TikTok algorithm to be retrained on US user data under Trump deal (www.bbc.com)

The White House says a deal has been reached that will copy and retrain TikTok’s recommendation algorithm using only US user data, with the recommendation system audited by Oracle and operated by a new US-based joint venture of “patriotic” investors (including Silver Lake) to satisfy national security concerns. President Trump is expected to sign an executive order outlining the arrangement and a 120-day pause on enforcement to allow the sale/transition to close. TikTok already stores data for about 170 million US users on Oracle servers under “Project Texas,” but it’s unclear whether China has formally approved the cross-border aspects required to complete the deal. Technically, the plan calls for inspection of source code and the recommendation system, then rebuilding or retraining a US-only model using US data—effectively creating a separate recommendation pipeline and operational control inside the US cloud. That could reduce foreign data access and lower some national‑security risks, but it also risks fragmenting TikTok’s global feed: analysts warn a US-only algorithm or content isolation could change user behavior, reduce creator value, and degrade experiences. The move underscores tensions between data governance, model security, and product utility as regulators push to localize AI-driven recommendation systems.