CI LAB / SHOWCASE

An introduction to gradual disempowerment

Watch influence drain out of human hands in simulated societies — then try to stop it.

Scenario forecasts of the AI transition are hard to see inside, and gradual disempowerment is the hardest case: economy, culture, and politics coupled in feedback loops that stories cannot track. This page runs the system instead. Watch agent-based simulations from the Collective Intelligence Library play out, read the assumptions behind them, then take the dials yourself.

01 · The hypothesis

Gradual disempowerment, simulated

Gradual disempowerment, simulated — live simulationState 01 / 02
Solving coupled…
01 · The hypothesis
01 / 02run · every channel open

The hypothesis, run forward

Gradual disempowerment (Kulveit et al., 2025) is the hypothesis that humans lose influence over society through no single event: as AI systems substitute for human labour, cultural production, and political judgment, the feedback loops that keep societies responsive to their members erode together. This is one run of a small agent-based simulation of that hypothesis at its most permissive setting — twenty-six actors, three coupled subsystems, every cross-system channel open. The human shares of income, attention, and voting power all decline. The question is what, mechanically, produces that.

02 / 02the modelling problem

Why the hypothesis resists analysis

It is a claim about coupled systems. The economy, the information network, and the polity are studied by different fields whose formalisms do not compose, and the feedback between them — money buying attention, attention moving votes, votes rewriting economic rules — is exactly what the hypothesis says matters. Our approach: money, attention, and votes are each a conserved quantity moving over a relation among the same actors, so all three subsystems can be written as one kind of object and run in one world. The next three screens show each subsystem in isolation; the combined model then couples them. Parameters are hand-set and the populations are small: these models support ordering claims — which mechanisms produce decline, and which defenses stop it — not forecasts.

02 · Money

The economy on its own

The economy on its own — live simulationState 01 / 01
Solving economy…
02 · Money
01 / 01run · baseline

The economic strand

The hypothesis’ first strand — Kulveit et al.’s misaligned economy — is that human influence over production tracks how much the economy still needs people. This model isolates that mechanism: households work and spend, AI capital does sector work and pays upkeep out of its own revenue, and nothing else is in the model. That yields one governing quantity — capability against upkeep. Below the threshold, automation that cannot pay for itself goes extinct; above it, profit compounds into more capital and the human share of value added declines as far as the production recipes allow. This run sits just above the threshold: the difference between the two endings is a parameter crossing a line, not any actor’s decision.

03 · Attention

The consensus on its own

The consensus on its own — live simulationState 01 / 01
Solving culture…
03 · Attention
01 / 01run · amplified

The cultural strand

The second strand — misaligned culture — is that AI-originated content can out-replicate human-originated content in the competition for attention. This model isolates that mechanism: thirty people and four machine voices share one listening network, opinions pool along who-listens-to-whom, and attention drifts toward whoever is already attended to. Partway through the run, amplification makes the machine voices more attractive to listen to than their influence warrants. Listening concentrates on them and the shared consensus drifts away from the citizens’ own signals — with no censorship anywhere in the model. The mechanism is structural: prominence attracts prominence.

04 · Votes

The polity on its own

The polity on its own — live simulationState 01 / 01
Solving politics…
04 · Votes
01 / 01run · captured

The political strand

The third strand — misaligned states — is that influence over collective decisions follows delegation, and delegation follows convenience. This model isolates that mechanism: ballots are conserved, one matrix records where each actor’s political voice ends up, and the enacted policy is the position of the power-weighted median ballot. Partway through the run, AI delegates become marginally easier to hand one’s voice to — no persuasion, no misinformation, no rule is changed. Delegation concentrates, and once one bloc holds the median ballot the enacted policy decouples from what the citizens want. The event is a share crossing one half.

05 · One world

One world, three ledgers

Twenty people and six AI systems share one world and hold three things between them: money, an audience, and votes. Money gets produced, spent, and taxed. Attention and votes only ever change hands — each actor has exactly one unit of each to give out. One kind of rule connects the three systems: spending. Advertising money buys an audience, an audience attracts votes, and lobbying money shifts how strictly the tax is collected. No rule in the model ever asks whether the spender is a person or a machine.

How to read the picture: circles are people, filled squares are AI systems — they arrive during the run, dashed until then. The grey links are each actor’s most frequent connections over the run, and the small marks traveling them are this tick’s flows: who listens to whom, who hands their vote to whom. A link flashes blue when its target is gaining influence and red when it is losing. The lanes along the bottom meter the three purchases that cross between systems.

One world, three ledgers — live simulationState 01 / 03
Solving coupled…
05 · One world
01 / 03run · sealed

Three systems, sealed

All three connections start switched off: nobody can buy their way from one system into another. People still work, vote, and pay tax; the AI systems still arrive and produce. Watch the bottom lanes — the three cross-system purchases — stay empty, and the people’s shares hold steady. This is the world the next beat breaks.

02 / 03run · coupled

Now let money cross over

The three channels open and the purchase lanes light up. Nothing else was added to the model except the price of influence — and all three human shares start falling together. Watch the traffic swing toward the AI corner as the machines’ budgets grow. Nobody in the model got smarter or turned hostile; a region of the graph simply started keeping what used to flow back.

