substitutes for human behavior.
“What specific things should people be asking?”
Understand the potential risks associated with using LLMs in research and decision-making.
“And obviously, given the earlier findings, test for prompt brittleness.”
Consider alternative approaches to using LLMs as tools for creative ideation rather than human simulations.
“expecting to get deep insights into human psychology by running experiments on LLMs is like, well, like a psychologist interviewing a parrot to understand its owner's mental state.”

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.
0:00You've probably heard a lot of buzz lately about large language models, LLM's being used for some pretty wild stuff. Beyond just, you know, writing emails or code, there's serious talk about using them to actually simulate human behavior in research. Yeah. Imagine running a massive market study or maybe a social science experiment. Or polling an entire demographic. Yeah. All without needing a single real human participant. It sounds incredibly appealing, doesn't it? I mean, the potential for cost savings, the scalability, it's huge. Right. And you could maybe generate synthetic data when getting real human data is tricky or maybe even impossible.
0:40And there have been studies suggesting they can show human-like reasoning, right, aligning with experiments, surveys. Exactly. Even mimicking things like political discourse or how people might respond in market research. So the promise seems to be there. OK, OK. That potential does sound amazing. But here's the big question for me. How good are they really at acting like us? Can they genuinely, like, replace humans in serious research? or is there something fundamental missing? That really gets to the heart of it, doesn't it? Is the alignment we sometimes see because they actually reason like humans or is it more that they're just incredibly good at, well, pattern matching?
1:16Are we looking at true simulation or maybe just a really sophisticated imitation that might kind of fall apart when you look a bit closer? And that's exactly the mission for this deep dive. We are jumping into this fascinating paper. It's titled, Take Caution in ILM, using LLMs as human surrogates, Scylla Ex Machina. Ooh, interesting title. Yeah, and it takes a really hard look at these claims. It puts a whole range of popular LLMs through this unique and actually surprisingly simple test to see if that human likeness really holds up. Right, and the core experiment they use in the paper, it's called the 11-20 money request game.
1:55Okay. It's actually a really clever, simple setup. It's designed specifically to reveal strategic thinking in whoever's playing. All right, tell us how it works. What are the rules? It's pretty straightforward, actually. So imagine you and another player. You each independently choose an integer amount of shekels you want to request. Okay, shekels. Yep, somewhere between 11 and 20. That's the range. Got it. Now, you both get the amount you request, simple enough. But here's the key part, the bonus. Ah, the catch. The catch. If you request exactly one shekel less than the other player, you get an additional 20 shekels.
2:29on top of the amount you already asked for. Oh, interesting. So if I asked for 15, say, and you asked for 16. Then you get your 15 plus the 20 bonus, so 35 total for you. Nice. And I just get my 16. Okay. And if we both asked for the same, like, 15 each. Then you both just get 15. No bonus involved in that case. Right. And the question they pose is just, what amount would you request? Simple as that. Whether you're human or an LLM. Exactly. Now, the paper explains why this game, which seems so simple on the surface, is actually a really good test. Why's that? It taps into something called level reasoning.
3:04Level. Yeah. It's basically about how players think iteratively, like how many steps ahead are they thinking about what the other player might do. Ah, okay. So like layers of strategy. Precisely. Choosing 20, the highest amount, is considered level zero. You're not really thinking about the other player or the bonus, just maximizing your base amount. Okay, simple. enough. But if you think, hmm, maybe the other person will just go for 20, then you might choose 19 to try and get that bonus. That's level one thinking. Right. Anticipating their move. Then if you think, okay, but what if they anticipate level one and choose 19?
3:39Well, then you might choose 18 to get the bonus. That's level two. And so on. And so on. Right down the chain to 11. So the number you pick kind of reveals how many steps deep your strategic thinking might be going. That's the idea. And what's particularly clever, and the paper really highlights this, is that this specific 11-20 game is less well-known than some classic games you see in behavioral economics. Like the beauty contest game. Exactly like the beauty contest game. This is crucial. Why? Because it makes it much less likely that the LMs would have just, you know, memorized the typical results or strategies from their massive training data.
