Why Does Online Consensus So Easily Look Like Knowledge?

Geoffrey Chen

How a Person Builds a World · Season Two: “Others, Dialogue, and Shared Knowledge” · Article Eleven

Open a topic and the first ten items express similar positions. Highly rated comments confirm one another. A screenshot announces that “everyone finally understands”. Within minutes, it is easy to form two judgements: most people really do think this, and their agreement shows that the matter is probably settled.

What the user has actually observed is what a platform displayed to this account at this moment.

Online consensus looks like knowledge for more reasons than a generic human tendency to conform. Digital platforms compress “How many people believe this?”, “Which content became visible?” and “Which expressions received interaction?” into the same interface signals. Likes, ranking, reposts and repeated appearance simultaneously suggest visibility, group approval and algorithmic assessment. A user cannot readily tell whether the display represents a social distribution or a surface produced by distribution machinery.

Seeing a majority does not establish that a majority exists

To judge that “everyone thinks this”, we would need to know the population, how the sample appeared and where silent people are. Social platforms usually provide none of these conditions.

Following relations are not random samples. People select accounts with related interests and views. Active users produce more material than silent ones. A platform predicts which information is likely to retain a user's attention based partly on earlier behaviour. What a person sees is therefore a sample filtered repeatedly through self-selection, network position and recommendation.

Network structure itself can produce a majority illusion. Kristina Lerman, Xiaoran Yan and Xin-Zeng Wu showed that even when an attribute is globally rare, uneven connectivity can make it appear common in many people's local neighbourhoods. Their open-access study of the majority illusion in social networks used synthetic and real networks to analyse this possibility. It does not imply that every apparent online majority is false. It shows that locally visible prevalence can systematically differ from global prevalence.

Highly connected accounts appear in many local environments. A relatively small active group occupying central network positions can give many users the independent experience that “most people around me think this”. Each local observation can be real while the shared inference about the whole is wrong.

A like records evaluation and also changes evaluation

If ratings merely aggregated independent judgements, a larger number of ratings could ordinarily provide more information. Online ratings also influence the people who arrive later.

Lev Muchnik, Sinan Aral and Sean Taylor conducted a randomised experiment in a large online community. They applied small manipulations to the initial evaluation of comments and observed later users. Prior ratings causally influenced subsequent rating behaviour, with asymmetric herding effects. Their paper, “Social Influence Bias: A Randomized Experiment”, shows that the final number is not only a sum of independent opinions. It contains the influence of earlier numbers on later judgement.

Consensus signals can therefore become circular. Content becomes more visible because of an early advantage. Visibility produces more interaction. Interaction yields further distribution. In the end, a high rating may reflect quality, publication time, initial chance, account position and platform mechanisms together.

“Many people approved” does not entail “Many people independently inspected the evidence and reached the same conclusion.” Epistemically valuable consensus depends on some independence of judgement. Ten thousand repetitions of one unverified report do not automatically become ten thousand pieces of evidence.

Repetition makes familiarity resemble confirmation

The same claim can recur in a feed as a news headline, reposted commentary, a short video and an AI summary. Each expression differs slightly and appears to provide another source. Tracing them may reveal that all derive from one report or one study.

People cannot easily distinguish repeated sources from repeated content by feel alone. Multiple accounts making the same statement are experienced as social corroboration. Platforms rarely display the genealogy: who reported first, who merely repeated, and which accounts cite one another.

This is where lateral reading becomes valuable. Professional fact-checkers do not evaluate a page solely from its internal appearance. They leave it to investigate the source's identity, other coverage and original material. The research on lateral reading suggests applying a similar question to apparent consensus: how many independent evidential paths exist beneath these ten visible sources?

Familiarity is not worthless. A claim independently observed across different settings can accumulate genuine support. The difficulty is that an interface shows repetition without showing independence.

When consensus genuinely deserves trust

The possibility of manufactured consensus does not entail that consensus is generally untrustworthy. In science, medicine and professional practice, convergence among relevant experts after different investigations, public criticism and review provides powerful second-order evidence for non-experts.

Epistemically valuable consensus usually has structural conditions. The participants are relevant to the question. Their judgements do not all copy a single source. Methods and evidence can be inspected. Dissent has a channel of expression. Conclusions can change with new evidence. The scope of agreement is accurately stated.

Online popularity does not automatically meet these conditions. It tells us that an expression is receiving visibility and response. It need not tell us how relevant experts judge, still less how an entire population is distributed.

Nor does minority status make a view automatically true. “Everyone has been deceived and only I see the truth” is another story capable of exploiting identity. Majorities require structural inspection; minorities require evidence too. Epistemic judgement cannot consist of oscillation between conformity and anti-conformity.

AI compresses online visibility into “It is widely believed”

When asked about a disputed subject, generative AI often organises an answer around phrases such as “It is widely believed”, “Many experts point out” and “The mainstream view is”. Sometimes these descriptions are accurate. Sometimes they compress frequently encountered language in training material or search results.

Text encountered by a model is not a population survey. Material that is digitised, published, searchable and included has already been selected. Highly visible sources become more likely to appear in later texts. If an AI does not show its source set and selection process, the user cannot easily tell whether “widely” refers to a professional consensus, a media convention, a dominant view in the English-language internet or simply a phrase used to make the answer sound balanced.

An AI summary of consensus therefore warrants follow-up. Which group is meant? Is there a survey or formal consensus statement? Do relevant experts and the public differ? What is the distribution of disagreement? How current is the material? Are the cited sources independent?

The most useful role for AI is not to announce consensus but to map its structure: core agreement, principal disagreements, sources of evidence, scope and unresolved questions. If it cannot provide those, “widely believed” should remain an unverified description.

Turn the feeling of consensus back into inspectable claims

When “everyone online says so”, the phrase can be unpacked:

  • Which platform, language and period does “online” describe?
  • Does “everyone” mean all users, active posters, accounts I follow or accounts recommended to me?
  • Does the visible number count people, interactions or a ranking output?
  • Do different expressions derive from independent sources?
  • How are people who did not speak represented?
  • Does the consensus concern a fact, a value or a signal of group identity?

These questions do not give an ordinary user complete population data. They prevent an interface feeling from becoming knowledge prematurely.

Online consensus is first a fact about an information environment: within a particular distribution structure, this expression is prominent. Only after sampling, independence, relevant expertise and formation process are known can it become stronger evidence about the world.

A person's world is inevitably influenced by shared opinion. Without social feedback, we cannot calibrate which questions deserve attention or benefit from checks other people have already performed. But when a platform chooses which “other people” become visible, a psychological interval must remain. The visible neighbourhood is not the whole society. Repeated language is not independent evidence. Popularity is not a property of a fact.

The consensus worth relying on is not a picture without disagreement. It is a process that lets disagreement enter, displays the sources of evidence and changes after error. The internet can give us the picture with extraordinary speed. A knowledge life must continue to look for the process.

Primary sources and further reading

Continue reading: Explore the How a Person Builds a World series.

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.