AI Personalization: What Gets Measured, and What You Can Switch Off

Personalization is the part of the internet that adapts to you without announcing it. Two people can open the same app on the same morning and see almost nothing in common. This piece stays on two questions: what is being measured, and what European law obliges the largest platforms to offer people in the EU.

What gets measured

Two of the largest platforms publish an explanation of their own ranking, and reading them is more instructive than reading commentary about them. TikTok’s newsroom description and Instagram’s ranking explainer are both short, and both say something that surprises people: clicks and follows are in the mix, but they are not where the weight sits.

TikTok’s account is explicit about the principle. “A strong indicator of interest, such as whether a user finishes watching a longer video from beginning to end, would receive greater weight than a weak indicator, such as whether the video’s viewer and creator are both in the same country.” Instagram’s explainer names the five Feed interactions it looks at most closely, and the first is how likely you are to spend a few seconds on a post. For Reels the list is headed by resharing and by watching a reel all the way through. Completion and dwell time carry the load. They are also the two signals people are least aware of producing.

The same TikTok page is equally clear about what does not count for much, which is the part that gets left out of most summaries: device and account settings such as language, country and device type “receive lower weight in the recommendation system relative to other data points we measure since users don’t actively express these as preferences”. A good deal of folk wisdom about being tracked by device is aimed at a signal the platform itself describes as weak.

The remaining mechanism is older than any of these products. Systems also predict what you will do from what people with similar behavior did next, which is how something arrives in your feed that has no connection to anything you have watched. In the research literature this goes by collaborative filtering, and it has been standard in recommender systems for decades.

None of this requires knowing anything about you in the ordinary sense. A system does not need your job or your politics to predict which of two videos will hold you longer. What it models is behavior.

The question that matters

Personalization is usually described as a system learning what you want. That is not quite the arrangement. The system is trained toward a target somebody chose, and that target is a measurable proxy: time spent, sessions per week, items added to a basket, subscriptions retained. Those proxies overlap with your interests without being the same thing.

The gap between them explains most of the familiar complaints. A proxy for interest can be satisfied by something you find irritating enough to keep watching. It can also be satisfied by more of what you already engaged with, which narrows the range over time, because the system only ever observes your response to what it already chose to show you. That feedback loop is a structural property of the method, not a scandal, and everyone building these systems knows it is there.

So the question to put to any personalized surface is not whether it knows you. It is what it was told to maximize.

What the Digital Services Act requires

What follows is a summary of published rules, not legal advice. It also applies to people in the EU and to platforms the Commission has designated, which is a narrower set than “big websites”. If something specific matters to you, the regulation itself and a qualified lawyer are the right sources.

The Digital Services Act has applied since 17 February 2024, with some provisions from November 2022, and it reached the platforms designated as very large earlier still under its anticipated-application rule. Four of its provisions bear on personalization directly.

Article 38 is the one most people have not heard of, and it repays reading in the original, because the word doing the work is “profiling”. The GDPR defines that as automated processing of personal data to evaluate personal aspects of a person, such as their preferences, interests, behavior or location. An option that is not based on profiling can still be shaped by things you have stated openly, like the accounts you chose to follow. It is a specific removal, not a blank page.

Article 27 does the quieter job. It means the ranking parameters are written down in a document you can read, which is the paper trail behind any argument about how a given service ranks things.

Where to look

Doing this once on the service that takes most of your time beats reading about it.

1. The explanation attached to a single advertisement. For ads this is required: Article 26(1)(d) obliges platforms to make the main parameters used to select you accessible from the ad itself, and where applicable how to change them. Several services offer something similar on ordinary posts and videos, Instagram and TikTok among them, but that part is voluntary and coverage is uneven.
2. The advertising settings. Usually under advertising or ad preferences in your account settings, and this is where the inferred interest categories are listed. On Meta and Google properties the list can run to hundreds of entries. Elsewhere it may be coarse or absent.
3. The non-profiling option. On designated very large platforms it exists because Article 38 requires it. In practice services have implemented it as a chronological or non-personalized view, though the law asks for an option not based on profiling rather than for a chronological feed specifically. Where it sits, and whether it stays selected after you close the app, differs by service.

What switching it off does and does not do

A chronological feed is still a ranking. It favors whoever posts most, and whoever posts while you happen to be awake. Removing a profile removes one specific kind of shaping and leaves the surface sorted by something else.

What you gain is legibility. A chronological feed can be described in one sentence, and you can predict what it will do to what you see. Whether that is worth the loss of a feed tuned to your attention is a question only you can answer, and it is reasonable to answer it differently for different services.