Field station 01 Platform incentives × attention design

Your attention is not simply taken. It is bid for, shaped, and allocated.

This observatory follows one scarce minute through the machinery of a feed: the business model that values it, the ranking system that predicts it, the interface that extends it, and the person who can still redirect it.

Enter the simulator Download the Field Guide

No diagnosis. No tracking. The experiment runs entirely in your browser.

Fig. A A minute has no built-in destination. Systems assign value; interfaces arrange the path; people and institutions can change the route.

SCARCE INPUT: HUMAN ATTENTION OPTIMIZATION TARGETS ARE CHOICES ENGAGEMENT IS NOT THE SAME AS VALUE AGENCY IS SHAPED, NOT ERASED

Interactive exhibit / 01

The Attention Budget Simulator

Configure a fictional platform and watch a 120-minute attention budget move between chosen use and momentum use. The model is a transparent thought experiment—not a behavioural prediction.

Illustrative · unvalidated · local only

1. Set the incentive Primary funding logic Ranking objective

2. Shape the route

Daily sandbox 120 min

Return cues per day 6

Natural stopping cues 1

AutoplayNext item begins without a fresh decision. Continuous feedNo visible end to the available inventory. Session intention promptAsk “what are you here to do?” before entry.

Settings are stored only in this browser. Clear them with “Reset model.”

Illustrative clearing result

A high-momentum route

Advertising logic, watch-time ranking, six return cues, and few stopping points concentrate the model toward continued use.

Each token represents one twenty-fourth of the selected sandbox, not measured behaviour.

Chosen-use minutes

—

Minutes the model leaves aligned with the visitor’s starting purpose.

Momentum minutes

—

Minutes the model attributes to cues, continuation, and weak stopping points.

Return attempts acted on

—

A fictional fraction of return cues, adjusted by friction and intent prompts.

Platform yield proxy

—

A direction-of-incentive label, not revenue or welfare.

Show the causal model behind this result

ABase pull begins at 8% of the sandbox. This is a design assumption, not an empirical baseline.

BContinuation pressure rises with cues, autoplay, continuous inventory, ad funding, and dwell ranking.

CExit support rises with stopping cues, a session-intent prompt, chronology, and non-ad funding.

DMomentum share = base pull + 55% of net pressure, capped between 5% and 78%.

Momentum = selected budget × illustrative momentum share.

Changing a control shows the direction this model assumes—not the size of a real-world causal effect. Real effects vary across people, cultures, purposes, content, devices, and time.

Static worked scenario

JavaScript is off, so the live controls are unavailable. Here is the core experiment in print: begin with 120 minutes. Add directional pressure for advertising, predicted watch time, six return cues, autoplay, and continuous inventory. Subtract pressure for one stopping cue. The model then labels a larger share “momentum use.” If you instead choose subscription funding, chronological ranking, four stopping cues, and an intention prompt, the direction reverses. This demonstrates the model’s assumptions—not a forecast of any person.

Use the downloadable Field Guide to calculate a scenario by hand and challenge any coefficient you disagree with.

Field notes / 02

What is actually being traded?

“Attention economy” is a useful lens, not a complete theory of human life. It names a recurring allocation problem: information is abundant, but the time and cognitive capacity available to receive it are limited.

Herbert Simon’s 1971 account of an “information-rich world” treated attention as the scarce counterpart to abundant information. The point was not that attention literally behaves like money. It was that every message imposes a receiving cost: someone must notice, interpret, compare, or ignore it. Digital platforms made that old allocation problem measurable at enormous scale. A view, pause, replay, comment, swipe, or return can become a signal in a ranking system and, under some business models, part of the path to revenue.

Four layers, one feed

Incentive layer. A platform has to sustain itself. Advertising can make additional opportunities to display or target ads economically valuable. Subscription models can prioritize retention and perceived ongoing value. Public-interest or cooperative systems may define success around access, learning, or community goals. None of these logics mechanically determines a design, but each makes some metrics easier to defend internally than others.

