I have talked about “the Instagram algorithm” for years.
The algorithm liked this post.
The algorithm killed that one.
The algorithm isn’t showing my content.
The algorithm wants Reels.
The algorithm doesn’t like this kind of photograph anymore.
I used the word constantly.
Then I realized something slightly embarrassing.
I wasn’t entirely sure what I meant by it.
So I started asking a very basic question: What actually is an algorithm?
At its simplest, an algorithm is just a procedure for making a decision. Give a computer information, apply a set of processes to that information, and produce an answer.
Instagram is obviously far more complicated than a simple set of rules. Its recommendation systems use artificial intelligence and machine-learning models to make enormous numbers of predictions about what people are likely to do.
And once I understood that, something about Instagram suddenly became much clearer to me.
Instagram doesn't know my photograph is beautiful
That sentence sounds almost ridiculous when I write it.
Of course a computer doesn’t look at my photograph the same way another person does.
But I think, somewhere along the way, I had unconsciously started treating Instagram as if it did.
I would create an image I absolutely loved—the lighting, the location, the expression, the clothes, the composition—and then watch it perform poorly.
Meanwhile, another photograph I considered much more ordinary would take off.
And I would wonder:
How can Instagram not see that this one is better?
The answer is surprisingly simple.
It can’t.
At least not in the way I mean when I call something beautiful.
Instagram’s AI can understand a tremendous amount about an image. Modern machine-learning systems can recognize subjects, similarities, visual characteristics and relationships between different kinds of content.
But recognizing what is in a photograph is very different from experiencing it.
The system doesn’t look at a mountain sunset and feel awe.
It doesn’t see tenderness in an expression.
It doesn’t understand why a photograph reminds someone of being twenty years old, falling in love, losing someone, going home, growing older or wanting desperately to be somewhere else.
It doesn’t admire an outfit.
It doesn’t appreciate the hours I spent getting an image exactly the way I wanted it.
It doesn’t experience beauty.
What it can do extraordinarily well is predict behavior.
Prediction is not understanding
Meta has publicly described Instagram ranking systems that predict things such as the probability that someone will like, comment, follow or otherwise interact with a piece of content.
That distinction has become incredibly important to me.
Instagram doesn’t necessarily need to understand why you like my photograph.
It only needs to become good at predicting that you probably will.
Imagine Instagram has learned that a particular person frequently engages with photographs containing some combination of fashion, outdoor scenery, blonde women, mountain settings and certain types of creators.
I post something matching many of those characteristics.
Instagram doesn’t have to think:
“This is a beautiful portrait of Rissa.”
It can effectively think:
“Based on everything I have learned, this person has a high probability of responding positively to this.”
And if that person does?
The prediction was successful.
That is an impressive technological achievement.
But it is not the same thing as understanding a human being.
And then I noticed the loop
This is where it became much more interesting to me.
Machine-learning systems learn from previous behavior.
If one type of my content consistently performs well, Instagram gains evidence that this kind of content works for certain people.
So Instagram has more confidence recommending similar content.
People respond.
That produces more evidence.
Instagram becomes even more confident.
And the cycle continues.
At first that sounds completely reasonable.
If something works, do more of it.
But there is a problem hidden inside that logic.
What happens to the things the system doesn’t already believe will work?
Suppose Instagram predicts that Content A will perform well and Content B probably will not.
It distributes A heavily.
A performs well.
It cautiously distributes B.
B performs poorly.
The system can now conclude:
“My prediction was correct.”
Except A and B were never really given the same opportunity.
The prediction helped create the result that later appeared to confirm the prediction.
That fascinates me.
And troubles me a little.
The algorithm may be measuring a world it helped create
This is where recommendation systems stop looking completely neutral to me.
Instagram isn’t merely standing on the sidelines watching what people like.
It decides what people get the opportunity to see.
That matters enormously.
If the system repeatedly shows someone a certain type of my content because that person previously responded to it, that person will naturally generate even more behavioral evidence about that kind of content.
Meanwhile, something completely different that I create may never receive enough exposure for Instagram—or even me—to discover whether that same person would have loved it too.
