Can AI Really Think Like Us? A Deep Dive into LLM Simulations

Large Language Models (LLMs) like ChatGPT and DeepSeek are revolutionizing AI, but can they truly simulate human thought, emotions, and social behaviors? Researchers Qian Wang, Zhenheng Tang, and Bingsheng He tackle this question in their latest study, exploring the strengths and pitfalls of using LLMs to model human societies.

Let’s break it down and see if AI can really think like us—or if it's just good at faking it.

The Promise of LLM-Powered Simulations

LLMs are trained on vast amounts of text, allowing them to generate content that mimics human conversation. This makes them attractive for simulating complex human behaviors in economics, psychology, and even social sciences.

Imagine being able to test political campaigns, economic policies, or social dynamics without needing real-world experiments. LLMs offer a scalable, cost-effective way to model such scenarios. Unlike traditional simulations, which require predefined rules and expensive human participants, LLM-based simulations can generate dynamic interactions in real time.

But there’s a catch—several, actually.

Why LLMs Struggle to Truly Mimic Us

While LLMs can mimic human-like responses, they lack the depth of real human cognition. Here's why:

1. No Inner Psychology or Personal Experience

Humans make decisions based on emotions, personal experiences, and psychological states—all things LLMs lack. AI models “think” using patterns derived from training data, but they don’t experience life, form memories, or have personal biases shaped by past events.

For instance:
- A person who has experienced betrayal may behave cautiously in new relationships.
- An AI model, on the other hand, can only infer how a cautious person might behave based on the data it was trained on—it doesn't feel that caution.

This missing personal history makes AI-generated personalities feel generic, even when researchers prompt them to behave differently.

2. No Real Motivations

Why do humans make choices? We have motivations: survival, happiness, social acceptance, personal growth—you name it. LLMs, however, operate purely on statistical probabilities, predicting the "next likely word" rather than acting with intent.

This lack of intrinsic motivation means that even the most advanced AI isn’t making choices in the way humans do. It’s simply calculating the most expected response.

Think about how you negotiate a salary—your decision is influenced by past experiences, your financial needs, and your future goals. An LLM, by contrast, only considers patterns from negotiation conversations found in its training data.

3. Biases in Training Data Skew AI Behavior

LLMs inherit biases from the datasets they’re trained on. A model trained predominantly on Western, English-language data may struggle to accurately represent non-Western cultural perspectives.

Some major biases include:
- Cultural biases: Narrow viewpoints due to unbalanced training data.
- Socioeconomic biases: Missing perspectives of lesser-digitized industries or communities.
- Gender biases: Reinforcing stereotypes found in online data.

These biases lead to distorted simulations that don’t fully reflect real-world human diversity.

4. One AI, Many "People" – Is That Possible?

If multiple characters in a simulation are generated from the same LLM, can they truly behave like distinct individuals? Chances are, not entirely.

LLMs don’t have separate "minds" or "personalities" for different users unless given extensive prompting. This can create an illusion of diversity in a simulation but lacks genuine independent thinking among the simulated characters.

A great example is social simulations where AI agents participate in elections or make business decisions. While they can simulate such processes, they might not reflect how real humans would behave because their "thought processes" are essentially the same.

So, Why Use LLM Simulations if They’re Not Perfect?

Despite their flaws, LLMs still offer huge advantages. In many cases, they provide insights that traditional human-based simulations can’t:

• Scalability: AI models can simulate thousands of interactions quickly, covering large-scale social dynamics that would be impossible to test with human participants.
• Cost Efficiency: Running AI simulations is much cheaper than recruiting and compensating human subjects.
• Emergent Behaviors: Unlike rule-based simulations, LLMs sometimes exhibit unexpected behaviors that weren’t explicitly programmed, providing fresh insights into human-like interactions.
• Ethical Research: Simulating real-world issues like social bias or psychological stress without subjecting real people to uncomfortable experiments.

In fields like psychology, market research, and policy planning, LLMs are already being used to test theories, predict outcomes, and explore alternative futures.

Aligning LLM Simulations More Closely with Humans

If LLM-based simulations are to truly reflect human behavior, researchers suggest some improvements:

1. Better Training Data

To enhance realism, AI needs more diverse and personalized datasets, including first-person narratives, psychological insights, and cultural experiences.

2. More Human-Like Incentives

Future AI models should integrate external motivation structures, allowing them to simulate decision-making that’s driven by goals, fears, or aspirations.

3. More Realistic Simulation Environments

Building richer virtual worlds where AI agents must compete, collaborate, or survive in dynamic, resource-driven environments could make decisions more lifelike.

4. Improved Metrics for Evaluating AI Behavior

New ways to assess AI simulations—such as measuring psychological alignment with real human behaviors—could help refine these models further.

Case Study: LLMs in Cryptocurrency Trading

LLMs are already being used in real-world simulations beyond just academic research.

Take CryptoTrade, an LLM-powered trading agent that simulates human traders making buy and sell decisions in cryptocurrency markets.

Some key takeaways from this system:
- It struggles to outperform simple trading strategies. In a bear market, CryptoTrade fell behind the "buy and hold" strategy.
- It’s biased toward factual over sentiment-based analysis. This might work well in some cases but fails in markets driven by speculation and emotion.
- It creates herd behavior. If all AI traders are using similar models, they tend to make the same decisions, amplifying market trends instead of diversifying trades in human-like ways.

This experiment highlights that AI simulations still lack essential human-like unpredictability, intuition, and emotional decision-making—key factors in real financial markets.

Final Thoughts

LLMs provide an exciting tool for simulating human behavior, but they don’t actually think or feel like us—they only predict language patterns based on data. While this makes them useful for research, their limitations mean we can’t yet rely on them to fully model human psychology or society.

As AI improves, researchers will need to bridge the gap by incorporating richer training data, better incentive structures, and more diverse decision-making approaches. For now, though, AI remains a powerful reflection of humanity—but not quite a perfect replica.

Key Takeaways

• LLMs offer a low-cost, scalable way to simulate human interactions, useful in psychology, finance, and social science studies.
• However, LLMs lack real emotions, intrinsic motivations, and personal experiences, making their simulations only approximations of human behavior.
• Biased training data leads to distorted simulations, failing to represent the full range of human psychology and cultures.
• LLM-based trading bots struggle to outperform basic strategies, indicating that AI doesn’t yet think like real human traders.
• For better AI simulations, future research should incorporate more personal, psychological, and diverse datasets to align AI behavior more closely with real human societies.

What This Means for You

If you're using AI for tasks requiring nuanced human behavior—whether that’s chatbots, market research, or creative writing—keep its limitations in mind. While AI is an amazing tool, it isn’t a perfect substitute for human insight.

Want to improve your AI-related projects? Consider crafting better prompts or fine-tuning AI with rich, diverse data to push the boundaries of what’s possible!