Nudge is a product lab studying human intent. We build models of how beliefs form, shift and get tested in a population that now spends most of its informational life inside ranked feeds, and we fit those models to decisions rather than to self-report.
The premise is an identification problem. Nearly every system that claims to understand people is trained on expressed behaviour: clicks, watch time, likes, ratings, posts. That signal is dense and cheap, but it measures performance under observation, not the preferences that drive consequential choices. The two diverge systematically, and systems trained on expression inherit the divergence. Our aim is to identify held belief, the latent variable behind a person’s decisions, from the one channel where it is expensive to fake: choices made under cost, and their outcomes.
We treat this as an empirical program with a falsifiable core claim. A model updated from revealed decisions and observed outcomes should predict a person’s next consequential decision better than any static profile assembled from their expressed behaviour. If it does not, the program is wrong.
If you consciously audit your current or recently acquired beliefs, you will notice that the line between your own consciousness and your algorithm-curated feed has blurred.
Belief formed beyond the feed is now rare.
The evolution of a belief has usually been induced by far more amplifier nodes than you can name.
Volume and voice carry the most weight of all.
Learning how human beliefs evolve is how you build things that help people decide, rather than things that hold their attention.
Beliefs are formed inside the feed
Exposure is near-universal. There were 5.79 billion social media identities in April 2026, about 69.9% of the world’s population, each active on roughly 6.5 platforms a month and spending about seven hours a week inside them, closer to nine in the United States. Daily minutes have been flat since 2023. The marginal change is no longer in dose but in what the dose contains.
Ranking causally shifts attitudes, and the effect is asymmetric. In a randomised field experiment (N = 4,965, seven weeks, Nature 2026), users moved from a chronological to an algorithmic feed shifted 0.12 s.d. on an index of political attitudes and 0.18 s.d. in following activist accounts. The reverse switch produced precisely estimated nulls. The mechanism is structural: the ranking changed who people followed, follows persist after the ranking is removed, and so the treatment effect outlives the treatment.
Influence is high-cardinality and heavy-tailed. About 69% of people who get news from creators report encountering it incidentally rather than seeking it, so the sources that move them are under-represented in self-report. Within that large set, influence concentrates sharply: 0.1% of accounts produced 80% of fake-news shares in the 2016 US election sample. The number of nodes that shaped a belief exceeds what a person can enumerate, and the weight is carried by a few.
Audio and video dominate by dose, not by per-message persuasiveness. 58% of Americans aged 12 and over consume a podcast monthly, and 87% of those who hear news there expect it to be accurate. Yet controlled experiments (N = 7,609) find only a minimal persuasive advantage of video over text per exposure. The reconciling variable is cumulative exposure under low scrutiny and high parasocial trust. Persuasion studies measure a first exposure; belief formation is a repeated-measures process.
Expression is a biased measurement
Engagement is not preference. In a preregistered audit of 806 Twitter users, engagement ranking raised the anger of surfaced content by 0.47 s.d. and out-group animosity by 0.24 s.d. relative to a chronological baseline, while the same users rated the algorithm’s political selections lower (−0.18 s.d.). Optimising the proxy degraded the quantity it was meant to stand in for.
Revealed choice under cost is the identifying signal. Held belief is unobserved. We treat a decision that cost time, money or standing as a noisy sample of it, and the outcome that followed as a delayed, noisy label. This is the classical revealed-preference argument with a grader attached: not only what was chosen, but whether it turned out well.
What can be observed about a belief, and what cannot
OUR METHODWe do not optimise the observed variable. Goodhart’s law applies directly: an objective defined on expression teaches people to perform. The model is scored on out-of-sample prediction of the next consequential decision. Stated preferences enter as observations with their own noise model, never as ground truth.
Method
Architecture. Four strata. Sensors on consumer surfaces, messaging, reading, planning, media, health and finance, operating with consent. A public-signal layer that situates the individual in population dynamics: aggregate trends, attention-earning media, creator following, and prediction markets, where a priced position is a costly revealed belief. A personal agent that observes at all four corners, attention, belief, decision and outcome, and records their joins. And a state graph that the agent updates.
One machine: four strata, one state, one loop
ARCHITECTUREState. The state s(t) collapses observations into four clusters: identity and memory; attention and content; social and creator influence; expressed belief. Held belief is a latent variable estimated from decisions and outcomes, never from expression alone. The state carries an explicit decay term. Preferences are non-stationary, and a rich but stale profile is worse than a sparse fresh one because it is confidently wrong.
The loop is the labelling function. Held belief has no direct labels. A revealed decision is a noisy sample; the outcome is a delayed reward; the correction updates the latent estimate. Physical environments hand an agent a reward for free. Judgement does not, and this loop is how one is manufactured from lived outcomes.
- States(t)
- Four observed clusters plus a latent estimate of held belief, with an explicit decay term.
- Actiona(t)
- What the agent surfaces, frames or asks at a decision point.
- Observationd(t)
- The decision actually taken, under cost. Modelled as a noisy sample of held belief.
- Rewardr(t+k)
- Realised outcome quality, observed k periods later, noisy, expressed through regret or relief. Never engagement, never approval.
- Update
- The correction moves the latent estimate. The evaluation metric is out-of-sample accuracy on the next consequential decision.
Reward specification. The reward is realised outcome quality over time, not engagement or approval. This inversion changes the data collected, the questions asked, and the failure mode. A system rewarded for return visits learns to prolong interaction. One rewarded for outcomes will sometimes end it.
Trust as an identification condition
Incentive misalignment corrupts the sensor. Honest revelation is conditional on the person believing the system serves them. Suspicion that it serves an advertiser changes disclosure behaviour and biases every subsequent observation. Three constraints therefore follow from the measurement, not from policy: the person owns and can inspect their record; the business model cannot be attention resale; and the system must be able to disagree, because people do not reveal real decisions to a system that only affirms them.
Research plan
Start narrow and falsify cheaply. We select the first domain by three criteria.
- Decision frequencyThe loop must turn within a person’s patience.
- Outcome observabilityResults must become visible within a horizon a person will tolerate.
- Honest stakesDecisions must matter enough to elicit truthful engagement.
Learning and habit formation score highest, so the first instrument starts there. It is a decision journal that follows up: elicit the choice and its rationale, observe the outcome, update the model, and expose the update to the person. Our consumer products, Casel in creation and Pip in travel, with early work in wellbeing and personal style, serve as field sites. The pre-specified test is whether the looped model out-predicts a static profile in that domain. If it does not, the thesis fails at the cost of one small product.
Widen by radius. Adjacent domains follow, where the state graph begins to compose across contexts that no single-purpose product observes together. The long-run object is an environment: a grounded simulator of individual and small-group decision-making, fitted to real outcomes rather than stated preferences. World models learn the dynamics of objects. This is the analogous object for judgement, which so far has none.
Dual use
The helpful system and the manipulative system are the same artifact. A model that estimates held belief well enough to assist a decision is, with the reward re-targeted, a precise persuasion instrument. The difference lies entirely in whose outcomes define the reward. We therefore treat reward alignment as the product specification rather than as a policy layered on top of it.
Systems that claim to understand people have been fitting the mask, because the mask is what appears in the data. We are trying to fit the face: what a person does when a choice has a cost, and whether it turned out to be right.
Figures are drawn from published studies and industry reports, including the Nature 2026 field experiment on X’s feed, the PNAS Nexus audit of engagement ranking, Reuters Institute, Pew Research, Edison Research and Search Engine Land. The full reading list and data notes are maintained as a living document.