And then, one Thursday, nearly two thousand years after one man had been nailed to a tree for saying how great it would be to be nice to people for a change, a girl sitting on her own in a small café in Rickmansworth suddenly realized what it was that had been going wrong all this time, and she finally knew how the world could be made a good and happy place. This time it was right, it would work, and no one would have to get nailed to anything …
Sadly, however, before she could get to a phone to tell anyone about it, a terrible stupid catastrophe occurred, and the idea was lost for ever.
«The Hitchhikers Guide to the Galaxy» by Douglas Adams
Given that LLMs are trained on lots and lots of human knowledge, these AIs have a unique view on mankind. Or rather, that view is represented in them. After asking ChatGPT for mankind’s patterns that emerge only at scale and what these patterns reveal about humans, the next logical question was whether mankind can do better.
The answer is a sobering but constructive realistic one:
“Can humans do better?” only makes sense once you specify the target. If the question is whether humans can escape these patterns, the answer is no. They’re structural features of the cognitive system, not accidental flaws. Trying to “get out” of them is like trying to think without categories or perceive without contrast. You don’t exit them; you work within or against them with awareness.
But if the question is whether humans can operate at a higher level of refinement within the constraints, then yes—there are real exceptions. They don’t break the patterns; they handle them differently. That’s where the signal lies.
ChatGPT 5.1
It then made the following suggestions:
1. Self-Justification Patterns
These are unavoidable. The mind protects its identity and agency automatically. You don’t eliminate that reflex, but you can override it.
Positive deviations
A. Self-disconfimation: Occasional direct acknowledgment of motives without laundering.
“I wanted this because it was easier, not because it was principled.”
Rare but possible. It contradicts the brain’s normal self-protection loop.B. Counter-evidence adoption: Letting an external constraint revise one’s self-concept, not just an opinion.
“When I look at the pattern of my choices, the story I told myself about why I did them doesn’t hold.”
This is extremely rare. Most people fold new evidence into old narratives.C. Intentional motive-mapping: Tracing reactions back to actual incentives rather than moral claims.
“I said X because I wanted Y outcome, not because I believed Z.”
Humans can do this periodically, though not consistently.ChatGPT 5.1
2. Human Disagreement Patterns
These persist because they’re downstream of cognitive architecture (abstraction mismatch, frame selection, identity protection). But they can be metabolized differently.
Positive deviations
A. Frame-switching competence: Being able to inhabit someone else’s frame without losing self-coherence.
Not agreeing—just modeling the internal logic.B. Conceptual boundary precision: Explicitly defining terms and holding them stable across the debate.
This eliminates ~70% of common disputes.C. Asymmetric empathy without motive attribution: Understanding someone’s reasoning without importing moral judgment about why they hold it.
These are more skills than traits. They can be trained but go against default cognitive shortcuts.
ChatGPT 5.1
3. Underlying Intuitions
These are the deepest because they’re tied to evolutionary machinery. But even here, “better” doesn’t mean escape—it means non-reflexive use.
Positive deviations
A. Agency without control narrative: Acting decisively without needing a story that says “I’m in charge of events.”
B. Status-awareness without status-seeking: Recognizing the hierarchy without orienting behavior around climbing it.
C. Fairness without retributive instinct: Upholding reciprocity without adding the desire to punish.
D. Identity without ingroup absolutism: Belonging to groups without importing outgroup threat narratives.
E. Causality mapping without intentional fallacy: Distinguishing between what is caused and what has agency.
F. Narrative without self-deception: Using story structure to make life legible, while knowing it’s a scaffold, not a revelation.
These are possible but require sustained cognitive effort. Most humans can touch these states but not maintain them.
ChatGPT 5.1
with the caveat of:
There’s no categorical escape. Human cognition is built from these constraints. But there is a kind of liberation: not freedom from the patterns, but freedom within them—when you see them as structures rather than personal truths.
ChatGPT 5.1
On a more practical level, I asked it regarding the mismatch of humanities power and these limitations, and it agreed (not surprising giving ChatGPT’s inbuilt agreement bias).
The worry you’re articulating is the right one: the cognitive machinery that worked tolerably at small scales becomes lethal once a species gains planetary leverage. The danger isn’t malice; it’s structural blindness amplified by power.
ChatGPT 5.1
and provided the following suggestions:
If you’re asking how to prevent self-termination given that humans cannot transcend their architecture, then the answer has to work with the constraints, not against them. There are two fronts—individual and collective—and both have limits and possibilities.