Over time, this can become self-reinforcing.
A creator discovers what performs.
She makes more of it.
Instagram becomes better at finding people who respond to it.
Those people become a larger part of her audience.
That audience responds strongly when she produces more of it.
She learns that this is apparently what her audience wants.
Instagram learns the same thing.
Around and around we go.
Eventually everyone involved has an enormous amount of data proving that this is what works.
But we rarely get to see the alternate history.
What would have happened if different content had received the same opportunity?
This doesn't mean Instagram never experiments
To be fair, recommendation systems aren’t completely trapped by their own history.
They have to experiment.
If Instagram only showed us things its systems were already confident we would like, our feeds would eventually become painfully repetitive.
Meta itself has acknowledged this problem in some of its ranking systems. Pure engagement optimization can overprioritize creators or content types a person has engaged with before, so additional mechanisms can be used to introduce diversity.
In machine learning there is a useful way of describing this tension:
Exploitation versus exploration.
Exploitation means:
“I know this works. Keep doing it.”
Exploration means:
“I’m not certain about this, but let’s try it and see what happens.”
Every creator lives somewhere inside that tension.
The problem is that experimentation carries risk.
Instagram has a limited amount of your attention. Showing you something uncertain means not showing you something the system already believes has a high probability of keeping you engaged.
So there is always mathematical pressure toward the safe prediction.
Toward what worked yesterday.
That may explain why social media becomes repetitive
Once I understood this, something else began making sense.
Why do successful accounts so often become increasingly repetitive?
Why does a creator discover one particular style, subject, pose, format or personality and then seemingly produce variations of it forever?
We usually blame the creator.
Sometimes that’s deserved.
But the system is also sending a very strong message:
This worked.
So the creator repeats it.
It works again.
See?
Repeat.
Eventually creativity can turn into optimization.
Not necessarily because the creator stopped having ideas, but because every measurable signal is encouraging her to stay inside an increasingly well-defined box.
And the irony is that the box may partly consist of assumptions the recommendation system made about her in the first place.
A machine can be right without being wise
I don’t think Instagram’s recommendation systems are stupid.
Quite the opposite.
The scale of what they accomplish is extraordinary. They make billions of predictions in an environment involving enormous amounts of content and human behavior, and many of those predictions are remarkably accurate.
But accuracy and understanding are different things.
A system can correctly predict that I will click something without understanding why it mattered to me.
It can correctly predict that one of my photographs will outperform another without knowing which one means more to me.
It can identify patterns in human behavior without experiencing anything human.
That is the part I think I was missing whenever I casually complained about “the algorithm.”
I had subtly personified it.
I treated it like a critic.
It isn’t.
I treated it like an audience.
It isn’t.
I occasionally treated poor reach almost like a judgment on the creative value of what I had made.
It certainly isn’t that.
It is a collection of mathematical systems making very sophisticated guesses about human behavior.
And when those guesses are correct, they become evidence used to make the next guesses.
So what does Instagram actually know about me?
Probably a lot.
But perhaps less than I once imagined.
It can learn patterns associated with my content.
It can learn which people tend to interact with me.
It can learn what kinds of posts have historically generated particular responses.
It can become increasingly good at predicting who might respond to something I publish.
But that isn’t the same as knowing me.
It doesn’t know why I chose that mountain.
It doesn’t know why I loved that dress.
It doesn’t know that an image I created one afternoon meant more to me than one that received ten times the reach.
It doesn’t know when I’m proud of something.
It doesn’t know when I’m disappointed.
It doesn’t know when I’m experimenting, growing, remembering, grieving, flirting, laughing or simply trying something because creating it sounded fun.
It knows the measurable traces those things leave behind.
That is very different.
And perhaps understanding that distinction is healthy for anyone who creates on social media.
Because an algorithmic system can tell me a great deal about performance.
It can tell me what generated likes.
What generated comments.
What held attention.
What gained followers.
What it cannot tell me is whether something was worth creating.
That decision is still mine.
And maybe that is the part of being a creator that I should never have handed to an algorithm in the first place.
