The three jobs AI does in social media

"AI on social media" gets used for three jobs that share almost nothing. Different models, different builders, different ways of being wrong. Keeping them apart is most of the battle.

1. Distribution. Deciding which posts you see, in what order. This is a recommender: a ranking system.
2. Creation. Helping make the post: drafting a caption, generating an image, cutting a clip, translating a voice-over. This is generative AI.
3. Moderation. Deciding what comes down, gets labelled, or gets limited. This is classification at scale.

When somebody says "the AI is pushing this", they mean job one. When they say "this is AI slop", job two. When they say "the AI banned me", job three. They are unrelated systems that happen to share a label.

The reason to separate them: they fail differently, and a fix for one does nothing for the others. Confusing them produces most of the bad arguments about this subject.

The oldest and by far the most consequential use of AI in social media is ranking. It predates the current wave of generative models by more than a decade, and it is the part that actually shapes what a platform is.

The shape is simple to state. A system predicts, for every candidate post, how likely you are to engage with it. Then it sorts by that prediction. Everything else is detail.

What is worth internalising now: this system does not understand quality, truth, or whether something is good for you. It predicts a behavior, because a behavior is the only thing it can measure. Every characteristic people attribute to "the algorithm" follows from which behavior it was told to predict.

This job has its own path on AnyLearn, How social media algorithms decide what you see, which covers the funnel, the objectives, and cold start properly. The rest of this path is about jobs two and three.

The newest job is generation: models that produce text, images, audio, and video rather than sorting it.

In practice this shows up as caption and hook drafting, image generation, background removal and editing, auto-cutting long footage into clips, subtitling and translation, and voice cloning for dubbing.

The economics are what matter. Each of these used to cost either money or hours. Subtitling a video was a task; now it is a button. When the cost of producing something falls to near zero, the volume rises and the average quality falls, because the constraint that used to filter out low-effort work has gone.

That is the mechanism behind the complaint about "AI slop". It is not that the models are incapable. It is that removing a cost floor changes what gets made and posted. Anyone can now produce a plausible-looking post about anything, whether or not they know anything.

Moderation is the least visible job and the one with the hardest constraints. Platforms receive volumes of content no human review could touch: the arithmetic simply does not permit a person to look at each item.

So the first pass is always automated. Classifiers score each item for categories like spam, nudity, violence, and self-harm. Most content is cleared automatically. A small fraction is removed automatically. A middle band is escalated to human reviewers.

The defining constraint is that these classifiers are wrong at scale in both directions, and the two errors trade off against each other:

• False positive: something fine gets removed. A creator loses a post and trusts you less.
• False negative: something harmful stays up. A person gets hurt and you are in the news.

You cannot minimise both. Tightening the threshold to catch more harm inevitably removes more innocent content. Where a platform sets that dial is a policy decision wearing an engineering costume, and no model choice makes the tradeoff disappear.

Because the jobs differ, so does what failure looks like and who notices.

Read across the bottom row. Distribution failures are diffuse: the feed feels off, nobody can point at one thing. Creation failures are obvious the moment you look. Moderation failures are invisible to everyone except the one person they landed on, which is exactly why they generate so much anger and so little agreement.

Same word, three different problems.

The jobs are separate systems, but they are pointed at each other, and that is where it gets interesting.

Creation feeds distribution. Generation makes content cheap, so far more of it enters the pool competing for the same fixed number of slots. Ranking does not care that an item was machine-made; it predicts engagement either way. A generated post that people watch will be promoted.

Moderation polices creation. As generation improves, moderation is handed a job it was never designed for: deciding whether something is real. That is a much harder question than whether something is spam.

Distribution shapes creation. Creators observe what gets reach and make more of it. The objective does not just sort content, it teaches people what to produce.

The honest summary: three systems in a loop, each responding to the others' output, with no one component in charge of the result.

The payoff of separating the jobs is that vague claims become checkable questions.

"The AI is suppressing my posts." Which job? If distribution, the question is what the ranker predicts and whether your item cleared its test audience. If moderation, the question is whether a classifier labelled you, which is a different system with an appeals path.

"AI is ruining social media." Which job? Generation raising volume and lowering the average is a real, specific claim. Ranking optimising a bad proxy is a different real, specific claim. They need different fixes.

"Just make the AI better." Better at what? A more accurate ranker optimising for clicks gives you more effective clickbait. Capability is not the lever; the objective is.

That is the whole point of a map. Not to settle the arguments, but to make them precise enough to be settleable. The next two lessons take jobs two and three and get concrete.

Distribution, creation, and moderation are separate systems with separate failure modes, but each one's output becomes another's input.