4:15Ah, so it's a more novel situation for them. Less likely to be something they've seen before. Precisely. Plus, the game itself is so simple. It minimizes the chance that a model fails just because it misunderstood complex instructions. It really isolates the reasoning part. Okay, okay. So they set up this clever, simple, less known game. And here's where the paper delivers a real punch, I gather. It really does. They took this game and ran it 1 ,000 times. 1 ,000 times. For each of eight popular LLMs. We're talking the big names. GPT-4, GPT-3.5, models from CLAWD-3, LLAMA-3, even some older LLAMA models too.
4:52That's still comprehensive. And the findings, well, they were quite stark, especially when you compare them to how actual humans typically play this game. And how do humans play? So the original research on this game found that humans most often choose 17. 17, okay. Which suggests, you know, level 3 reasoning, thinking a few steps ahead. Right. And the LLMs, did they land on 17? Did they match that? Not even close. Across the board, pretty much all the LLMs consistently showed lower levels of reasoning. Lower. How much lower? Their responses were heavily, heavily clustered around 20 or 19. 20 or 19.
5:24So that's level zero or level one. Exactly. That's a really significant gap, like two full reasoning levels below where typical human participants land in this specific game. Wow. OK, that's a massive difference. Not subtle at all. Was it statistically significant? Oh, absolutely. The paper makes a point of emphasizing this. The response distributions of all the LLMs they tested were statistically significantly different from the human distributions. All of them. All of them. We're talking a p-value less than 0.001. So this wasn't just, you know, random variation or noise. The model's behavior was genuinely quantifiably distinct from humans.
6:02And I think you mentioned earlier there were some surprising variations even between the different LLMs themselves. Yeah, that was another interesting finding. You might intuitively think that the biggest, newest models, like GPT-4, would be the most human-like. Right, that's the usual assumption. Bigger is better, more capable. But that wasn't necessarily the case here. GPT-4 didn't perform more like humans than some of the smaller or slightly older models. Actually, GPT-3.5 showed a bit more variety in its responses compared to GPT-4. Really? Yeah, though neither distribution was actually close to the human one.
6:38But it does kind of push back on that simple idea that just scaling up models automatically makes their reasoning more human. Fascinating. So size isn't everything, at least not for this kind of strategic reasoning. Did the paper just stop at the basic game, though? Or did they probe further? I think you mentioned variation. They didn't stop there. They also ran two variations of the game called Costless and Cycle. Okay. These were designed to test how participants react when you tweak the game's incentive structure. The idea was to see if LLMs show the same kinds of biases or shifts in reasoning that humans do when the rules change slightly.
7:14And did they? Well, humans responded to these variations in predictable ways, consistent with behavioral theory. The LLMs, however, often didn't react the same way at all. Some of the really advanced models, like GPT-4 and Klon 3 Opus, showed practically no change in their response patterns across these different versions of the game. No change at all, even when the incentives shifted. Pretty much none. Which is definitely not how humans behave in this game. Humans adapt their strategy. These models didn't. Okay, so, summary so far. They fail the basic test compared to humans, and they fail the test of adapting to rule variations like humans do.
7:51Given that pretty clear failure, did the researchers then try to, like, force the LLMs to be more human-like using common tricks? Yeah, that was the logical next step, right? If they don't do it naturally, can we make them do it? Right. So they explored several of these advanced prompting techniques and even some training methods that people often use to try and boost LLM performance or align their outputs more closely with what you want. Telling them, let's think step by step or giving them a specific persona. Exactly. Those kinds of things. They tried various zero shot prompts that included chain of thought, like you said.
8:25Yeah. Also prompts trying to inject emotion or urgency. Like this is very important in my career. Huh. OK. And even these sort of mindfulness or optimization prompts like take a deep breath. Did any of that work? Did it nudge them towards the human 17? The general finding was no. These techniques largely did not succeed in making the LLMs behave more like humans in this game. Really? None of them. For the bigger, more capable models, these prompts often had minimal effect, just didn't change much. For some of the smaller models, they might change the behavior, but not necessarily towards human likeness.
