Who decides what AI says? Musk, Andreessen and Zuckerberg on AI censorship
September 27, 2026 · 26 min read
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.
The rules, numbers and steps worth keeping in front of you.
- Andreessen's view: any moderation system becomes a ring of power that everyone wants to use.
- Andreessen accepts limits: incitement, terrorist recruitment and child abuse material are not protected.
- Andreessen estimates a 20-60-20 split: true believers, dissenters, and 60% who follow whoever seems to be winning.
- Andreessen's remedy is political: elections, transparency, investigations and prosecutions.
- Zuckerberg says ideological censorship pressure began with the 2016 election and COVID in 2020.
- Zuckerberg: fact-checkers were seen as biased in what they chose to check.
- Zuckerberg: the First Amendment binds the government, not companies' moderation.
- Zuckerberg's threshold example: at 90% precision, one in ten removals is a mistake.
- Meta is not easing off on terrorism, child exploitation, fraud or foreign influence networks.
- Zuckerberg wants AI widely distributed through open source.
- Musk warns that an AI rigidly programmed with one set of values could reach catastrophic conclusions.
- Srinivas: an answer engine should say nothing it didn't retrieve, and cite sources.
- Caution: these are the speakers' perspectives; Andreessen's account is openly partisan and Musk's claims are not settled fact.
- Compare several models and test them on questions where you know the answer.
This is the short form. The full insight has every expert's reasoning, their own words and the sources.