When Whip Media killed TV Time to pivot toward enterprise AI, they didn't just sunset an app. They deleted a decade of user preference data in service of building systems that, according to a new academic paper, are doing alignment entirely wrong. The irony is thick enough to stream.
Fixed Preferences Are a Fiction, and AI Keeps Building on Them
A 2026 paper on arXiv by Max Kanwal and Caryn Tran, Constructive Alignment: Governing Preference Dynamics in Human-AI Interaction, argues that most AI alignment frameworks treat human preferences as static targets to be inferred and optimized. TV Time was, at its core, a preference-mapping machine: what you watched, when you bailed, how many times you rewound. That granular behavioral signal is now gone, replaced by a vague promise of enterprise intelligence. The company is building AI products on the assumption that preferences can be abstracted, while the paper argues the entire premise is broken.
Taste Is a Moving Target, Not a Dataset
This matters beyond TV recommendations. Digitas CEO Amy Lanzi told The Verge at Cannes that AI won't save advertising precisely because the industry keeps trying to mechanize something that shifts under pressure. Preference is performative, contextual, and contested. Kyle Chayka's work on algorithmic homogenization maps the same terrain from the culture side: when platforms optimize for inferred taste, they flatten it. TV Time's death is a case study in what happens when a company decides the data it spent years collecting is less valuable than the inference layer built on top of it. That bet may be wrong on both the business and the epistemological level.