Prediction layer. A recommender does not need to “understand” a person in the human sense. It estimates which candidate item is likely to produce a chosen outcome, given available signals. That outcome might be a click, dwell time, a completed video, a reply, a survey response, or a multi-objective blend. The metric is a proxy: a measurable stand-in for something the platform values. Proxy choice matters because systems can become very good at the measured target while missing the broader goal.

Interface layer. Ranking determines what appears; interaction design determines how the next choice arrives. Autoplay reduces the effort between episodes. A continuous feed removes a natural stopping boundary. Notifications move the platform’s cue into another context. Visible endpoints, “you’re all caught up” notices, batching, and quiet defaults can restore moments for reconsideration. These elements are not automatically manipulative. Their meaning depends on transparency, symmetry, user expectations, age, vulnerability, and whether declining is genuinely easy.

Human and social layer. People are not empty vessels. They bring purposes, habits, relationships, moods, literacies, constraints, and values. Their networks shape what is available before a ranking system acts. A group chat can maintain care; a creator can find an audience; a recommendation can save search effort. The same feature can support one person and exhaust another. A serious analysis therefore asks not only “How long did someone stay?” but “What did the session enable, displace, or make easier to choose?”

The metric gap

Engagement is observable; benefit is harder to observe. A long session might mean delight, deep learning, confusion, conflict, or inability to find an exit. A short session might mean failure—or perfect service. This is the metric gap. When an organization optimizes a convenient signal, the signal can gradually become the definition of success. The central governance question is not whether metrics should exist. It is who chooses them, what counter-metrics constrain them, and whether people can understand or alter the resulting choice environment.

Mechanism map / 03

The feed is a loop, not a slot machine

The slot-machine metaphor captures variable rewards but hides the organizational decisions upstream and the social choices downstream. A loop is more useful: every interaction supplies new data that can tune the next prediction.

01
Define value
An organization selects objectives and guardrails: revenue, retention, satisfaction, safety, relevance, or a weighted blend.
02
Generate candidates
Eligible posts or videos are assembled from subscriptions, networks, catalogs, and retrieval systems.
03
Predict response
Models estimate likely outcomes from signals. The available signals and target encode consequential choices.
04
Arrange the path
Ranking and interface turn predictions into an order, a default, a cue, and a moment—or absence—of stopping.
05
Learn from traces
Clicks, dwell, skips, hides, surveys, and reports become feedback. What is not measured can disappear from view.

Documented cases / 04

When the target changes, the system changes

These cases use public technical papers, platform statements, and independent scholarship. Platform accounts document stated designs and policies; they do not independently prove their total effect.

CASE
01

YouTube recommendations · watch time, satisfaction, and contested “rabbit holes”

A ranking objective is never just a line of code.

A 2016 paper by YouTube engineers described a large-scale recommendation architecture with separate candidate-generation and ranking stages. The ranking model used expected watch time rather than click probability alone: a click on a video that was quickly abandoned should count differently from sustained viewing. This was an engineering response to clickbait, but it also illustrates the incentive question. When watch time becomes a target, content and interface choices that extend watching can gain system value.

In 2019, YouTube publicly said recommendations had expanded beyond views to include satisfaction signals such as likes, dislikes, and surveys, and announced reduced recommendations of “borderline” content. That statement documents a change in the platform’s declared objective and moderation strategy. It does not, by itself, establish the system’s effects across every topic, country, or user.

Structural concern

Optimizing predicted watch time may reward material that keeps attention even when the experience is not beneficial. Personalization can make exposure paths difficult for outsiders to audit.

Counter-evidence

Ledwich and Zaitsev’s 2020 snapshot of nearly 800 political channels found recommendation traffic moving toward mainstream sources rather than toward the most extreme categories. Their result disputes a universal radicalization “rabbit hole,” while leaving questions about personalization, other domains, and changing systems.