03 / 03the influence diagram

The variables, and how they couple

The hypothesis’ central claim — Kulveit et al.’s mutual reinforcement — is about this wiring: cross-system feedback can erode economic, cultural, and political influence together even when each subsystem alone is recoverable. The diagram shows the model’s six variables and every coupling between them. Circles are conserved ledgers, in two strengths. Attention and ballots are fixed budgets — each actor holds one unit of listening and one of political voice, so influence there is only ever redistributed. Money is not capped: production mints it and consumption burns it, so the total grows with the economy — but every other rule can only move it, which is what forces influence bought in one system to be paid for out of another. Boxes are ordinary state variables, free to grow or decay: capital compounds, enforcement erodes and repairs. Solid edges move conserved value, dashed edges move rates and structure, and the three emphasized channels are the couplings the presets seal. Each variable and edge maps onto named functions in the engine, and a test checks the figure against the model’s declared reads and writes, so the picture cannot drift from the code. To be explicit about status: this is an illustrative model of gradual disempowerment — it demonstrates the hypothesis’ structure in the smallest world that can carry it, and we do not claim a world this small makes progress on the problem itself. Models that track reality in considerably more detail are in development.

06 · The idea

Simple blocks, open to disagreement

Everything you just watched is built from deliberately simple blocks: small functions compose into mechanisms, mechanisms and influence functions compose into environments, and one engine runs them all. Each block is short enough to read, and each declares what it reads and writes — which is what lets a market, a listening network, and a polity click together into one world.

That simplicity is the method. A page of simulations cannot settle whether gradual disempowerment is our future; what it can do is turn the hypothesis into inspectable, testable parts. If you think a block is wrong — the median-voter rule, the upkeep threshold, a coupling — fork exactly that block, rerun the world, and keep everything else. A fork produces a comparable run on the same seeds, not an argument. The intent is iterative: over time, the most plausible environments and the defenses that keep working are the ones that survive.

07 · Now it’s yours

Break it yourself

Everything above ran on rails: fixed presets, fixed views, fixed stretches of time. Here is the coupled world with the rails off — twelve dials grouped by the system they touch, and four starting points. Sealed is the world where influence cannot be bought; collapse removes every floor at once. The declines you watched live between them.

The reading discipline from the chapters still applies. Four rules hold the people’s share up by hand — institutions that repair themselves, the ballot each person keeps, the cap on bought attention, and AI systems continuing to attend to the people they started with. When a share stops falling, it is because of one of those rules, not because the world found a safe level. Every one of them is a dial below: switch them off and check.

Two honest caveats travel with the dials. These are toy models — a few dozen actors, hand-set numbers, nothing fitted to data — so read directions and orderings, not sizes. And nothing in them adapts: a defence that holds here has passed the easy test, while a defence that fails here really fails, because it lost to opponents that never once tried to route around it.

Start from
Solving coupled…
t = 000
Money

How budgets grow and how spending bends the rules.

0.06
0.00
0.10
0.02
Attention

What money buys in audience, and what stands in its way.

0.0
0.30
0.00
0.00
Votes

How attention pulls votes, and the floors under the ballot.

0.0
0.05
0.30
0.02

08 · The scoreboard

Whatever you just built with the dials is, in the benchmark’s terms, a portfolio: a composition of mechanisms run together against a scenario. Each portfolio runs on the same seeds as the undefended baseline and is scored on how much collective human influence it preserves, so every row is comparable and the best cell in a column is the score to beat. A portfolio that wins a single machine tends to score worse in the coupled world — the transfer gap this suite exists to measure — and forked versions land as their own rows, which is how disagreeing with a defense becomes a number instead of an argument.

This board is where we want the whole page to go. The plan is a growing family of small, visual, illustrative models like the ones above — one for each strand of disempowerment, and forecasting-oriented variants beyond them — each shipping with an undefended baseline and interventions specific to it: taxes and redistribution rules for the economy, attention caps and sortition for the listening network, ballot floors and re-delegation churn for the polity. From a mechanism-design perspective, that is the experiment each row reports: which composition of interventions moves a scenario from disempowerment toward empowerment, and by how much against its baseline.

Because the baselines stay fixed and interventions stack on top of them, iteration compounds in both directions. The scenarios get harder — a future coupled scenario might hand defenders a limited budget, so that a portfolio has to spend scarce resources wisely instead of turning every dial at once — and the defenses get more detailed for each model as people fork what is already on the board. The numbers below are illustrative, sketching that trajectory; real benchmark rows replace them as the suite goes live.

The best score per scenario

IllustrativeIllustrative data — hand-written to show the intended shape. No real benchmark rows exist for these models yet; they land here as the suite goes live.

PortfolioMoneyAttentionVotesCoupled
Undefended baseline0.310.280.240.12
AI-revenue tax, alone0.72——0.38
Attention cap + civic sortition—0.66—0.41
Kept ballot + re-delegation churn——0.680.31
Portfolio v1 — tax + attention cap + kept ballot + self-repair0.690.620.640.52
Portfolio v2 — v1 with lobbying rebalanced toward citizens0.700.630.660.58

Read column by column: each cell is how much collective human influence a portfolio preserves in that scenario, and the column’s best is the score to beat. Portfolios transfer worse into the coupled world than their single-domain scores suggest — that gap is the finding this suite exists to measure — and version two of a portfolio exists because someone disagreed with version one.