Use recommendation systems and platform audits to teach the distinction between a system’s objective, its measured behavior and its effects on people.
Ask students to identify who produced the evidence, what was measured and which inference goes beyond the source. Company descriptions, independent audits and experiments answer different questions. Ranking and personalization cases do not by themselves establish a generative AI architecture or a learning benefit.
These historical records retain their original source-inspection limits. Their dates describe the cases; they are not a fresh review of present-day systems.
8 entries
Netflix released over 100 million ratings and offered one million dollars to any team that beat its Cinematch recommender by a further 10 percent. The contest framed recommending as predicting star ratings, a framing Netflix later moved away from.
Two Netflix employees described the company's recommender, said it influences about 80 percent of streamed hours and estimated its value at over one billion dollars a year. These are company claims; the paper does not show how the estimate was made.
Covington, Adams and Sargin (2016) describe YouTube's two-stage recommender, whose ranking stage predicts expected watch time. The paper is a description of the 2016 system hosted by Google Research, not an independent audit.
Netflix posts describe testing several images per title and then choosing the image shown to each member from their viewing history. The performance claims are Netflix's own. In 2018 users alleged racial targeting; Netflix denied it.
Studies disagree on whether YouTube's recommendations push viewers toward extreme content. Ribeiro and colleagues found movement from milder to more extreme channels, while Hosseinmardi and colleagues and Chen and colleagues found small, concentrated audiences and no systematic recommendation effect.
TikTok's own account of its For You feed lists three groups of signals and says strong signals count for more than weak ones. It is a company description, not an audit, and it gives no weights.
Researchers ran scripted accounts on the TikTok website to test what changes the For You feed. Following a creator had the strongest effect. The abstract and main text disagree on the order of liking and watch time. This is a limited web audit, not a full test of the algorithm.
Spotify researchers reported an online A/B test of a contextual-bandit method for mixing music and podcast shelves on Home. Their paper describes the action, feedback, comparison, and engagement results. This is company-associated experimentation, not evidence about Wrapped or addiction.
No cases match. Try a broader phrase or clear the filters.