Reading rule Do not infer a universal outcome from an objective, and do not infer safety from one audit. Ask which version of the system, which population, which content domain, and which outcome was measured.

CASE
02

Facebook News Feed · “meaningful interactions” and the agency debate

Replacing one proxy means choosing another.

In January 2018, Facebook announced that News Feed would prioritize posts predicted to spark conversations and “meaningful interactions,” while showing less public content. The company described comments, shares, and likely back-and-forth discussion as ranking signals. This made a value judgment explicit: interaction between people would be treated as more desirable than passive consumption. Yet “meaningful” remained a proxy inferred from observable actions, not a direct reading of relationship quality.

A related question is how much ranking itself determines what people encounter. A 2015 Science study of 10.1 million U.S. Facebook users reported that individual choices played a stronger role than algorithmic ranking in limiting exposure to ideologically cross-cutting news. The study was observational, restricted to self-identified liberal and conservative users, and conducted by Facebook employees—important context when interpreting it.

Four years later, a randomized experiment published in the American Economic Review found that deactivating Facebook before the 2018 U.S. midterms produced a mixture of outcomes: more offline activity and improved reported well-being, but less factual news knowledge. Participants also valued access to the service. The lesson is not that Facebook is “good” or “bad.” It is that welfare is plural, and removing a platform can create benefits and losses at once.

Reading rule Platform power and user agency are not opposites. Networks, ranking, interface, and deliberate choice can all contribute to the same observed feed.

Evidence ledger / 05

Evidence, interpretation, hypothesis, myth

Strong critique needs calibrated language. The labels below separate what research directly supports from what remains an interpretation or a design hypothesis.

StatusClaimWhat the evidence permits

EvidenceNotifications can divert attention even without a response.A controlled study found performance costs from receiving phone notifications. Generalizing the exact effect to all contexts would exceed the study.

EvidenceBrief smartphone checking can become habitual.Oulasvirta and colleagues combined usage data and diary work to describe short, repetitive checking habits. Habit is not the same as addiction.

InterpretationAd-funded platforms face stronger pressure to extend measurable engagement.This follows from the availability of ad inventory and behavioral signals, but company strategy, ad load, quality goals, and regulation can constrain it.

HypothesisMore stopping cues will reduce momentum use in this simulator.The direction is plausible and testable. The model’s coefficient is invented for learning and must not be presented as an estimated effect.

MythAlgorithms hypnotize passive users into whatever platforms want.This erases human purpose, social networks, competing platform goals, and large differences in context. Influence is real; total control is not established.

The structural perspective

Choice architecture is not neutral. Platforms control defaults, ranking objectives, data collection, notification schedules, and the ease of exit. These systems operate at scales that make small average effects consequential. Power is asymmetric because the platform can observe and experiment across a population while an individual sees only one interface.

The agency-and-benefit perspective

People actively select, resist, repurpose, and leave platforms. Digital services reduce search costs, sustain relationships, enable creative work, and offer entertainment people value. Time spent is not automatically time harmed. Interventions that merely reduce minutes can also remove information, community, or pleasure.

A+B

A better synthesis

Evaluate the quality of agency: Are objectives legible? Are alternatives real? Is refusal easy and durable? Are benefits and burdens distributed fairly? Can independent researchers inspect outcomes? The aim is not minimal attention; it is attention allocated under conditions people can understand and contest.

Design and governance / 06

Design for a dignified exit

An ethical platform does more than offer a settings page. It makes the preferred action understandable, gives alternatives equal visual dignity, and preserves a person’s decision over time.

Name the objective

Explain the main parameters that shape ranking and why an item appears. The EU Digital Services Act makes recommender transparency a legal obligation for covered platforms; explanation quality still matters.

Offer a real alternative

Provide chronology, subscription-only views, or other non-profiled options where appropriate. Make the choice reachable at the point of use, not hidden several levels deep.

