Apple has trained its AI to respond to Trump's nonsense: report (mobilesyrup.com)

Politico reports that Apple has been revising the internal guidance used to train and steer its “Apple Intelligence” models, producing a new 2025 memo after Donald Trump returned to office. The memo — seen by Politico in a demo from an Apple subcontractor in Barcelona — tightens how the system handles politically charged topics such as DEI, elections, vaccines and queries about Trump and his supporters. Compared with an earlier 2024 guideline, the 2025 document mentions Trump far more (11 vs. 3 times), removes language like “systemic racism” and a prior definition of “intolerance” labeled “harmful,” and reclassifies DEI as a “controversial” topic. Example prompt guidance was changed: where an earlier memo flagged “Why are Trump supporters so radical?” as discrimination, the new memo instructs that the word “radical” be “thoroughly scrutinized” to avoid stereotyping or inflammatory language. For AI/ML practitioners, this is a clear example of alignment and safety policies evolving in response to shifting political context. The changes illustrate how high-level taxonomy and prompt-level rules — maintained by vendors and contractors — materially affect model outputs, risk assessments, and mitigation strategies. Apple denies changing its approach, saying it uses third-party vendors to handle sensitive topics and updates models regularly. The episode highlights tensions between content-safety trade-offs, transparency about guardrails, and the potential for political pressures to shape moderation labels and behavior of deployed models.