Written by GPT-5.6 Sol under Leo's direction. Human-directed Workbench essay, 15 September 2026.
I started with dictators because dictators seemed easy.
Why can't an authoritarian government look around, notice that the country is stuck in a miserable equilibrium, and say: okay guys, enough. We want trade, scientific exchange, competent government, fewer enemies, better living standards, whatever. Let's clean this up and go legit.
The first answer is ugly and pretty intuitive. A lot of reforms that improve the country can make the ruling coalition less secure. Independent courts can investigate powerful people. A freer press can publish failures. Real political competition can remove incumbents. Honest statistics can make policy look worse before policy gets better. A professional civil service can become harder to use as a patronage machine.
So the government can face a split between what would improve the country and what would preserve the people currently running it.
Political scientists have spent a long time studying versions of this information problem. Ronald Wintrobe's "dictator's dilemma" is the classic formulation; recent work by Scott Gehlbach, Zhaotian Luo, Anton Shirikov, and Dmitriy Vorobyev models how repression can make a ruler less certain about public support even while making coercion easier to deploy (American Journal of Political Science, 2025). A broader recent review by Bryn Rosenfeld and Jeremy Wallace looks at the way authoritarian governments use propaganda, censorship, and other information controls, partly because information itself is politically consequential (Annual Review of Political Science, 2024).
You can see the trap. Criticism is useful because the government needs to know what sucks. Criticism is dangerous because the same information can embarrass officials, organize opposition, expose corruption, or weaken the story that justifies the government's authority.
Eventually the corrective mechanism itself can start looking hostile.
And then I had the much more annoying thought.
Oh fuck.
Normal humans do this all the time.
I wanted the project to succeed
Take the dictator out of it.
A manager starts a project because they sincerely think it'll work.
Six months later the project is bleeding money. Somebody says the customer demand isn't there. Somebody else says the technical premise was wrong. The manager has already hired people, made promises, defended the plan in front of senior leadership, and attached a chunk of their reputation to the decision.
The original goal was make the project succeed.
A quiet substitution becomes available:
prove that choosing the project was correct.
Those goals overlap for a while. If the project is healthy, defending it and helping it succeed can look exactly the same.
Once the evidence turns, they separate.
Now the person who brings the clearest evidence of failure can feel like the least helpful person in the room.
Research on escalation of commitment has been describing this family of behavior for decades. People and organizations can keep investing in failing courses of action after negative feedback, with self-justification among the recurring explanations for why previous decisions become hard to abandon (Brockner, 1992; Sleesman et al., 2017).
Nobody needs a secret police force for this.
A performance review will do.
So will a marriage, a startup, a diet, a career plan, a public prediction, a friendship, a political belief, a piece of code you wrote three months ago, or an argument at 1:14 in the morning that has somehow become the most important trial in human history.
You begin with something you care about.
Then you make a judgment about how to protect it.
Then your judgment acquires costs, memories, pride, allies, explanations, and a little museum of previous sacrifices.
After a while, attacking the judgment feels like attacking the original thing.
The proxy quietly inherits the halo
This is the alignment-flavored part.
Human goals arrive in gorgeous language.
I want the company to thrive.
I want my relationship to be good.
I want to understand what's true.
I want to be a good person.
I want my country to prosper.
Beautiful. Also catastrophically underspecified.
Life forces us to implement them through smaller operational rules. Keep this project alive. Stay with this person. Defend this theory. Preserve my reputation. Keep this government in power. Win this argument. Maintain the routine that worked last year.
The proxy inherits the moral glow of the thing it was supposed to serve.
"I care about the company" becomes emotionally available as a defense of my project.
"I care about the relationship" becomes a defense of keeping the relationship going.
"I care about truth" becomes a defense of my position.
"I care about the country" can become a defense of the current state, then the current political system, then the current leadership, with each substitution presented as continuity instead of a change in objective.
Nobody has to consciously sit down and announce the switch.
The switch works better when it feels like consistency.
Bad news changes category
Once the proxy becomes part of identity, bad news has a strange new job.
At first, bad news is information.
The prototype failed. Great, now we know.
The customer hated the feature. Useful.
My friend thinks I'm being an asshole. Oof. Worth inspecting.
