When beliefs start behaving like organisms

A rumor travels through a town. It gathers followers, changes its wording, develops defenses against criticism, and survives the people who first repeated it. A scientific method does something similar: it recruits practitioners, trains descendants, and carries instructions for correcting its own mistakes.

What kind of thing are we watching? A wave? A machine? An organism?

Each answer invites a different experiment. Waves suggest transmission and interference. Machines suggest feedback and repair. Organisms suggest inheritance, competition, and the possibility that a message’s success differs from its host’s success.

Let’s climb a ladder of models. We will begin with matter that responds to its surroundings, give it memory and motion, then ask what changes when its patterns reproduce.

This is a ladder of explanatory tools, rather than a universal definition of life. Viruses, institutions, and software already make neat boundaries difficult. For beliefs, the productive question is which mechanisms are present and what they predict.

Give each participant an evidence ledger

Suppose a participant models an unknown probability $p$ with a Beta distribution, $B\sim \mathrm{Beta}(\alpha ,\beta )$. Both parameters are positive. Under a Bernoulli sampling model, independent supporting and opposing trials add to the two channels; the starting prior remains included.

The familiar Epistemic Physics coordinates are

$$\tau =\alpha +\beta ,x=\alpha -\beta ,v=\frac{x}{\tau },m=2\sqrt{\alpha \beta }.$$

Here $\tau$ is prior-inclusive trial amount, $x$ is the channel imbalance, and $v$ is the centered mean. The ordinary mean is $(1+v)/2$. The mass-like coordinate $m$ satisfies the exact identity

$${\tau }^2-x^2=m^2.$$

Two participants can have the same $v$ and radically different malleability. One has just begun considering a question. The other has weighed thousands of trials on both sides. An incoming trial shifts the first ledger much more.

Nothing in that identity tells us how neighbors influence one another. We must choose an interaction rule. That freedom is useful: the same individual bookkeeping can support several very different collective worlds.

A community that remembers a reversal

Imagine a room of people deciding whether to adopt a convention. Each person has a private inclination, feels some pressure from the room, and hears the same public message. A simple magnetic comparison assigns each person a state $s_i\in \{-1,+1\}$ and updates it toward the sign of

$$JM+H+h_i.$$

$M$ is the average state of the room, measures the influence of that average, $H$ is the public pressure, and $h_i$ is person $i$’s fixed private inclination. These binary states are an additional coarse description of behavior, separate from the continuous Beta mean.

With Gaussian private inclinations of standard deviation $R$, the infinite, fully connected model has the self-consistency relation

$$M=2\mathrm{\Phi }\left(\frac{JM+H}{R}\right)-1,$$

where $\mathrm{\Phi }$ is the standard normal cumulative distribution. At $M=H=0$, the slope of the right-hand side is $J\sqrt{2/\pi }/R$. It reaches one at

$$R_c=J\sqrt{\frac{2}{\pi }}.$$

That calculation exposes a tipping mechanism. When local inclinations are sufficiently similar, one person’s switch can push others across their thresholds. A small change in public pressure can release a large collective jump.

Slowly reverse the pressure and the room need not retrace its path. This history dependence is hysteresis. In the zero-temperature random-field Ising model, ferromagnetic couplings and suitable slow, monotone driving also support return-point memory: nested excursions return to the same state when their previous turning points are revisited. The theorem belongs to those dynamics. It offers a precise pattern to look for in human data. Sethna and colleagues’ original model develops the avalanches and memory together.

Before predicting an avalanche, ask what kind of resistance you are seeing:

A single survey cannot reliably distinguish these mechanisms. Repeated observations and carefully varied inputs help. Delayed opposition can even synchronize anti-conformists, under the conditions studied in Touboul’s model.

Competing positive and negative couplings add frustration: satisfying one relationship can prevent satisfying another. Even symmetric couplings can produce this conflict. A rugged landscape of long-lived factions becomes a plausible model to explore, with response depending on how long the group has occupied its present state. Whether real communities display such glass-like aging is an empirical question.

