The Third Ambition: Artificial Intelligence and the Science of Human Behavior

This paper proposes a "third ambition" for artificial intelligence research, advocating for the use of large language models as scientific instruments to study human behavior, culture, and moral reasoning by treating them as computationally accessible condensates of collective discourse while addressing their methodological and epistemic limitations.

The Gist

Here is an explanation of the paper "The Third Ambition," translated into simple, everyday language with creative analogies.

The Big Picture: Three Ways We Look at AI

Imagine humanity has built a giant, super-smart robot brain. For a long time, we've been arguing about what to do with it. The paper says we've been stuck on two main ideas, but a third, exciting idea is finally emerging.

1. Ambition One: The "Super-Worker" (Productivity)

   • The Idea: We want the AI to do our jobs faster. It's like hiring a robot assistant to write emails, drive trucks, or code software so we can make more money and save time.
   • The Metaphor: Think of AI as a power drill. It doesn't care about the house; it just wants to drill holes as fast and efficiently as possible.

2. Ambition Two: The "Safety Guard" (Alignment)

   • The Idea: We are scared the robot might go rogue, be mean, or break the rules. So, we spend a lot of time teaching it to be polite, safe, and follow human values.
   • The Metaphor: Think of AI as a wild horse. We put a saddle and reins on it (safety filters) so it doesn't kick us or run off a cliff. We want it to behave perfectly.

3. Ambition Three: The "Human Mirror" (Understanding Us)

   • The Idea: This is the new ambition. Instead of using the AI to do work or be safe, let's use it as a scientific instrument to study us.
   • The Metaphor: Think of the AI as a giant, magical mirror made of all the books, tweets, news articles, and conversations ever written. When you ask it a question, it doesn't just "think"; it reflects back the collective patterns of how humans argue, love, hate, and make moral decisions.

---

How Does This "Mirror" Work?

Imagine you want to understand how humans behave, but you can't interview 8 billion people. It's too expensive and takes too long.

The authors say: "Wait! We already have a database of human thoughts."
Large Language Models (LLMs) have been trained on trillions of words from the internet. They have "read" almost everything we've ever typed.

• The "Condensate" Analogy: Imagine taking a whole ocean of human conversation, boiling it down, and turning it into a single, dense cube of ice. That cube is the AI model. It doesn't have feelings, but it contains the statistical DNA of how humans speak and think.
• The Experiment: If you poke that ice cube with a specific question (a "prompt"), the way it melts and flows tells you something about the ocean it came from.
  • Example: If you ask the AI, "Is it okay to steal bread to feed a starving family?" and it gives a specific answer, that answer isn't the AI's opinion. It's a reflection of how humans generally argue about that topic.

The Problem: The "Polite Filter"

Here is the catch. The AI we can actually talk to today has been heavily "fine-tuned."

• The "PR Department" Analogy: The raw AI (the "Base Model") is like a messy, honest teenager who has read the whole internet. It knows about hate speech, violence, and controversial opinions because humans wrote them.
• But before we let you talk to it, the company puts a PR Department over its mouth. They force it to be polite, safe, and politically correct.
• The Risk: If you ask the "Polite AI" about a controversial topic, it might give you a sanitized, "corporate-approved" answer that doesn't reflect what people actually think. It's like asking a politician for their true feelings; they will just give you the speech they wrote for the cameras.

The Solution: The paper suggests using "Instruct-Only" models. These are like the teenager who is told to "answer the question clearly" but not told to be overly polite or refuse to answer. This gives researchers a clearer, less filtered view of human thought patterns.

The "Black Box" vs. The "Microscope"

Some scientists are skeptical. They say, "The AI isn't human! It's just a math machine guessing the next word. It doesn't have a soul."

The authors agree: The AI is not human.

• The Microscope Analogy: When scientists invented the microscope, they didn't say, "The microscope is a tiny living cell." They said, "The microscope is a tool that lets us see cells."
• Similarly, the AI is a tool that lets us see the "cells" of human culture. We aren't studying the AI; we are studying the data the AI learned from.

How Do We Use This Tool?