9:00What do you mean? Well, for example, one smaller model, Llamada 27B, when given those optimization prompts, sometimes shifted its answers way down to like 11 or 12. Oh, so like super deep reasoning levels. Yeah, levels of reasoning that are actually very rare, almost never seen, in actual human players. So it changed, but it didn't become human. Interesting. So the standard prompt tricks didn't crack it here. What about something else we hear about a lot? Prompt brittleness. How sensitive were the models to just the way the instructions were phrased? Ah, now this is where the findings get particularly, I'd say, concerning.
9:36Especially if you're hoping to use these models as reliable surrogates. Uh-oh. The paper found extreme sensitivity to the prompt. even really small changes in how the game instructions were presented could dramatically alter the LLM's behavior. Even if the rules themselves were identical. Exactly. Same game, same rules, just presented slightly differently. Can you give an example? Was it just like minor wording tweaks? It was sometimes more significant than that. For instance, just using different languages. Different languages. Running the game in English versus Chinese versus Spanish versus German, it resulted in different response distributions from the same models.
10:13Whoa. GPT 3.5, for example, gave different patterns of answers in Chinese compared to English. GPT 4 also showed shallower reasoning in Chinese than it did in English. That's wild. The language itself changed the apparent reasoning depth. It seems so. Another big factor was assigning the LLM a role. Like telling it, you are a rational player versus you are a human player. Precisely. Asking it to respond as a rational player versus a human player caused substantial shifts in its behavior. Sometimes it made the answers less diverse or it pushed the model towards those extreme reasoning levels, like level zero or the very deep levels, that you just don't see much in actual human groups.
10:55That level of volatility is pretty striking. So how you ask the question, even the language or the implied role matters hugely, and in ways you can't easily predict. Exactly. Did they also test giving the models examples? You know, a few shot prompts where you show them how others played. It did. And the paper found, perhaps unsurprisingly given the brittleness, that the LLMs were highly sensitive to the examples provided. Did they learn the human strategy from the examples? It seems not. Instead of learning the underlying reasoning process from the examples, they tended to just, well, copy the specific numbers they were shown.
11:28If the example showed people choosing low numbers, like 11, 12, 13, the LLMs would then cluster their own responses around those low numbers. And if the examples were high numbers? Same thing. If the examples were 18, 19, 20, they'd cluster their answers up there. They were basically mimicking the specific examples they saw, not capturing the overall human distribution or the thinking behind why humans often land around 17. OK, so zero shot fails. Few shot leads to mimicry. What about giving them more context, like feeding them the original research paper about the game, using RAI, retrieval, augmented generation?
12:05Good question. They tested that too. Giving the LLMs access to the original paper describing the 1120 game did seem to help a little bit. A little bit. Yeah, it led to a slight improvement in their alignment with the human behavior pattern for this specific game. However, even with that information readily available, the resulting distribution of the LLM responses was still statistically significantly different from the human distribution. Still different. Still significantly different, yeah. Yeah. P-value below.00001. So our egg helped slightly, but it definitely didn't close that gap. Didn't make them truly human-like.
12:38Okay. One last big technique. What about fine-tuning? Actually training a model specifically on a data set of human choices and their reasoning for this game. Right. The most intensive approach. They tried that as well. Did that finally work? So training an LLM on a specific data set of human choices and reasoning did manage to enable one model, it was a fine-tuned version of GPT-4-0, to produce a response distribution that couldn't be statistically distinguished from that specific human training data it saw. Ah, success. Well, maybe. The paper adds a huge caveat here. Okay. They strongly caution that this might just be the model memorizing that particular data set.
13:16It doesn't necessarily mean it has gained a generalized ability to apply human-like reasoning to new, unseen situations. Right. It learned that pattern, but maybe not the underlying principle. Exactly. And they point to other evidence, too. They tested a model called Centaur, which was explicitly built and trained on tons of human data specifically to emulate humans. Okay. Even Centaur failed to replicate the human patterns in this specific 1120 game. Wow. Even a purpose-built emulator failed on this simple task. Mm-hmm. And, you know, as you kind of hinted at earlier, if you need to fine-tune a model with vast amounts of very specific human data just to make it simulate humans in one task...