Try steering a small room through a reversal. Raise the public pressure, release the response, and watch whether one person’s switch unlocks others. Bookmark a settled pattern before reversing direction: a return of the headcount is weaker than a return of every individual state.

A community with memory

Turn the room, one voice at a time

Reset the room

Raise public pressure toward zero, then let the room respond. Can one switch unlock a crowd? Tap any person to inspect their private inclination.

−

Public pressure H-0.2

+

−Member 1decline−Member 2decline−Member 3decline−Member 4decline−Member 5decline−Member 6decline−Member 7decline−Member 8decline−Member 9decline

Member 5: private inclination 0.00, current total push -1.00.

Let the room respond

0 of 9 adopt. Start with a settled room that declines.

Try a memory loop

Settle at H = 0.1, bookmark the pattern, lower H to 0 and settle, then return to 0.1. Try a larger reversal too. A remembered pattern includes every member, not only the headcount.

Bookmark this settled pattern

Interaction rule and free exploration

Binary behavioral choices are an additional toy model, separate from a Beta mean. Member i responds to H + hᵢ + J times the average of the other eight states. An unhappy member switches; repeat until stable. Exact ties retain their state. All interactions are positive; no random force is used.

Neighbor coupling J = 0Neighbor coupling J = 0.4Neighbor coupling J = 0.8

Each response settles at fixed pressure. The displayed sequence chooses the first unstable member; the gold set shows the whole avalanche. History retains the last 24 responses. A collective jump follows these chosen thresholds, not from the Beta cone identity alone.

Let the particles choose where to linger

A magnet’s sites stay put. People move between conversations, publications, and feeds. Exposure becomes part of the dynamics.

In active matter, self-propelled particles can separate into dense and dilute regions when they slow sufficiently strongly in crowded places. Attraction is unnecessary in suitable models: more neighbors cause slower escape, which attracts still more occupancy. Tailleur and Cates developed this connection for run-and-tumble particles.

Here is a proposed social counterpart. Suppose a crowded discussion takes longer to leave because it contains more replies, rituals, or engaging exchanges. Longer residence increases occupancy; occupancy increases residence. A busy place can become sticky through that feedback.

The qualifier “sufficiently strongly” matters. In a local speed-density approximation, with speed $s(\rho )$ at density $\rho$, the instability condition is

Merely slowing a little does not guarantee separation. And a cluster of people becomes an echo chamber only when content, exposure, and belief also become correlated. Dwelling on agreeable content already introduces a content preference, even if nobody explicitly selects like-minded companions.

The experiment follows from the mechanism. Measure arrivals, departures, content, and residence times. Then compare changing recommendations with changing the conditions that make a conversation sticky. Similar-looking clusters can respond differently because one formed through whom people seek and another through how long they stay.

A membrane needs a maintenance budget

A living cell maintains a boundary while material flows through it. A community can also maintain selective permeability: some sources enter readily, others face scrutiny, and repeated practices repair a shared interpretation.

Attention is a useful candidate for the resource that funds this work. Retelling, teaching, checking, and rebuttal take time. Forgetting and competing demands erode the pattern. Calling this a metabolism suggests measuring both flows.

We can make one small part exact. Let background evidence arrive at unit trial rate with centered incoming bias $u$. Let a corrective source supply extra positive trials at rate $r$. The ledger then obeys

$$\dot{\tau }=1+r,\dot{x}=u+r,\dot{v}=\frac{u+r-v(1+r)}{\tau }.$$

To hold a chosen mean $v$ fixed when , the required rate is

$$r=\frac{v-u}{1-v}.$$

Near the positive edge, maintaining the same discrepancy from the background can be expensive. This is a rate balance under an explicit evidence rule. Converting it into attention, money, or metabolic energy requires a measured cost per accepted trial.

There is another catch: repetition can preserve memory without supplying independent evidence. Ten retellings of one observation share a lineage. They cannot become ten fresh trials merely because ten people repeat them. A tracelet records the evidence increment and its place in that lineage; an account of social maintenance needs to preserve that provenance.