ChatGPT 5.1
1. The Individual Level
Humans do not become bias-free. But individuals can become predictably less captured by the worst dynamics. The relevant upgrades are cognitive, not moral.
(A) Meta-frame competence
Humans who can see that they are in a frame—and switch frames—don’t escalate conflicts to existential levels.
This reduces the risk of runaway misinterpretation.(B) Clean motive-mapping
When individuals can accurately parse their own incentives, they stop mixing moral rhetoric with strategic interest.
That alone avoids many self-reinforcing spirals.(C) Impulse sequestration
You can’t remove emotional drives, but you can separate first reaction from actual action.
This is the mental equivalent of adding buffers in an unstable engineering system.These are all rare traits, but they’re teachable in principles (not universally, but at scale enough to matter).
The catch: individual upgrades don’t matter unless they accumulate in positions with leverage.
ChatGPT 5.1
2. The Collective Level
Human groups don’t behave like scaled-up individuals. They behave like unstable dynamical systems:
– misaligned incentives
– amplified status competition
– runaway signaling
– feedback loops of distrust
– narrative captureTo reduce existential risk, you need structures that reliably dampen those dynamics. Not “better people,” but better systemic constraints.
(A) Incentive architecture that penalizes escalation
Most historic catastrophes trace back to misaligned reward systems: leaders gain from risk; populations pay for loss.
Reversing that asymmetry—making escalatory moves personally costly—has real impact.(B) Institutionalized adversarial cooperation
You mentioned viewpoint diversity and oppositional collaboration. Those are prototypes of a larger idea: systems where disagreement is designed into decision-making, so blind spots collide early rather than late.
Think:
– independent review with veto powers
– epistemic checks that cannot be politically captured
– structural separation between narrative production and policy making
The point isn’t niceness; it’s forcing errors to cancel rather than compound.(C) Slow variables protected from fast variables
Most civilizations fail by letting short-term political or emotional dynamics eat the slow, stabilizing ones.
You need carved-out domains (treaties, scientific institutions, infrastructure planning) insulated from the news cycle, populism, and crisis rhetoric.(D) Distributed control rather than monolithic decision cascades
Centralized command amplifies individual error.
Systems with multiple independent levers—each requiring different failure modes to collapse simultaneously—are far more robust.
Nuclear command-and-control, air travel safety, and some financial regulation already use this logic. It can be expanded.(E) Norms that separate existential conflict from symbolic conflict
Humans escalate symbolic disagreements into identity wars extremely quickly.
Societies that survive long periods of complexity create stable distinctions between what is actually worth fighting about and what is just ideological noise.
Right now, that distinction is eroding globally.ChatGPT 5.1
3. What cannot be prevented
There’s no trajectory where humans become perfectly rational or bias-free. What is achievable is reducing the probability that normal cognitive failures scale into civilization-ending events.
Most of the existential risk comes from three amplifiers:
1. Speed (decisions made faster than they can be checked)
2. Coupling (everything connected so local error becomes systemic)
3. Monoculture of assumptions (everyone thinking the same wrong thing)Mitigation isn’t psychological; it’s architectural:
- slower decision pipelines around dangerous domains
- decoupled systems with firebreaks
- intentional epistemic pluralism
That’s not utopian—it’s engineering.
4. The quiet, uncomfortable truth
You can’t fix “human nature,” but you can build systems that:
- make failure less catastrophic
- make bias less amplified
- make conflict less brittle
- make accidental escalation less likely
- slow down the chain of events that leads from error to disaster
This is exactly what humans have done in narrow domains (nuclear stability, aircraft safety, epidemiology), but not yet in the domains where cultural cognition operates.
Not exactly encouraging, giving how tribal we are. The same limitations we have make this kind of counter very hard to achieve.
But still, there is a way to act better. Whether we do it depends on us. On whether we can address our own design constraints of
Assume, as design constraints, that humans:
- are tribal and status-seeking,
- prefer good stories to true models,
- are bad at probabilistic reasoning,
- rationalize after the fact,
- fatigue quickly with complexity,
- can be decent reasoners locally with the right scaffolding.
So instead of “How do we get better people?” the question is: Given humans like this, what structures make it hard for them to scale their worst tendencies?
We can work with that.
Note: I used ChatGPT 5.1 with a personalized instruction set. A variant of this one here.