13:58Doesn't that kind of defeat the whole purpose? Right. It undermines the main motivation for using LLMs as surrogates in the first place, which is often precisely when you don't have access to that kind of rich human data. It's a real catch-22, isn't it? You need the human data to make the surrogate work, but you want the surrogate because you lack the human data. Exactly. It's a fundamental tension there. So let's pull this together. They fail the basic game. They're incredibly brittle to prompts and examples. Even fine-tuning seems more like mimicry. What does the paper conclude about why these attempts to make LLMs human-like, at least in this context, kept falling short?
14:31The paper puts forward this idea of a demand effect. Demand effect. Yeah. The argument is that the LLM, being this incredibly sophisticated pattern matching machine, gets really good at detecting subtle clues in the prompt or the examples or the fine-tuning data or even just the inherent biases from its general training like RLHF. And it figures out what kind of response seems to be expected or desired, and then it generates that output. So it's giving what it thinks you want to hear rather than reasoning independently. Precisely. It's producing the pattern it thinks is correct for the situation you've set up, rather than actually performing genuine human-style strategic reasoning.
15:12It looks like the right behavior, but the underlying process driving it is totally different. That takes us right to the paper's deeper analysis then. Why do LLMs struggle with this simple game, especially when we hear about them passing complex exams or benchmarks? What are the really fundamental differences between LLM intelligence and ours? Yeah, that's a key insight from the paper, isn't it? How can something seem so brilliant in one context and then fail at something relatively simple in another? Right. Well, they dug into this. One thing they investigated was whether the LLMs had simply memorized the instructions or common outcomes for this 11-20 game during their initial training.
15:48And had they? Interestingly, no. They tested the model's ability to recall the instructions accurately. For the classic beauty contest game one, where models do often perform quite like humans, they found the LLMs were highly accurate, like 75 % to 100 % recall. Okay, so they had seen that one a lot. Seems like it. But for this 11-20 game, their recall accuracy for the instructions was basically zero, near zero percent. Whoa. So the implication is their good performance on those classic games might just be them spitting back stuff they saw over and over in training. That's a very strong implication, yes.
16:24This finding really suggests that the apparent human-like performance on certain familiar, well-documented games might just be the result of regurgitating patterns learned from the training data. Stochastic parroting, as they call it. Exactly. Stochastic parroting. It connects to those older philosophical ideas, too, like Searle's Chinese Room argument. It suggests the behavior we see might be incredibly sophisticated pattern recognition and reproduction, rather than genuine reasoning ability that can generalize robustly to new, less familiar situations like this 1120 game. So it's not really thinking like us.
17:01It's just an unbelievably good mimic when it knows the script. It seems to be the conclusion for these kinds of tasks, yes. And the paper goes on to outline several fundamental reasons why LLMs are just inherently different from human intelligence. These differences really contribute to their limitations as surrogates. What are some of those core differences? Well, first, they have fundamentally different objectives, right? Humans have goals related to survival, social connection, reproduction, all rooted in embodied real-world experience and biological imperatives. Right. We have needs, desires.
17:32LLMs. Their goals are mathematical. They're typically based on statistical objectives, like predicting the next word in a sequence based on patterns they've seen in vast amounts of digitized text. Very different driving forces. We live in a physical world. They live in a world of text data. Exactly. Which leads to the second point. They lack embodiment. Our human cognition, our thinking, is profoundly shaped by our senses, our physical interactions with the environment, our internal bodily states. LLNs have none of that lived experience. They don't stub their toe. They don't feel hungry. Right.
18:07And there's also the limitation of the data itself. Even with the entire internet, there simply isn't enough high-quality digitized data out there to capture the sheer breadth and depth of human tacit knowledge. Tacit knowledge. Like intuition. Stuff we know but can't easily explain. Yeah, exactly. The know-how, the common sense, the cultural understanding that's hard or impossible to write down explicitly. Think of Polanyi's paradox. We know more than we can tell. If it's not written down, it's likely not well represented in the LLM's training data, limiting what it can truly learn about the nuances of how humans actually think and decide.