The biological vocabulary now becomes suggestive. A stored rebuttal resembles immune memory; a small, answered challenge can resemble inoculation; an indiscriminate defense can injure the host it protects. These comparisons propose interventions and measurements. Their mechanisms must be specified before importing epidemiological thresholds.

And texts resemble spores: an interpretation can persist in a dormant record and later recruit a new community of readers. Storage still has physical costs, but active rehearsal can pause. A library becomes a wonderfully literal place to search for dormant lineages.

Enter the copying ecology

Replication adds a new question: which variants produce more copies?

Imagine a message with $L$ binary symbols. Each symbol is copied incorrectly with probability $ϵ$, independently of the others. The probability that the entire message arrives unchanged is

$$q=(1-ϵ{)}^L.$$

Let the intact message produce copies for every one produced by a mutant. Put every mutant in one background class and omit accidental mutation back to the original. If $z_t$ is the intact fraction at generation $t$, reproduction produces $\sigma z_t$ intact-parent copies and $1-z_t$ background copies. Only the fraction $q$ of the intact-parent copies stays intact. Therefore

$$z_{t+1}=\frac{\sigma q{z}_t}{\sigma z_t+(1-z_t)}.$$

This two-class model gives us an inspectable version of the mutation-selection problem introduced by Eigen. Its simplicity lets us see every moving part.

Predict before copying: if each symbol is 98% reliable and the intact message enjoys a fourfold reproductive advantage, will an 80-symbol message survive? What if each symbol gets three independent copies and a majority vote?

Build · send · revise

Pack a message for its descendants

Reset ecology

Your expedition wants at least 5% intact messages after 30 generations. Keep at least four cargo tiles, then choose whether to pay for repair. Tap a tile to trim the message to that length.

▤10▤20▤30▤40▤50▤60▤70▤80+90+100+110+120

80 binary symbols · every tile contains ten symbols

Ordinary copyingOne copy per symbolMajority repairThree copies per symbol

Send through 30 generations

10.0% intact

The colony starts with 10% intact messages. Choose a design, then send it.

See both populations and the persistence boundary

Rose: ordinary copying. Teal: majority repair. Both start at 10%. The selected design tends toward 0.0% intact in the infinite-time limit; a finite run can still be transient.

Copy onceCopy 10 generationsRun to 100

Ordinary persistence needs length below 68.6. Repair needs length below 242.8.

Change the channel and inspect the model

Message length in free exploration Error per symbol 0.5%1.0%2.0%4.0%Copying advantage 4×8×12× Charge repair its threefold copying cost

Repair error: 0.1%. Repair advantage: 1.33×. The 5% mission and four-tile cargo minimum are game choices, not natural thresholds.

$$q=(1-ϵ{)}^L,z_{t+1}=\frac{\sigma q{z}_t}{1+(\sigma -1)z_t}$$

Infinite deterministic population; independent symbol errors; one favored exact sequence; equally fit mutants; no back mutation. Repair assumes independent errors and perfect majority decoding. A design change restarts both populations.

Why there is a length ceiling

When the intact sequence is rare, the denominator in the update is almost one. Its initial growth multiplier is approximately $\sigma q$. It can invade only if

For , this gives the strict bound

The approximation uses small $ϵ$. The exact boundary for 2% error and fourfold advantage is about 68.6 symbols. Thus 68 symbols can persist in this model; 69 cannot. At equality, a positive initial intact share still decreases, although increasingly slowly.

Solving the fixed-point equation gives the long-run intact share, starting with some intact messages:

$$z_{\ast }=max(0,\frac{\sigma q-1}{\sigma -1}).$$

An error threshold concerns the persistence of the favored sequence. The population itself can continue as mutants. Allowing many fitness levels, back mutation, finite populations, or selection for meaning rather than exact wording changes the problem. Cerf’s finite-population treatment illustrates how population size enters a sharper analysis.

