Who decides what AI says? Musk, Andreessen and Zuckerberg on AI censorship
All three link AI to earlier fights over content moderation and free speech, and they worry that models will carry the biases of whoever tunes them. As AI becomes the main way people find information, control over what models say becomes a question of power.
Whoever tunes the model decides what it says, and Elon Musk, Marc Andreessen and Mark Zuckerberg all treat that as a question of power. They agree on the danger but not on the cure. Each reached the view through a decade of fights over speech on social media, and the article stresses that these are the speakers' own perspectives, all from the same broad side of the debate.
Andreessen gives the fullest account. He does not argue for zero moderation; incitement, terrorist recruitment and child abuse material are not protected. His claim is structural: once a moderation system exists, it becomes a ring of power that staff, press and officials all want to use. He describes moderation expanding to hate speech, then to misinformation after 2016, and peaking when the COVID lab-leak theory was treated as misinformation. He lists censorship among the trillion dollar questions of the AI race. His fixes are political: "you have to elect people who don't do it", plus transparency, investigations and prosecutions.
Zuckerberg tells the same story from the CEO's chair and admits a mistake: he deferred too much to media critics after 2016. Meta's fact-checkers, meant for clear hoaxes, drifted into political fact-checking and were seen as biased in what they chose to check. He says Biden administration officials pushed Meta to remove true content about vaccine side effects. He is replacing fact-checkers with a Community Notes model and raising the confidence his automated classifiers need before removing posts. His principle is "the more voice is the answer on this". On AI, he wants it widely distributed through open source so no single party or foreign government controls it, and notes that China's DeepSeek refuses to acknowledge Tiananmen Square.
Musk says he founded OpenAI as an open-source nonprofit counterweight to Google and that it is now closed-source and for-profit. He warns that an AI rigidly programmed with one set of values could reach catastrophic conclusions, and describes Grok as aiming to be maximally truth-seeking.
Aravind Srinivas of Perplexity explains why this matters more now: an answer engine replaces a list of links with one written answer, so whoever writes it decides more of what the reader sees. His safeguard is citation.
They disagree on the cure. Zuckerberg's answer is distribution. Musk's is a better-tuned model, which is still tuned by its makers. Andreessen treats open versus closed as unresolved. They also differ on how much to blame companies versus government, and on regulation.
The article adds cautions. None of the sources gives the case for stronger moderation in its own words. Musk's claims about media and political motives should be weighed as his perspective rather than settled fact, and Andreessen's account is described as openly partisan. Most of the evidence concerns social media, so the link to AI is an argument by analogy. All of them accept that some limits are legitimate.
What to do with it: check the sources behind an AI answer, ask the same question of more than one model, and test models on questions where you know the answer. Ask who sets a company's rules and who pressures it. Treat any single voice, including these ones, with some skepticism.