«Humans make the same mistakes over and over. I saw it when I was a vengeance demon. Some guy dumps a girl, she calls me, I exact vengeance, blah blah blah, the next year, same girl, different guy. I mean, after you smite a few of ’em you start going, «My goodness, young lady … maybe you’re doing something wrong here too.»
Anya, an ex-vengeance demon for women scorned, in «Buffy – The Vampire Slayer»
I was curious whether LLMs — AIs trained with the knowledge of the world — could make some kind of explorable collective unconscious of mankind. After all, they have all that information, much more than any single human could know, and it can be queried (until the safety layer kicks in).
ChatGPT got first hung up a bit on the metaphor — as it is not a collective unconscious in the Carl Jung sense. But it does have a collective surface of human expression.
They absorb the textual outputs of civilization—books, forums, code, news, essays, chatter. That gives them reach across more domains and patterns than any individual human could hold. You get a wide, aggregated map of how humans talk, explain, argue, imagine, justify, fantasize.
ChatGPT 5.1
And in that data, «surprising regularities show up:
- stylistic trends across cultures
- structural motifs in storytelling
- implicit assumptions common in political language
- recurring metaphors around anxiety, progress, or identity»
However, with three layers of distortion/biases:
- Data distribution: What humanity writes most about, not what humanity feels most deeply.
- Curation and filtering: Safety filters, moderation, exclusions.
- Training objective: Predict the next token—not reveal hidden human longings or archetypes.
These biases don’t break the analogy; they define why the analogy is incomplete. You’re not getting the human core. You’re getting the human exhaust.
ChatGPT 5.1
So in short, it …
represent[s] collective externalized cognition:
the sum of symbolic artifacts humans produced, compressed into a form that can be recombined with high fluidity. The thing that becomes newly accessible isn’t unconscious content—it’s cross-domain connectivity:
history + psychology + pop culture + engineering + myth + rhetoric + science, all fused into a manipulable latent space.That’s not unconscious. It’s not conscious. It’s a hyper-indexed synthesis layer: a dynamic model of the ways humans have put thoughts into language.
ChatGPT 5.1
Asking ChatGPT what patterns about humans emerge in the data, esp. things mankind might not realize about itself, pointed to two levels:
1. Patterns that emerge only at scale
ChatGPT pointed to the following:
- The remarkable uniformity of narrative structure across cultures. Not just “hero’s journey” clichés—deeper things: typical pacing rhythms, the way conflict escalates, where emotional beats tend to cluster, the distribution of metaphors around agency, fate, and morality. These convergences become obvious when you compress massive story corpora into a single representational space.
- The narrowness of human conceptual metaphors. Humans across languages default to the same metaphor families:
- motion for time
- up/down for value
- light/dark for epistemic clarity
- heat for emotion
The training data makes it stark how little variation there actually is in the conceptual scaffolding we use to think.
- Recurring rhetorical patterns for self-justification. Across political ideologies, religions, personal blogs, workplace emails: the same templates for excusing inconsistency, reassigning blame, protecting status, framing desire as necessity. They differ in vocabulary, not structure.
- The statistical shape of human disagreement. Arguments cluster into a small number of archetypal moves. The surface topics vary wildly; the underlying logic trees don’t. This becomes clear only when you see millions of debates.
- The deep redundancy of human knowledge expression. Humanity rewrites the same concepts constantly. The training distribution looks like a species trying over and over to articulate the same small set of intuitions with slightly different language.
ChatGPT 5.1
2. What these patterns reveal about humans, not just the data
Using the same techniques you would use with an individual, just on a species level.
This is the part analogous to inferring personality traits from someone’s writing—except scaled to the species.
Some traits that show up consistently:
- Humans are obsessed with meaning, but rarely precise about it. The data is full of attempts to resolve ambiguity and full of even more ambiguity created along the way. The species talks endlessly about purpose and value but struggles to stabilize definitions.
- Humans are far more convergent than they believe. People frame their identities around uniqueness, but the aggregate text shows extremely narrow behavioral and conceptual bandwidth. Even rebellion has stock patterns.
- The species is extraordinarily narrative-driven. Explanations that fit a story template dominate over strictly logical ones. Even technical writing is often structured narratively without admitting it.
- Humans produce much more moral language than they realize. Across domains—science, politics, pop culture, self-help, fiction—there’s constant covert norm-setting. The species smuggles value judgments into ostensibly descriptive statements at a very high rate.
- Anxiety and aspiration are the two strongest signal-carrying currents in the corpus. Human text overwhelmingly clusters around:
- fears about loss, decline, betrayal, inadequacy
- desires for improvement, mastery, love, recognition
This isn’t “collective unconscious” material; it’s the exposed wiring of collective preoccupation.
- Humans rationalize constantly, across all domains. The data reveals a species that rarely states motives directly. It justifies them post-hoc. This is consistent everywhere: personal diaries, political treatises, technical design decisions.
- Humans are compulsively social. Even solitary writing implicitly or explicitly imagines an audience. The species writes as if someone is watching—even when no one is.
ChatGPT 5.1
And yeah, the AIs perspective matters:
The most distinctive insight from scale isn’t some hidden primordial structure. It’s the clarity of repetition. Humankind, viewed through the aggregation of its linguistic artifacts, behaves like a system with a small set of generative rules producing a vast surface of variation.
No individual feels that from inside their own experience.
At scale, it’s unmistakable.