Now repair the channel. Three independent binary copies disagree with the original after majority decoding only if two or three copies are wrong:

$${ϵ}_{\mathrm{repair}}=3{ϵ}^2(1-ϵ)+{ϵ}^3=3{ϵ}^2-2{ϵ}^3.$$

Below a 50% error rate, this improves reliability. It also triples the transmitted symbols. The lab first compares reliability with that extra budget available, then lets you charge the intact repaired sequence by dividing its reproductive advantage by three. This is an explicit cost model; different shared repair systems would distribute costs differently.

Rhyme, repeated phrases, comparison against an archive, and communal correction invite similar questions about cultural transmission. They can help preserve a message, but the three-copy formula applies only to the independent binary-error scheme we constructed. Real language can preserve meaning despite changed words, and a single mistaken source can corrupt all three copies together.

The interesting historical hypothesis is therefore conditional: better copying and repair may permit more complex traditions to remain recognizable. Testing it requires defining the unit of information, estimating channel errors, and deciding which variations count as the same tradition.

Whose success are we measuring?

A message can be excellent at reproducing while offering little benefit to the people who carry it. A chain letter makes that distinction particularly visible. Conversely, an inconvenient scientific procedure may be costly to transmit while greatly improving decisions.

It helps to place two ledgers side by side:

One proposed audit compares predictive gains with maintenance costs. Forecast improvements can be measured using a proper score, a rule that rewards accurately stated probabilities. Attention can be measured in time. Combining those quantities requires an explicit valuation; minutes cannot simply be subtracted from a probability score. Nor does low predictive value establish low human value: poetry, rituals, and friendships serve purposes beyond prediction.

This distinction also sharpens biological metaphors. A fixed text resembles a genotype, while its enacted interpretation resembles a phenotype. Borrowing between traditions resembles horizontal transfer. A package of mutually supporting claims invites an epistasis model, in which the value of the combination differs from the sum of its parts. These are useful hypotheses about transmission and interaction, rather than labels that settle a tradition’s worth.

The organism can also be a belief engine

The analogy has a lovely reverse direction. An organism continually predicts and acts to keep itself within survivable conditions. Friston’s free-energy principle connects perception, learning, and action through an inference framework.

That makes Epistemic Physics an interesting neighbor, with a distinction worth keeping visible. For the centered Beta random variable $2B-1$,

$$\mathrm{Var}(2B-1)=\frac{1-v^2}{\tau +1},\text{inverse variance}=\frac{\tau +1}{1-v^2}.$$

Trial amount $\tau$ and inverse variance are different quantities. Even at equal trial amount, the mean affects the spread. Precision-weighted prediction errors in an active-inference model therefore require a specified probabilistic bridge before they can be identified with an evidence coordinate.

The opportunity is to build that bridge explicitly: which observations arrive, what the agent can change, what it tries to maintain, and how its action changes the next observations.

Forecasting a world that responds

Three complications make this journey especially fertile.

First, copying preserves a pattern without transferring away the original. Evidence coordinates have no automatic global conservation law under replication. Attention is a limited competing resource, but it still needs its own budget equation.

Second, influence can be directed. If one person listens more than the other, a single symmetric energy landscape may fail to describe the motion. Cycles and persistent currents become possible. Ordinary statistical physics already has tools for such driven systems.

Third, forecasts can change the population they describe. A prediction of a run on a resource may help cause it. Performative prediction formalizes learning when deploying a model changes the data distribution. A prediction stable under that response need not be the prediction that optimizes long-run outcomes.

These complications suggest practical research: distinguish slow drift from threshold switching with repeated observations; test memory using reversals; measure residence-time feedback; track copying lineages; and evaluate predictions after people have reacted to them. Logical relationships between claims also matter: social alignment alone cannot tell us whether two propositions are consistent.

The next time an idea seems alive, watch its boundary, its maintenance, and its descendants. Each offers a different window into what keeps the pattern here.

Further reading

• Sethna and colleagues: hysteresis, avalanches, and return-point memory.
• Tailleur and Cates: motion, density, and phase separation.
• Eigen: self-organization and replicating information, alongside Cerf on finite populations.
• Friston: an inference account of perception and action.
• Perdomo and colleagues: performative prediction.