You pause over an outrageous post for twelve seconds.
Not because you like it.
Because you cannot believe a person assembled those words, inspected them and still pressed Post.
The platform records twelve seconds.
It does not record disbelief.
Tomorrow, it brings you more.
From the system’s point of view, your moral horror has excellent retention.
The system sees traces, not reasons
Recommendation systems learn from behaviour:
- what we click;
- what we finish;
- what we replay;
- what we hide;
- what we share;
- whom we follow;
- and how long we pause.
Platforms can use many signals and objectives, not one crude “engagement” counter. Meta, for example, describes ranking systems that predict several possible actions and combine those predictions when selecting content. (Meta, 2023)
But behaviour remains ambiguous.
A long pause can mean delight, confusion, attraction, anger or that the kettle started screaming in another room.
The system must infer a preference from the trace.
Then it acts on the inference.
The loop
The basic relationship looks like this:
attention → inferred interest → more exposure → changed attention → new inference
Working model—not a claim of direct control
You watch one video about running shoes.
The platform shows more running content.
Running becomes easier to think about. You learn the vocabulary, recognize the creators and perhaps begin running.
Your new behaviour confirms the original inference.
Was the platform correct about a preference you already had?
Did it cultivate a weak interest into a strong one?
Or did it merely put a pair of shoes in front of a person who needed a hobby?
The loop makes these explanations difficult to separate.
Researchers call one technical version algorithmic confounding: recommender systems learn from data partly produced by earlier recommendations. In simulations, this can homogenize behaviour without increasing user utility. Simulations are not proof that every social platform produces the same outcome, but they reveal the structural problem. (Chaney, Stewart & Engelhardt, 2018)
The observer is inside the experiment.
So is the algorithm.
Repetition changes the available world
The platform does not need to convince you that an event is common.
It can show you the event repeatedly.
What comes easily to mind begins to feel important. A rare crime, unusual relationship disaster or fringe opinion can occupy the psychological space of a national census.
This is not unique to algorithms. News editors, neighbours and family stories have always selected reality.
The difference is speed, personalization and feedback.
My feed can learn which version of the world keeps my thumb moving, then test another version minutes later.
The result is not necessarily a false belief.
It may be a badly sampled reality.
Creators train too
The loop includes more than viewers.
Creators learn which openings retain attention, which emotions travel and which version of themselves receives distribution.
A thoughtful qualification performs poorly.
A confident declaration performs well.
The next post loses the qualification.
Millions of creators make similar local adjustments. The platform then learns from content already adapted to the platform.
The algorithm trains creators.
Creators train audiences.
Audience behaviour trains the algorithm.
Nobody needs a master plan. Reciprocal adaptation is enough.
Users are not passive
The helpless-user story is tempting and incomplete.
People choose whom to follow, what to search for, when to leave and what to create. Recommendation systems can reveal rare expertise, small communities and art we would never find through our existing social circle.
Algorithmic selection is not automatically manipulation.
Nor is a chronological feed a neutral window onto reality. It privileges recency and the posting frequency of accounts we already chose.
One large field experiment during the 2020 US election changed Facebook and Instagram users from algorithmic ranking to reverse-chronological feeds. The change substantially altered what people saw and reduced engagement, but did not produce detectable changes in measured political attitudes or behaviours over the three-month study. (Guess et al., 2023)
That is important pushback.
Changing the feed did not simply reprogram the person.
Short-term platform exposure is one influence among friendships, institutions, prior beliefs, other media and the irritating human ability to ignore what researchers expected us to notice.
The real concern is legibility
I do not need to believe that an algorithm controls me to want the loop to be more visible.
Three questions are currently entangled:
- What did I actively choose?
- What did the system infer from my behaviour?
- What did repeated exposure teach me to choose next?
A useful platform might help separate them.
Why am I seeing this?
Which action was interpreted as interest?
Can I say “I want to understand this, but I do not want more of it”?
Can I ask for novelty rather than similarity?
Can I inspect the themes the system believes define me?
The ability to reset or tune recommendations helps, but it still asks users to manage a model they cannot fully see.
A small personal experiment
For one week, before interacting with a recommended post, write down:
- why it caught your attention;
- whether you want more of it;
- and what action the platform is likely to observe.
“I watched the whole video because the explanation was excellent” and “I watched the whole video because the advice was medically alarming” may look identical in a log.
The exercise will not reveal the algorithm.
It may reveal the gap between preference and engagement.
That gap is where much of the reciprocal training happens.
What I still cannot figure out
Preferences do not arrive fully formed from somewhere outside the world.
Books, friends, schools, cities and families shape what we learn to want. Social media is another environment—just unusually responsive to our smallest behaviours.
If an algorithm learns what keeps our attention, how long before we begin wanting what it has learned to show us?
Until the next strange question,
Osagie
My current hunch: Social platforms do not simply discover preferences; through repeated, personalized exposure they participate in forming them.
Most likely reason I am wrong: Users’ durable interests and off-platform lives may constrain this loop far more strongly than feed-level behaviour suggests.