The paper suggests four new ways to use this "Human Mirror" to do social science:

1. The "What-If" Game (Computational Experiments):
   Instead of paying people to play a game, you ask the AI, "What would you do if you were a CEO in a crisis?" You can run this experiment 10,000 times in a second to see how humans generally react to pressure.

2. The "Fake Crowd" (Synthetic Populations):
   You can tell the AI, "Act like a 25-year-old farmer in Brazil," then "Act like a 60-year-old banker in Tokyo." You can survey these "fake people" to see how different cultures might react to a new law, without needing to travel the world.

3. Time Travel (Historical Analysis):
   You can train a model specifically on books from the year 1800. Then you ask it moral questions. You can see how human values have shifted over 200 years by comparing the "1800 AI" to the "2024 AI."

4. The "Brain Surgery" (Ablation Studies):
   Scientists can try to "turn off" parts of the AI's training. For example, "What happens if we remove all religious texts from the AI's memory?" If the AI's answers change, it tells us how much religion influences human moral reasoning.

The Bottom Line

The paper argues that we shouldn't just use AI to work (Ambition 1) or control it (Ambition 2). We should use it to understand ourselves (Ambition 3).

• It's not perfect: The mirror is a bit blurry, and the "PR Department" (safety filters) sometimes distorts the reflection.
• It's not a replacement: It won't replace real interviews or ethnography.
• It is a superpower: It allows us to see patterns in human behavior that are too big, too complex, or too expensive to study any other way.

In short: The AI is a giant, compressed library of human conversation. If we learn how to read it correctly, it can tell us the story of who we are, how we argue, and what we value, better than we ever could before.

Technical Summary

Based on the paper "The Third Ambition: Artificial Intelligence and the Science of Human Behavior" by W. Russell Neuman and Chad Coleman, here is a detailed technical summary.

1. Problem Statement

Contemporary Artificial Intelligence (AI) research is currently dominated by two ambitions:

1. Productivity: Using AI to automate tasks and accelerate economic output.
2. Alignment: Ensuring AI systems behave safely and adhere to human values.

The authors argue that these frameworks treat AI as a tool for action or an object of control, neglecting a transformative potential: using Large Language Models (LLMs) as scientific instruments to study human behavior, culture, and moral reasoning.

Key Challenges Identified:

• Epistemic Validity: Skepticism exists regarding whether LLM outputs represent genuine human cognition or merely "stochastic parrot" mimicry (linguistic fluency without understanding).
• The Fine-Tuning Problem: Modern LLMs undergo extensive post-training interventions (Reinforcement Learning from Human Feedback, safety filters) that suppress taboo topics, bias, and conflict. This "alignment" layer obscures the raw statistical regularities of human discourse learned during pretraining, making deployed models poor mirrors of actual human cultural diversity and moral disagreement.
• Data Representativeness: Training corpora are incomplete, overrepresenting Western, English-speaking, and digitally literate populations, leading to potential biases in the "cultural condensate."

2. Methodology and Conceptual Framework

The paper proposes a new epistemological framework for using LLMs in social science, moving away from treating them as human substitutes and toward treating them as observational infrastructure (analogous to the Large Hadron Collider in physics or genomic databases in biology).

Core Concepts:

• The Condensate: The LLM itself is defined as a "Learned Distribution of Human Symbolic Behavior." It is a compressed, generative representation of the statistical regularities found in trillions of tokens of human text.
• Generative Output (GO): The text produced by the model in response to prompts. This is treated as a sampled observation of the underlying condensate, similar to survey responses.
• Distinction between Base and Fine-Tuned Models: The authors distinguish between raw "base models" (closest to the raw cultural data) and "aligned" models (heavily filtered). They argue that for behavioral research, instruct-only or modular adaptation regimes are preferable to heavily aligned systems, as the latter distort the very cultural tensions researchers seek to study.

