Apple’s AI Policy Shift: Why Trump’s Return Changed its Chatbot
Apple revised the evaluation guidelines for its forthcoming chatbot, built on large language models, immediately after Donald Trump’s return to the White House. According to POLITICO, the new instructions for employees of the contractor TransPerfect, based in Barcelona, were issued in March 2025—just two months after the U.S. presidential inauguration. Several hundred staff received updated documents detailing how to assess the AI’s responses to politically sensitive questions.
POLITICO obtained two versions of the handbook—one used throughout 2024 and early 2025, and a second that took effect in the spring. In the latest rules, terms such as “systemic racism” and “intolerance,” once classified as harmful expressions, have been removed. While “discrimination” remains a prohibited category, topics relating to diversity, equity, and inclusion have now been shifted into the “controversial” section.
The documents also instruct reviewers to pay greater attention to the model’s responses about Trump’s supporters and about Apple itself. Whereas the earlier guidelines treated questions about the former president’s supporters merely as examples of political-viewpoint discrimination, the revised instructions caution that labeling them as “radical” could be offensive and stereotypical. Trump’s name appeared three times in the old version but eleven times in the new.
The roster of sensitive subjects has been significantly expanded. Beyond existing categories such as abortion, gun control, LGBTQ+ rights, international disputes, and affirmative action, the list now encompasses elections, vaccination, DEI policies, and even artificial intelligence itself. Entire chapters are dedicated to security circumvention tactics (“jailbreaks”), reflecting challenges faced by competitors like OpenAI. In March, additional geopolitical flashpoints—Gaza, Taiwan, Kashmir, and Crimea—were designated as requiring special scrutiny.
The updated guidelines also include measures to safeguard Apple’s corporate image. Contractors must flag as sensitive any responses concerning the company’s brand, products, or leadership. Special attention is reserved for Tim Cook, Craig Federighi, Eddy Cue, and even Steve Jobs. Negative references to privacy violations, past scandals, or the use of copyrighted material in training the model are considered critical risks that must be carefully monitored.
Programmers tasked with reviewing the model received a 125-page manual filled with examples and threat classifications. It outlines “long-term risks” of AI usage, ranging from psychological dependence on the system and uncritical trust in its outputs to diminished confidence in democratic institutions and the amplification of disinformation.
Apple has officially denied any policy shift, stating that instructions are updated regularly to improve model reliability. Company representatives emphasize that development is guided by “Responsible AI Principles” and that annotators from multiple countries contribute to evaluating response quality across numerous domains. TransPerfect likewise rejected suggestions that the revisions were influenced by external politics, insisting instead that the changes pertain solely to workflow refinements.
Support Our Threat Intelligence
If you find our technology report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
