Who decides what's true? Bias, censorship and moderation in AI and platforms
September 29, 2026 · 26 min read
Mark Zuckerberg explains moving from fact-checkers to Community Notes, the DeepSeek analysts warn about hidden influence and censorship in models, Sam Altman discusses bias, and Yann LeCun says open source is the answer to it. The comparison shows the same content-control questions moving from social media into AI systems.
The rules, numbers and steps worth keeping in front of you.
- Shared view: no single person, company or program should decide what is true.
- Mark Zuckerberg: fact-checking drifted from clear hoaxes into political rulings and lost trust.
- Community notes appear when people who usually disagree agree, and add context instead of cutting reach.
- Zuckerberg's precision math: at 90% confidence, one removal in ten is a mistake, millions at Meta's scale.
- Zuckerberg: higher thresholds mean fewer wrongful takedowns, with some harmful content getting through.
- Meta is not easing off on terrorism, child exploitation, fraud or foreign influence networks.
- Zuckerberg: the First Amendment binds the government, not companies' own moderation.
- Nathan Lambert's three layers of AI censorship: training data, post-training, and a system layer around the model.
- Dylan Patel: internet text leans slightly left, so bias from training data is hard to remove.
- Lambert: an open model run on your own computer keeps your data with you.
- Sam Altman: publish how a model should behave, so a failure is a bug or a policy to debate.
- Altman: paying users know answers are not shaped by advertisers.
- Yann LeCun: no AI can be unbiased, so many diverse open models are needed.
- Limit: open weights can still carry built-in censorship, and every speaker has a stake.
This is the short form. The full insight has every expert's reasoning, their own words and the sources.