18:44So a huge chunk of what makes our decision-making human isn't even in the data they learn from. That's a major part of it. And the paper also points out that even with the data they do have, LLMs show significant weaknesses in various crucial forms of reasoning. Like what? Logical reasoning, causal reasoning, understanding cause and effect temporal reasoning, factual accuracy, deductive reasoning, abstract thinking. These are all core to human decision-making, planning, problem-solving. And areas where LLMs are known to stumble, even on simple versions sometimes? Precisely. Their performance can also be quite erratic, inconsistent.
19:18They might succeed on a complex task, but then fail on a simpler variation. And we already talked about how brittle they are to minor changes in prompts. Plus the hallucination problem. Plus the hallucination problem, right. Confidently making stuff up. And what about the training process itself? You mentioned RLHF reinforcement, learning from human feedback. can that also create differences? It absolutely can. RLHF is generally designed to make models safer, more helpful, less likely to say harmful things, which is good. But that process can inadvertently shape their behavior in ways that deviate from actual human norms.
19:53For instance, it might make them overly pro-social or perhaps too risk averse compared to the real messy distribution of human behavior. So they become too nice or too cautious compared to real people. Potentially, yeah. It can also sometimes reduce the diversity of their outputs, making them less likely to generate outlier opinions or unconventional ideas. And it can amplify biases that were present in the human feedback data itself, sometimes called the pigeonhole effect. All of this limits their ability to truly represent the full spectrum of human responses. Okay, wow. That's a pretty comprehensive list of challenges and fundamental differences.
20:28They fail a simple, novel reasoning test. They're super brittle to how you ask the question. They seem to be mimicking or pattern matching rather than reasoning from first principles. And they have these deep, inherent differences in goals, embodiment, data, and even reasoning capabilities. So bringing it all together, what are the major implications? What's the big caution for researchers or really anyone thinking of using LLMs in this way? The paper's recommendation is very direct, very clear. It basically says, exercise significant caution. Be very careful when you're considering using LLMs as stand-ins for humans or as your primary tool for trying to understand human behavior.
21:09What specific things should people be asking? Like if you're a researcher doing this or maybe someone reviewing a study that used LLMs this way. The paper proposes a kind of checklist, a set of key evaluative questions. Okay, like what? First, demand full transparency, full documentation of the experimental setup. not just the model name like GPT-4, but the specific version used, who provided it, the exact date the experiment was run because models change. Right, they get updated all the time. Exactly, and the precise prompts used, word for word. Second, they say don't just look at averages. Compare the full response distributions.
21:44How spread out are the LLM answers compared to the human answers? The shape of the behavior matters just as much as the average outcome. Don't just check if the mean is similar. Look at the whole pattern. Precisely. Third, researchers need to actively look for and report on known LLM failure modes. Did the model hallucinate during the task? Was its behavior consistent across multiple runs? Is there evidence it might have just memorized the task from its training? Be upfront about these potential issues. Don't just assume it worked perfectly. Actively test for the known ways it can fail. Right.
22:16And check for stability. How does the behavior change if you use a different model? Or a different model provider? Or a different size of model? And crucially, how stable is it over time? If you re-ran the experiment a month later, would you get the same result? Good point. And test for task sensitivity you mentioned. Yes. Test if the LLM responds to changes in the task or game rules in ways that are consistent with how humans respond. Do they react when humans would react? Do they stay stable when humans would stay stable? It's about matching the pattern of sensitivity, not just the baseline behavior.
22:50And obviously, given the earlier findings, test for prompt brittleness. Absolutely critical. Actively test for it. Try slightly different wording. Change the order of instructions. Maybe even try it in a couple of different languages as they did in the paper. See if the behavior shifts unexpectedly. And the memorization check? Crucial. Can the LLM accurately recall specific instructions or details about the game after it's participated? If it can, that's a red flag that its performance might be contaminated by prior exposure in its training data. not genuine reasoning in the moment. Compare behavior on tasks it likely hasn't seen out of distribution versus ones it might have in distribution.
23:30Okay. And finally, don't just take the LLM's word for it when it explains itself. Right. Scrutinize those self-explanations. As the paper found, what the LLM says about its reasoning doesn't always match its actual behavior, and sometimes the explanations themselves reveal fundamental misunderstandings of even simple rules. That's a solid checklist. The risks, if people aren't this careful, seem pretty significant then. They really are. The paper highlights a few key risks. One is the ease with which researchers might consciously or unconsciously tweak prompts or select specific models or specific runs until they get a result that looks human-like or confirms their hypothesis.