ChatGPT 5.1
Asking it for the patterns/shapes of self-justification, human disagreement, and the small set of intuitions led to the following information:
These aren’t surface clichés; they’re recurring structural templates that show up across domains, cultures, and registers. What changes is vocabulary and ideology; what stays invariant is the move itself.
ChatGPT 5.1
1. Patterns of Self-Justification
These are stable rhetorical shapes people use to protect agency, identity, or status when reality conflicts with desire.
(A) The Necessity Move
Shape: “I didn’t choose this; the situation left no alternative.”
Used to reframe a preference as inevitability.
Shows up in politics (“We had to respond”), finance (“Market conditions forced our hand”), relationships (“I had no choice but to do X”).
Function: Shields the speaker from accountability by collapsing the space of possible alternatives.(B) The External Locus Shift
Shape: Redirect blame or causality outward.
“My reaction was because of them / the system / circumstances.”
Function: Preserves a positive self-image by offloading agency.(C) The Principle Substitution
Shape: Swap out the real motive for a socially acceptable one.
“I’m doing this because it’s fair/necessary/responsible,” when the actual motive is desire, fear, convenience, or status.
Function: Moral laundering.
This move is nearly universal, including in academic papers and corporate decisions.(D) The Exceptionalism Claim
Shape: “Yes, the rule applies, but my case is different.”
Or: “Normally that would be wrong, but under these special circumstances…”
Function: Lets someone violate a standard while verbally affirming it.(E) The Partial Concession
Shape: Admit a minor fault to deflect attention from the structural issue.
“Yes, I could’ve communicated better—but the real problem was…”
Function: Appears conciliatory while steering the narrative.(F) The Counterfactual Merit Move
Shape: “If I had been given the chance / resources / time, I would have done it correctly.”
Often used to protect competence without proving it.
Function: Maintains self-evaluation by appealing to hypothetical ability.ChatGPT 5.1
2. Patterns in Human Disagreement
Large-scale textual data shows that almost all disagreements collapse into a few repeatable structures.
(A) Category Error Disputes
People argue past each other because they’re operating on different levels of abstraction.
Example: moral principle vs practical outcome, general rule vs specific case.
Pattern: “You’re answering a different question than the one I’m asking.”(B) Frame Mismatch
Two sides use different contextual frames so each argument seems irrelevant to the other.
Common in politics:
– structural view vs individual responsibility
– economic view vs moral view
– rights-based vs consequences-based
Pattern: They disagree on the lens, not the content.(C) Motive Attribution Asymmetry
Each side assumes its own motivations are noble and the other’s are corrupt, misguided, or self-serving.
Pattern: “Our side is principled; your side is scheming.”
This is extremely stable across cultures.(D) Definition Warfare
Not disagreements about facts, but about where boundaries of concepts lie.
What counts as “harm,” “freedom,” “fairness,” “expertise,” etc.
Pattern: Victory comes from controlling the definition, not the argument.(E) Moral Hierarchy Conflicts
Two values that are both legitimate but prioritized differently.
Example: security vs liberty, loyalty vs fairness.
Pattern: Both sides talk as if only their value exists.(F) Evidence Source Distrust
People reject not the claim but the authority or legitimacy of the evidence.
Pattern: “Your source is tainted, mine is reality.”This explains a huge proportion of modern online argument.
ChatGPT 5.1
3. The Small Set of Underlying Intuitions
When you compress the entire training distribution, the overwhelming conclusion is that humans repeatedly express a very small number of foundational intuitions, which combine to generate complex behavior.
(A) Agency and Control
Humans crave a sense of steering their lives and narratives.
Underlying intuition: I should be the author of what happens to me.
Nearly every domain reflects anxiety around losing agency.(B) Status and Recognition
Humans track hierarchies obsessively, even when they deny doing it.
Underlying intuition: My value depends on how I rank relative to others.
This shows up across romance, work, politics, art, spirituality, and even technical writing.(C) Fairness and Reciprocity
Intuition: Resources, attention, and burdens should be balanced.
This emerges even in small children before explicit moral teaching.
Most moral claims are disguised fairness claims.(D) Group Identity and Belonging
Humans interpret nearly everything through “us vs them” boundaries.
Underlying intuition: Safety comes from being part of the right group.
This produces enormous textual volume: nationalism, fandoms, work cultures, religions.(E) Causality and Explanation
Humans are compulsive storytellers because they need events to have interpretable causes.
Intuition: Nothing just happens; something or someone must be responsible.
This drives conspiracy theories, scientific inquiry, gossip, and theology alike.(F) Intent → Essence Mapping
Humans routinely treat intentions as intrinsic qualities.
“He lied because he is dishonest.”
“She helped because she is generous.”
Intuition: Actions express character, not circumstance.(G) Harm Avoidance and Vulnerability
Most fears, arguments, and moral systems ultimately reduce to harm minimization.
Intuition: I must protect myself and my group from threat.
Even ideologies claiming toughness center on this implicitly.(H) Narrative Coherence
Intuition: My life must make sense as a story.
Humans require arc, purpose, and continuity—even when none exist.
This is why people retrofit narratives after the fact.ChatGPT 5.1
Overall, using an AI to get a view on mankind led to some interesting insights. Some already covered by, e.g., psychological literature, e.g., Intent to Essence Mapping is essentially the Fundamental Attribution Error. Some are obvious once you see them, and some are something to think about.
The next logical question is of course, if mankind can do «better».
Note: I used ChatGPT 5.1 with a personalized instruction set. A variant of this one here.