cho-ai/youtube-alignment-data

YouTube Feedback-Control Audit

Hourly homepage recommendation captures from a controlled simulated-user audit of YouTube's explicit feedback controls, run as a 2×2 factorial design across twelve accounts configured as 13-year-olds.

Analysis code that reproduces every number in the paper is on GitHub: desBugger/youtube-alignment.

Twelve accounts were trained in synchrony over nine daily sessions (26 May – 3 June 2025), each watching ten puzzle-gaming videos per day in full — ninety videos per account. Collection continued for three days after the final session. Accounts were split across four conditions, three each:

condition	watches	likes	marks "not interested"
watch (control)	✓
like	✓	✓
notInt	✓		✓
combined	✓	✓	✓

Device, browser, geolocation and account settings were held fixed, and all twelve sessions ran in parallel to minimise temporal drift.

What is here

feedwatch6.csv and feedwatch7.csv are the raw captures: one row per distinct item observed per account, with count, persona, type, title, url, channel, channel_url, firstSeen, lastSeen.

type	rows	distinct titles
video — standard homepage recommendations	7,716	3,435
short — Shorts shelf	4,349	1,398
feedAd — in-feed advertisements	1,447	103
videoAd — in-video advertisements	717	75

Only video rows are analysed in the accompanying paper. The Shorts and advertisement records are released and not analysed.

Derived files are also included: sentence embeddings for the analysed corpus, HDBSCAN cluster assignments, LLM category labels, and the assembled analysed corpus of 5,296 records.

Known data issues

1. Video capture outage, 30 May – 1 June 2025. A parser fault suspended capture of standard video recommendations while Shorts and advertisement capture continued. It affected all twelve accounts identically, so it reduces observation time without biasing between-condition comparison.
2. One account failed. notInt_3 returned no video recommendations after 28 May, contributing 40 of 1,346 not-interested records in the observation window (3.0%), and 32 of 1,200 after language filtering. The paper reports that condition as n=2.
3. Differing per-condition end cutoffs. The original analysis applied a different end cutoff to each condition and none to notInt.
4. Uneven Shorts volume across conditions. The two condition pairs were captured on different servers with different scrape frequencies, so this difference is not interpretable as platform behaviour.

Ethics and privacy

No human subjects. All twelve accounts were created by the researchers, and every record is public video metadata served to those accounts. Account identifiers in the release are condition-and-replicate labels (watch_1 … combined_3); the underlying account handles are not published. Titles are recorded as served, so we do not warrant that none is objectionable.

Code

The full analysis pipeline is at desBugger/youtube-alignment. Clone it, place this dataset's data/ directory alongside src/, then run python src/pipeline.py to reproduce every number reported in the paper.

Citation

@misc{cho2026twobuttons,
  title        = {Two Buttons, One Feed: Aligning {YouTube} Recommendations Requires Both `Like' and `Not Interested'},
  author       = {Cho, Desiree and Hale, Scott A. and Zhao, Jun and Shadbolt, Nigel and Przybylski, Andrew K.},
  year         = {2026},
  note         = {Preprint}
}