For most of modern media history, information scarcity created editors. A newspaper had a finite number of pages. A broadcast had a finite number of minutes. Someone had to choose.
Digital media reversed the constraint. The problem is no longer scarcity. It is abundance. There is far more information available than any person can consume, so selection did not disappear.
It became automated.
An editor that does not look like one
Recommendation systems perform a function once associated with editors: they determine what receives attention. But they do it differently. There is not one front page. There may be hundreds of millions of them.
TikTok describes its For You feed as a personalized ranking system. Signals can include what a person watches, likes, shares and comments on, the accounts they follow, and information about the videos themselves. Some behaviors are weighted more heavily than others.
The important word is ranking. The system has a large universe of possible content. It predicts. Then it orders. That ordering determines experience.
It does not have to understand the argument
When a rebuttal answers an accusation, the rebuttal does not automatically confront the accusation. It becomes another piece of content. The system tests it separately.
That is deliberately simplified. Real recommendation systems optimize multiple objectives and operate under moderation rules, safety policies and other constraints, and should not be described as single-minded machines maximizing watch time.
But the underlying distinction is real. A recommender does not need to determine which side of an argument is correct in order to predict which item a user is more likely to engage with. Truth evaluation asks whether a proposition corresponds to reality. Ranking asks which item should appear next.
A system can become exceptionally good at the second without solving the first.
Selection is power
Most information never reaches most people. A platform does not need to censor an idea to make another idea more visible. It can rank one higher, recommend another repeatedly, introduce one to new audiences, or stop recommending another.
None of this requires an ideological intention on the part of the machine. It requires criteria. Every ranking system must have some basis for ranking, and the moment one item appears above another, the system has expressed a preference in operational terms.
The objective comes first
The Hamlet Evaluation System provides the bridge. It could perform its calculations correctly and still produce a misleading picture if the categories, weights or aggregation rules failed to represent the reality commanders needed to understand.
Recommendation systems inherit a related problem. Before the system can optimize anything, somebody has to decide what counts as success: watch time, completion, return visits, shares, comments, satisfaction, discovery, revenue, safety, or some combination. The mathematics can be extremely sophisticated. The objective remains a human choice.
They taught it what winning meant.
Engagement is not neutral
A 2026 Nature registered report randomly assigned participants to different feed-ranking systems around the 2024 U.S. election. Compared with reverse-chronological feeds, engagement-based ranking increased exposure to intergroup, moralized, emotional and toxic content. The largest increases appeared in moral outrage and political material.
This does not establish that every commercial recommendation algorithm behaves the same way. It demonstrates something more fundamental: the ranking rule changes the environment. Different arithmetic produces different attention.
The audience is not passive
It is tempting to describe algorithms as machines manipulating helpless users. That is too simple. Users train the system through their own behavior. They click, pause, skip, search, follow, share, argue and return.
- The system selects what the person sees.
- The person reacts to what the system selected.
- The reaction becomes evidence for the next selection.
Neither side operates independently anymore. The audience helps edit itself.
Correction enters the same machinery
A false claim can reach one million people. A correction appears. In personalized feeds, the people who receive the correction are not automatically the people who received the original claim. The correction has to earn distribution too. It needs its own signals, its own audience and its own momentum.
That is why the Estrada episode later in the novel matters. The original false identification receives 912,000 retweets. The correction receives 48,000. The underlying factual issue has been resolved. The distribution problem has not.
Correction does not assign itself.
There is no single editor anymore
The old information environment had identifiable centers of editorial power: publishers, networks, newsrooms, governments, campaigns. Those institutions had interests and intentions.
The recommendation environment is harder, because intention has been partially displaced by optimization. An information environment can develop systematic tendencies without requiring a person to sit in a room and choose every outcome. Millions of local ranking decisions accumulate, and the aggregate becomes culture.
Behavior before belief
Traditional persuasion often starts with beliefs. Recommendation systems have access to something more immediate. You stopped here. You replayed this. You skipped that. You watched this creator twice. You searched for this phrase.
The system does not necessarily need a philosophical explanation for why. It can learn that one behavior predicts another. That moves us beyond Simulmatics, which placed people into 480 fixed types. A modern category can be temporary, contextual, probabilistic, and perhaps invisible even to the person being categorized.
From prediction to environment
The three pages connect cleanly.
- Simulmatics: can we model the person well enough to predict behavior?
- Hamlet Evaluation System: can we measure the environment well enough to know what is happening?
- Recommendation systems: can we continuously alter the environment based on predicted behavior?
The final system closes the loop. Measurement informs selection. Selection changes behavior. Behavior produces new measurements. Those measurements inform the next selection. The model is no longer standing outside reality describing it. It is participating in the reality it measures.
Selection does not have to resemble censorship to shape reality. It only has to determine what rises, what falls, and what appears next. The editor did not disappear. The editor became arithmetic.