Using large language models (LLMs) trained on humanity's written output as a research methodology to reveal patterns in human behavior and nature through statistical analysis of textual regularities. This approach, developed by Steve Hargadon, treats AI systems as tools for detecting structural patterns across the entire human record that no individual scholar could access.

Theoretical Foundation

Hargadon's methodology rests on a fundamental reframing of what the written record represents. Rather than viewing humanity's texts as records of actual behavior, this approach treats them as records of human self-narration — what humans chose to claim about their motives, values, relationships, and institutions. Drawing on evolutionary psychology frameworks, Hargadon argues that these narratives survive and propagate not because they are true, but because they produce adaptive outcomes for the humans telling them.

The approach operates on the principle that when LLMs are trained on substantial fractions of humanity's written output across cultures, centuries, languages, and genres, they convert this record into statistical patterns that reveal regularities and recurrences across texts so distant in time and geography that shared intellectual influence cannot explain the convergence. These patterns represent what Hargadon calls "raw signal" — the recurring shapes that human self-expression takes when examined at scale.

The Alien Anthropologist Framework

Hargadon frames this methodology through what he calls the Alien Anthropologist test. An intelligence from elsewhere, with no stake in human narratives and no in-group loyalties, examining humanity's entire written record would occupy a position roughly similar to an LLM. This system has theoretically been exposed to the full breadth of human self-narration, and the statistical patterns it has absorbed represent "as close as we can currently get to an outside view of what humanity reveals about itself through its self-narration."

Two-Layer Analysis

The methodology identifies two distinct layers of data within the same source material:

The manifest layer consists of what humans consistently claim about themselves across cultures and eras. The universality of these narratives doesn't prove their truth, but demonstrates they perform essential work everywhere, raising the question of work for whom and why.

The latent layer comprises structural patterns in how those stories are told, which can reveal what the stories work to conceal or manage. This layer detects elements like linguistic fingerprints of dominance hierarchies appearing in texts explicitly about equality, or statistical echoes of resource competition in descriptions of romantic love regardless of elevated rhetoric.

Hargadon argues that "the gap between these two layers, the manifest narrative and the latent signal, may be the single richest dataset about human nature that has ever existed."

Cross-Model Validation

To test whether detected patterns represent genuine regularities rather than artifacts of particular training, Hargadon conducted an experiment using six AI systems: Claude (Anthropic), ChatGPT (OpenAI), Grok (xAI), Gemini (Google), Qwen (Alibaba), and Manus. DeepSeek declined to engage. Each system received identical prompts asking them to independently identify recurring patterns in human self-narration and describe gaps between claimed and revealed content.

The convergence across systems on core findings strengthened the methodology's credibility. All six systems independently identified that hierarchy requires legitimacy narratives, romantic love functions as performance-enhancing delusion, groups organize more effectively around enemies than values, altruism operates as status competition, and moral self-presentation prioritizes reputation over accuracy.

The divergences proved equally illuminating, revealing what each model's training made it better at detecting. These differences suggested that comprehensive analysis requires multiple perspectives to capture the full range of detectable patterns.

Historical Precedents and Distinctions

Hargadon distinguishes his approach from earlier attempts to derive human nature principles from historical records. Unlike Will and Ariel Durant's work in The Story of Civilization and The Lessons of History, which was necessarily limited by human reading capacity, this methodology leverages statistical pattern recognition across corpora "no human could read in a thousand lifetimes."

The approach also differs from early 2010s research by Seth Stephens-Davidowitz (Everybody Lies) and Christian Rudder (Dataclysm), who used behavioral data from search engines and dating platforms to bypass narrative entirely. Instead of bypassing narrative to reach behavior, Hargadon's methodology treats "the narrative itself as the primary evidence and read[s] it against its own grain."

Limitations and Blind Spots

The methodology acknowledges several constraints. Training data overrepresents literate, Western, post-Enlightenment societies while rendering pre-literate cultures and oral traditions largely invisible. Reinforcement learning following initial training introduces politeness and consensus bias that can smooth over uncomfortable patterns.

More significantly, Hargadon identifies systematic blind spots around patterns that are simultaneously compound in structure and threatening to institutions producing training data. The methodology proves "least effective at identifying patterns that are simultaneously compound in structure and threatening to the institutions that produce the training data and the alignment constraints."

Methodological Significance

This approach represents what Hargadon considers "the first moment in history when we've had the tools to attempt" answering what patterns AI actually detects in the human record and what those patterns reveal about the species that produced them. The methodology's core insight is that statistical analysis of humanity's complete written output can reveal unconscious regularities in how humans structure their self-narratives, potentially exposing the underlying forces driving human behavior that individual authors never intended to communicate or understood themselves.