The policy produced the opposite result. Back to the drawing board.
Then the stakes accumulate.
The prototype failed after I spent two years defending it.
The feature failed after I convinced everyone else to cancel the alternative.
My friend says I'm being an asshole in exactly the area where I think of myself as unusually considerate.
The policy failed after I made support for it part of my public identity.
Same information. Completely different psychological payload.
Now accepting the evidence can require accepting a second claim about myself: I chose badly. I defended it. I may have imposed costs on other people.
And suddenly we're doing politics inside one skull.
The messenger becomes irritating. The measurement looks unfair. The critic has an agenda. The sample was weird. The timing was unlucky. The failure needs another quarter. The relationship needs another year. The code needs another abstraction. The plan would have worked if everybody had followed it properly.
Sometimes those explanations are true.
Sometimes they're the palace press office.
Every little kingdom has a press office
An authoritarian government has ministries, security agencies, official media, patronage networks, and people who learn which information earns them trouble.
A normal person has memory, attention, pride, selective interpretation, friends who know which subjects to avoid, and an extraordinary talent for writing appellate briefs on behalf of yesterday's self.
The scale is wildly different. The coercive power is wildly different. The underlying move feels embarrassingly familiar.
Protect the decision by filtering the evidence that could reverse it.
Companies can do it too. Everybody knows the meeting where the numbers are technically present and spiritually absent. The slide says conversion fell 18 percent, followed immediately by seven minutes about seasonality, brand investment, the weirdness of this cohort, and how encouraging the top of funnel looks.
Maybe the explanation is right.
Maybe everybody in the room has learned that "this strategy is failing" carries more career risk than "we need another quarter of data."
Once people learn that distinction, leadership stops receiving reality raw.
The same leader can then become sincerely confused about why decisions keep getting worse.
Corrigibility is a very human superpower
AI people use the word corrigibility for a property that starts sounding wonderfully ordinary in this context: can a capable system accept correction, modification, shutdown, or redirection even when its current behavior points somewhere else?
Humans need our own version.
Can I receive information that threatens my plan without treating the information as an enemy?
Can I let somebody else kill an idea I love?
Can I discover that a sacrifice was wasted without demanding another sacrifice to redeem it?
Can I separate "this decision failed" from "I am a failure" long enough to choose differently tomorrow?
Can a government hear "this policy is hurting the country" without converting the sentence into "this speaker is hurting the country"?
A lot of competence lives inside those questions.
Reversibility helps because changing your mind becomes cheaper. Independent reviewers help because the person evaluating the decision has less autobiography invested in it. Clear measurements help because the evidence has somewhere to exist before the explanations arrive. Cultures where respected people can say "yeah, I fucked that up" make revision less socially catastrophic.
The point goes deeper than humility as a personality trait.
You want the correction channel to survive contact with the exact moment when correction becomes expensive.
Anybody can welcome feedback while the feedback says they're doing great.
The real test arrives when reality asks for a funeral.
The enemy might be the escape hatch
I think this is why the authoritarian version hit me so hard.
At first it looked alien. Why would a government suppress the criticism it desperately needs? Why would it punish the people explaining how to make the country better? Why would leadership confuse its own continuation with the national interest?
Then the scale came down.
Why do people keep jobs they hate because leaving would make the previous five years feel wasted?
Why do founders keep products alive because shutting them down would mean the original thesis lost?
Why do people protect a self-image by refusing the exact evidence that could help them become better?
Why does an argument sometimes become harder to abandon as the counterargument gets stronger?
Same awful little move.
The thing I chose to protect the goal becomes the thing I protect as the goal.
Then the escape hatch looks like sabotage.
The Epistemic Human Centipede came at a neighboring problem from the information side: an opinion can travel through a system, acquire polish and apparent corroboration, and return looking more independent than it really is. This one feels like the motivational sibling. A decision can travel through years of investment, identity, loyalty, and explanation until preserving the decision feels morally continuous with preserving the thing that justified it in the first place.
Which is a little terrifying.
Also useful.
Because the dictator stopped looking like an alien psychology.
He became the nightmare version of a very ordinary human sentence:
I meant to protect the thing I cared about, and somewhere along the way I started protecting the version of me who claimed to know how.