Restore stopping points

Use episode boundaries, feed endpoints, session summaries, or gentle time cues. A stopping point should invite reflection without shaming the person for staying.

Make refusal durable

“No” should not mean “ask again tomorrow.” Notification, personalization, and data-sharing choices should persist and be as easy to reverse as they were to accept.

Measure counter-outcomes

Pair engagement with hides, regrets, survey-based satisfaction, successful task completion, diversity, safety, and downstream well-being measures—while respecting privacy.

Audit distributions

An average can hide who bears the cost. Test effects across age, disability, language, socioeconomic context, and vulnerability, with meaningful external scrutiny.

Guided reflection / 07

Run a seven-minute attention audit

Choose one recent session. Do not judge yourself. Reconstruct the environment.

01 · ARRIVALWhat opened the session?
A deliberate task, boredom, a notification, another person, or an automatic check?
02 · CONTRACTWhat did you intend to do?
Write a verb: reply, learn, publish, rest, coordinate, laugh, or search.
03 · ROUTEWhich feature shaped the next action?
Ranking, autoplay, a badge, a recommendation, a streak, a visible endpoint, or a social obligation?
04 · LEDGERWhat value was created—and displaced?
Name both. Ten enjoyable minutes may be worthwhile; ten interrupted minutes may not be.
05 · REDESIGNChange one condition, not your character.
Batch a cue, add an endpoint, move the icon, use chronology, or state an intention before entry.

Then ask: If the platform were paid for my stated outcome rather than my continued presence, what would it change first?

Questions / 08

Six useful questions, without easy villains

Is attention literally a commodity?

Not in the same way as oil or wheat. Platforms do not own a standardized unit called “attention.” The economic metaphor is useful because access to audiences can be priced, and attention-related behaviours can create inventory and data. But attention is embodied, contextual, and selective. Minutes are only a rough proxy: the same duration can contain absorption, distraction, connection, confusion, or rest.

Does an ad-funded model always produce harmful design?

No. Advertising can subsidize broad access, and ad-funded services can adopt strong limits, quality goals, and user controls. The concern is directional: if more measurable exposure can produce more revenue, organizations need counterweights so continued use does not silently become the dominant definition of value. Governance, culture, market competition, regulation, and product strategy all mediate the result.

Are infinite scroll and autoplay “addictive”?

That word often collapses clinical, colloquial, and design meanings. These features reduce stopping friction and can extend a session, but the size and significance of that effect varies. A responsible critique describes the mechanism—automatic continuation or absent endpoints—without diagnosing the user. Clinical claims require clinical evidence; the simulator makes none.

Can engagement ever be a good metric?

Yes. Replies can indicate conversation; completion can indicate relevance; sustained attention can accompany learning or delight. The problem is construct validity: does the metric capture the outcome people actually value, across contexts and groups? A robust system uses multiple measures, checks for gaming and side effects, and allows qualitative feedback to challenge what the dashboard suggests.

Would chronological feeds solve the problem?

Chronology reduces one layer of predictive ranking, but it does not remove social selection, posting incentives, advertising, interface design, or unequal visibility. It can also overwhelm people when volume is high. Chronology is valuable as a legible option, not a universal cure. The relevant question is whether people can choose and understand the trade-offs.

Who is responsible: the user, designer, company, or regulator?

Responsibility is distributed but not equal. Individuals can shape habits; designers can surface alternatives; executives choose metrics and resources; advertisers and creators respond to incentives; regulators define minimum rights; researchers and civil society provide scrutiny. The party with greater information, experimental capacity, and control over defaults carries a greater duty not to exploit that asymmetry.

Source desk / 09

Spend the next minute on purpose.

Take the printable framework, question the simulator’s assumptions, or write with a platform design you want to examine.

Get the Field Guide Write to Sahil

Email fallback: sahil@sahiltayal.com