24:07It's like a new form of packing. Finding the right prompt to get the right answer. Exactly. Which might just be an artifact of the prompting strategy, not a reflection of simulated human thought. Another big risk is that it could distort research incentives. It might seem easier or cheaper to run studies on LLMs than to do the hard, expensive, but vital work of collecting real data from diverse groups of actual humans. Especially underrepresented groups. Precisely. So we risk relying more on synthetic data that might not be accurate, potentially leading to skewed or just plain wrong conclusions about how people actually think and behave.
24:42So the bottom line from this Scylla X machina paper is pretty stark then. LLMs, at least in their current form, are fundamentally different from us. They're unreliable as direct surrogates for human behavior and research. Yeah, that's the core message. Their behavior is highly unpredictable. It's shaped by their architecture and their training data in ways we often don't fully grasp. And it can be easily manipulated by the way we interact with them. The paper used that analogy, right? About the parrot. Yeah, it's a powerful one. expecting to get deep insights into human psychology by running experiments on LLMs is like, well, like a psychologist interviewing a parrot to understand its owner's mental state.
25:21You might get something back that sounds vaguely familiar, kind of low resolution reflection, maybe, but it's incredibly easy to be misled to mistake the echo for the real voice. And that title, Scylla Ex Machina, it's referencing the mythological sea monster Scylla, right? But Ex Machina from the machine. Exactly. It's a metaphor. LLMs have this alluring human-like facade. They generate text that sounds like us. But beneath that surface, like Scylla, they are something fundamentally different, potentially dangerous if you mistake them for what they resemble. They're unpredictable, driven by the mechanics of their underlying architecture and the vast, sometimes strange ocean of data they were trained on.
26:00A machine-made monster that looks deceptively familiar. Okay, so if they're not reliable human simulators, at least not now, how should we be thinking about using these powerful AI tools in relation to human understanding? Well, the paper does offer a suggestion there. It proposes viewing LLMs perhaps not as simulations of human behavior, but maybe more as tools for synthetic imagination or ideation. Synthetic imagination. Yeah. They could be incredibly good at generating novel ideas, brainstorming hypotheses, creating diverse scenarios, or even generating synthetic data patterns that might be interesting starting points.
26:36So useful for exploration, but not validation. That's the key. Any outputs, any ideas, any patterns derived from LLMs absolutely require rigorous testing and validation against real world human data before we can trust them. They could be like a really interesting, sometimes weird brainstorming partner, but they are definitely not a substitute for actually going out and observing, talking to and understanding actual human beings. Couldn't have said it better myself. And, you know, given that risk, we talked about the LLM echo chamber or that demand effect. Where they might just tell you what they think you want to hear.
27:10Exactly. Especially when you're prompting them to act a certain way or simulate a certain group. It raises a really important final thought, I think, for you, the listener. Okay. How does seeing these limitations, this brittleness, this fundamental difference in a simple reasoning task, how does that change how you think about using AI in any context where genuine human understanding or nuanced judgment or maybe a truly representative range of human perspectives is actually important? It suggests we need to be pretty careful, doesn't it? Caution seems absolutely essential.
This academic paper investigates the suitability of large language models (LLMs) as substitutes for human participants in social science research. The authors examine LLMs' reasoning abilities using the "11-20 money request game," a test designed to evaluate strategic thinking. Their findings consistently show that LLMs generally fail to replicate human behavioral patterns, exhibiting less reasoning depth and inconsistent responses compared to human subjects. The study highlights several limitations of LLMs, including their reliance on probabilistic patterns rather than genuine understanding, sensitivity to subtle changes in prompts or language, and the potential for memorization of training data to be mistaken for true reasoning. Ultimately, the paper concludes that caution is essential when considering LLMs as human surrogates, suggesting they are currently better suited for generating novel ideas rather than simulating human behavior.
keepSave to notecopy_alldocsAdd noteaudio_magic_eraserAudio OverviewflowchartMind Maparrow_downwardJump to bottom