Curiosity and preference are not the same thing.
Recommendation systems have to infer preferences from behavior, and behavior is ambiguous. A person watches a video about a conspiracy theory. Are they a believer, a critic, a journalist, a bored insomniac, or somebody who clicked the wrong thumbnail?
The system initially sees something much simpler: this account watched this thing.
YouTube’s current recommendation documentation says watch history and search history are major signals used to personalize future recommendations. Its own help page on managing recommendations and search results gives an unusually revealing example: users can turn off watch and search history when researching a subject for a school project that they are not personally interested in.
That advice exists because exploratory activity can otherwise influence what appears later.
One click can create the next opportunity
Suppose a user normally watches woodworking videos and clicks one documentary about an obscure cult.
The platform may now have evidence—weak evidence, but evidence—that this subject held the user’s attention. Another related video becomes slightly more plausible. If the user clicks that one too, the signal strengthens. Soon the feed may contain enough related material that the subject appears to have become a large part of the platform itself.
The loop is straightforward:
exposure -> click -> inferred interest -> more exposure -> more chances to click.
That does not mean every exploratory click causes a dramatic recommendation spiral. Modern systems use many signals, including dislikes, “Not interested” feedback, satisfaction surveys, subscriptions, search history, and repeated behavior. YouTube also lets users delete individual watched videos so they no longer shape future recommendations through watch history. See its watch-history controls.
Curiosity is hard to infer from telemetry
A recommendation system cannot directly observe the sentence in a person’s head:
I want to understand this once, but please do not turn my homepage into this subject.
It has to infer intent from actions.
A stronger model of lasting interest therefore needs more than a single event. Repeated voluntary selections, subscriptions, likes, searches, long-term viewing patterns, explicit topic choices, and negative feedback all provide additional context.
Users can test the effect themselves without assuming conspiracy. Record a recommendation surface, explore a narrow unfamiliar topic, then compare how the surface changes. Remove the relevant history or use a separate profile and compare again.
The Dead Internet Theory angle is not that the recommended material is fake.
It is that a personalized surface can rapidly become self-reinforcing.
After enough rounds, the feed may look like a spontaneous map of what the internet contains when it is partly a map of what the system thinks one experimental click meant.