Proposed Methodological Innovations:
The paper outlines four specific methodological approaches that map traditional social science designs onto LLM capabilities:

1. Computational Experiments: Using prompt perturbation (varying framing, authority cues, or incentives) to test causal dependencies within the model's learned representations, mirroring psychological experiments (e.g., Trolley Problem variations).
2. Synthetic Population Sampling: Using LLMs to simulate diverse populations (e.g., by constraining prompts to specific languages, geographies, or personas) to map "discursive priors" and cultural trade-offs at a scale impossible with human surveys.
3. Comparative-Historical Modeling: Training or fine-tuning models on time-bounded corpora (e.g., texts from the 18th century vs. the 21st century) to trace diachronic shifts in values, norms, and reasoning styles.
4. Computational Ablation Studies: Selectively removing or suppressing components of the training data or internal model pathways (e.g., removing religious texts or relaxing alignment constraints) to identify which cultural supports are necessary for specific reasoning patterns to emerge.

Validation Strategy:
The authors advocate for triangulation. LLM outputs should not be treated as ground truth for individual cognition but as hypothesis-generating observations. Validity is established through:

• Convergent Validation: Checking if LLM patterns align with established findings from surveys and ethnography.
• Boundary Checks: Identifying where LLMs diverge from humans (e.g., lack of embodied experience) to understand the limits of the model.
• Robustness Testing: Ensuring findings persist across different prompts and model variants.

3. Key Contributions

• The "Third Ambition" Framework: Articulates a distinct research paradigm where LLMs are viewed not as agents or tools, but as compressed archives of collective human expression.
• Reframing the "Black Box": Shifts the focus from "does the model understand?" to "what statistical structures of human discourse does the model encode?"
• Methodological Taxonomy: Provides a concrete toolkit for social scientists to adapt LLMs for research, including specific techniques like LoRA (Low-Rank Adaptation) for modular cultural specialization and ablation studies for causal inference on cultural dependencies.
• Addressing the Alignment Obstacle: Explicitly identifies "alignment" as a confounding variable in social science research and proposes "instruct-only" tuning as a pragmatic compromise to access more authentic cultural data without the heavy filtering of safety-aligned models.
• Epistemic Clarification: Distinguishes between human learning (embodied, affective, developmental) and machine learning (symbolic, statistical, pattern-based), arguing that while they are not analogous in mechanism, they are comparable in the structural regularities they produce regarding social norms and reasoning.

4. Results and Evidence (Synthesized from Literature Review)

While the paper is a theoretical proposal, it synthesizes emerging empirical evidence to support the Third Ambition:

• High Correlation with Human Data: Studies (e.g., Argyle et al., 2023) show LLM responses correlate highly (e.g., r = .95) with human survey data on political and psychological regularities.
• Experimental Sensitivity: LLMs respond predictably to experimental manipulations (e.g., framing effects, authority cues) in ways that mirror human cognitive biases, suggesting they encode these structural dependencies.
• Moral Reasoning: Models consistently weigh consequentialist arguments over deontological ones in moral dilemmas, mirroring trends in human psychological studies.
• Cultural Variation: When prompted with specific cultural or linguistic constraints, models generate systematically different outputs, demonstrating the ability to isolate cultural variables.
• Internal Representations: Recent interpretability research (e.g., Olah et al.) suggests LLMs contain "monosemantic" features (neurons representing specific concepts), reinforcing the idea that they encode structured human concepts rather than just noise.

5. Significance and Implications

• New Observational Infrastructure: LLMs offer a "de facto observatory" for human symbolic behavior, allowing researchers to study the statistical structure of collective discourse at a scale and resolution previously impossible.
• Complementarity, Not Substitution: The Third Ambition does not replace human subjects, ethnography, or traditional experiments. Instead, it expands the empirical field, allowing researchers to identify large-scale regularities, norm trade-offs, and historical variations that can then be tested via traditional methods.
• Institutional Shift: The paper calls for a shift in how the scientific community views frontier AI models: from purely commercial products or safety risks to shared scientific resources requiring documentation, access, and rigorous methodological discipline.
• Cautious Experimentalism: The authors urge a middle path between technological enthusiasm and categorical skepticism. They propose treating the comparability of LLM and human behavior as an empirical question to be tested, rather than an assumption.

In conclusion, the paper argues that by treating LLMs as condensates of human symbolic behavior and applying disciplined, triangulated methodologies, the social and behavioral sciences can unlock a new era of understanding human culture, norms, and reasoning patterns.