TikTok's "25+ AI Models Watching Every Scroll": What an Independent Researcher Claims, and What's Still Unverified

On 3 November 2026, Ayo Omolewa — known as @ayoships on X — posted a viral thread: according to Omolewa, TikTok runs more than 25 AI models, on-device and on its own servers, to watch in real time how a user scrolls the feed. The post, amplified by a repost (account MotionViz) that reached millions of views, describes a system able to predict whether a video will be skipped, finished, or liked before it even starts playing. None of these claims has been verified by an independent source, nor confirmed by TikTok. That is exactly why it's worth reading this not as news about what the app does, but as a case study in what happens when the transparency that European law requires is missing.
What Omolewa said, and who they are
According to Omolewa, the system described in the 3 November thread sends more than 50 signals with every request — battery level, screen brightness, silent/ringer status — and uses motion sensors to infer whether the user is lying down, walking, or in a car. The app, still according to the author, also collects messages typed and never sent (draft comments or DMs), and when it "thinks it's losing" the user, it tries to pull attention back: the source, reconstructed from multiple convergent citations, literally describes that if the user keeps skipping content, the app pulls a fresh batch of videos or inserts a livestream into the feed — a more sober formulation than the "emergency videos" phrasing that circulated in reposts. Omolewa further claims a model detects whether the user has fallen asleep, and that the system predicts when the user will reopen the app. In a second post, on 5 November, Omolewa adds a distinct claim: that TikTok also estimates age, gender, mood and an attractiveness score from the user's face.
Who the source is matters as much as the content. Omolewa is not an academic researcher, nor the author of a commissioned audit: they are a data scientist, co-founder of a0.dev (Y Combinator W25), with a public track record of reverse-engineering consumer apps — Candy Crush, Tinder, Subway Surfers, Duolingo, Block Blast. So this is a recognisable methodology, not an anonymous account — but it remains a one-sided claim: everything reported here should be read as "according to Omolewa," not as established fact.
What can be verified, and what can't
For the behavioural claims (the 25+ models, pre-play scoring, the 50+ signals, motion-sensor inference, message drafts, "attention recovery," sleep detection, reopen-time prediction), no paper, code repository, or independent technical audit has been found to support them. No secondary source consulted goes beyond repeating the original claim. TikTok, for its part, has not responded specifically to these allegations — not to be confused with the company's generic reply to the European Commission in the DSA proceeding discussed below, which concerns a separate file entirely.
The only independent technical corroboration found is a community project on GitHub that picked up the facial-recognition work: it confirms that something extractable from the app does exist, but only for the age/gender/mood/attractiveness models from the 5 November post — not for the behavioural scroll-scoring claims. A critical thread circulating in parallel raises methodological doubts, but it too is limited to the attractiveness score: it is not a technical rebuttal of claims 1-8, which simply remain unverified, not disproven.
If even part of it were true: the regulatory framework is already there
The interesting point for compliance readers isn't whether Omolewa is right. It's that, if even part of these claims were confirmed, they would fall squarely within an existing regulatory perimeter.
DSA Article 27 requires all online platforms — not just designated Very Large Online Platforms — to explain, in plain language, the main parameters of their recommender systems and the options to adjust them. A behavioural scoring system like the one Omolewa describes, if it existed as reported, is exactly the kind of mechanism that transparency obligation is meant to make visible.
DSA Articles 34-35, on systemic risk assessment and mitigation, are the same ground on which the European Commission has already opened a preliminary finding against TikTok — but for a different reason: the "rabbit hole" effect of the "For You" feed, which we covered in a previous article. The two episodes aren't the same story, but they share the same type of scrutiny: a recommender algorithm whose inner workings stay opaque to users, third-party regulators, and independent auditors alike.
On the privacy side, the GDPR is a framework of principle to invoke carefully. Article 5(1)(c) on data minimisation is relevant: if the signals Omolewa describes were collected, they would need to be proportionate to their purpose. Article 22, on the right not to face a decision based solely on automated processing, is a contextual principle here: applying it to a recommender system is questionable, because the required threshold — a legal effect or "similarly significant" impact — fits cases like credit scoring or hiring, not feed personalisation.
More specific is the ePrivacy angle, Article 5(3), on access to device state and sensors. The EDPB Guidelines 2/2023 clarify that the provision does not apply when processing stays entirely on-device. It does apply, however, when those signals are collected and transmittedto servers — exactly the scenario Omolewa describes for the 50+ signals sent with every request. If confirmed, this is the correct regulatory anchor — not a "sensor access requires a legal basis" rule.
Finally, the AI Act, relevant only to the distinct 5 November claim about mood and facial attractiveness. The prohibition in Article 5(1)(f) on emotion inference from biometric data applies only to workplace and education settings: TikTok falls outside that scope. The correct anchor is Article 50(3) (transparency toward people exposed), together with the high-risk classification under Annex III. Behavioural claims 1-8 on scroll-scoring are not emotion recognition and do not invoke these articles.
Why it matters even without technical confirmation
The point isn't whether there are really 25 models, or really 50+ signals. It's that, today, no user, third-party regulator, or independent auditor can verify it from the outside — and that is exactly the kind of opacity the DSA and AI Act aim to close with transparency and audit obligations. In the same month the European Commission contests TikTok's risk assessment of the "For You" feed, an independent reverse engineer describes — with no obligation to prove it, nor to be disproven — the inner workings of a different algorithmic system on the same platform. Two separate episodes, but the same kind of scrutiny missing from the inside and arriving from the outside — in one case from Brussels, in the other from an X account.
What changes for algorithm governance
"Ex post" transparency — a single reverse engineer's claim, confirmed or not — doesn't replace the obligation of ex-ante audit and risk assessment. Anyone running recommender or profiling systems can't wait for transparency to arrive via a viral post: it's worth documenting today which device or behavioural signals feed your system, under what legal basis, and with what risk assessment — before a regulator asks, or a reverse engineer with an X account does it for you.
Do you know which signals feed your recommender system?
Signals collected, legal basis, risk assessment: documenting them today is what separates an auditable algorithmic system from one you only discover when someone else takes it apart. We can help you map them and put them on paper.
Let's talk about your algorithmic governance →