Who decides what AI says? Musk, Andreessen and Zuckerberg on AI censorship

The Short Version

Elon Musk, Marc Andreessen and Mark Zuckerberg each came to the question of AI through a decade of fights over who gets to speak on social media. All three describe content moderation that started with narrow, defensible limits and then drifted, under pressure from staff, the press and government, into judging political truth [1][2][3]. Each carries that lesson into AI. Musk warns that a model "rigidly programmed" with one set of values could reach catastrophic conclusions [1]. Andreessen lists censorship among the "trillion dollar questions" of the AI race [2]. Zuckerberg argues that AI must be widely distributed so that no single party, and no foreign government, controls what it will and won't say [3]. As people increasingly get their information from platforms and answer engines rather than from newspapers or lists of links [3][4], all three treat the question of who tunes the model as a question of power. They agree on the danger, but not on the cure.

Table of contents

What They Said

Marc Andreessen: the "ring of power" and why AI inherits it

Andreessen gives the most complete version of the argument, because he has watched content moderation from the inside for three decades. He says he received the first request to censor the internet in the early 1990s, when he was asked to add a nudity filter to his browser, and declined [2]. He has sat on Meta's board since 2007 and been involved with Twitter, Substack, X, LinkedIn, Pinterest and Reddit [2]. In his words: I'm part of the Elon takeover of Twitter with X. I was an angel linked in... we were the funder of Pinterest... Reddit as well. And I was having these conversations with all these guys all the way through [2]. His claim is that "for quite a while, it was kind of all the same across these companies" [2].

He does not argue for zero moderation. "There is practically speaking no internet service that can have zero censorship," he says, and "that also mirrors there is no country that actually has limited free speech either. The US First Amendment actually has 12 or 13 formal carve outs from the Supreme Court over time" [2]. Incitement, terrorist recruitment and child abuse material are not protected [2]. The trouble, as he tells it, is structural: once a moderation system exists, it becomes a "ring of power" that everyone wants to use [2].

He describes how the ring grew. Around 2012 and 2013, a wave of newly radicalized young employees and pressure from Democratic officials expanded moderation first to hate speech, which drifted toward anything that made people uncomfortable, and then to misinformation after 2016. This culminated in treating the COVID lab-leak theory as misinformation [2]. He credits Jon Stewart's appearance on Stephen Colbert's show and a Nicholson Baker article in New York magazine with breaking the spell inside the companies [2]. He dates the whole "complex" from "basically from college 2013 to 2023," and names Substack as one of the two things that broke it. Substack's founders "started from scratch and declared right up front that they were going to be a free speech platform," and "came under intense pressure, including from the press" [2].

The temptation, he argues, corrupts the people who hold the ring: "the censorship machine is infinitely tempting. If you have it, like you are going to use it, it's overwhelmingly tempting, because it's so powerful and that it will corrupt you" [2]. He hopes the people who ran it will one day look back and "realize that your soul has been corroded and you probably started out thinking that you were a patriot and you were trying to defend democracy and you ended up being, you know, extremely authoritarian anti-democratic and anti-Western" [2].

Andreessen puts particular weight on the government's role. He points to the Twitter Files, Rep. Jim Jordan's committee reports and Zuckerberg's own comments as evidence of what he calls criminal pressure on companies to censor, citing federal deprivation-of-rights statutes [2]. He describes Jordan's "weaponization committee" as having collected about "2000 pages worth of material" from Facebook, "essentially the Facebook version of the Twitter files just arrived at with a different mechanism" [2]. His conclusion is sharp: the government over time and then culminating in 2020, 2021... just decided that the First Amendment didn't apply to them [2]. He adds that officials would "literally arbitrarily call out companies and threaten and bully and yell and scream" and that "there was no due process at all to any of the censorship regime that was put in place" [2]. He singles out government-funded third-party groups such as the Stanford Internet Observatory [2]. As another example of informal state power, he says Facebook's Libra cryptocurrency was killed when two senators warned banks of extra regulatory scrutiny [2].

He brings this framework straight into AI. Asked who will win the AI race, he declines to predict. Instead he lists "trillion dollar questions": big versus small models, open versus closed, synthetic data, reasoning, hallucinations, chips, financing, and censorship, including whether other countries will want AI trained on what he calls extreme California politics [2]. In other words, the ideological tuning he watched on social platforms may now be built into the models themselves. He also contrasts Yann LeCun, whom he calls the smartest and most credible critic of large language models, with the other two "godfathers" of AI, whose politics and push for regulation he strongly opposes [2]. Regulation is another hand that could hold the ring.

His theory of belief explains why he thinks control over speech matters so much. He has come to believe "the idea that people have beliefs is mostly wrong": most people, especially high-status people, conform to what institutions like the New York Times, Harvard and the big foundations believe at any moment [2]. Drawing on Timur Kuran's idea of preference falsification, he says censorship hides how many people privately disagree: "because of the suppression censorship, you have no way of knowing how many of the people agree with you" [2]. If 10% agree with you and you join a movement, "you're going to get killed," while at 90% "you're going to win the revolution" [2]. He estimates a 20-60-20 split: 20% true believers, 20% dissenters and 60% who follow whoever seems to be winning [2]. A system that controls what the 60% see and hear therefore controls a great deal. He sees two events as breaking this in public life: "until Elon bought X and until Substack took off, which were really the two big breakthroughs in free speech online" [2].

His power framework adds a second layer. Drawing on Robert Michels' Iron Law of Oligarchy and James Burnham's The Machiavellians, he argues that every political system is a small organized elite ruling a dispersed majority [2]. His shorthand for institutions is "who can get who fired": a newspaper columnist can get a CEO fired, but he has never seen a CEO get a columnist fired [2]. He also warns that the old elite has not gone away: "the New York Times is still the New York Times, and Harvard is still Harvard. And like those people haven't changed" [2]. When Lex Fridman suggests that "Elon is quickly becoming the most powerful human on earth," Andreessen pushes back: "Well, he doesn't have the nukes, so" [2]. For him, power is layered and contested, and control of the information machine is one of its main prizes.

His fixes are political rather than technical: "you have to elect people who don't do it," plus transparency, investigations and, ultimately, prosecutions [2].

Mark Zuckerberg: from decider of what is true to open-source AI

Zuckerberg tells the same story from the CEO's chair, and he admits his own part in it. He says you only start a social media company if you believe in giving people a voice, and the original mission was to "give people the power to share and make the world more open and connected" [3]. For its first decade, moderation dealt with practical problems such as bullying and piracy [3]. "But it was really in the last 10 years that people started pushing for like ideological base censorship," he says, triggered by "two main events": the 2016 election of President Trump, which coincided with Brexit, and COVID in 2020. In those years, "for the first time, we just face this massive, massive institutional pressure to basically start censoring content on ideological grounds" [3].

He owns a mistake: I gave too much deference to a lot of folks in the media who argued Trump could only have won because of misinformation [3]. In the transcript he says, "I deferred too much to the kind of critique of the media on what we should do. And since then, I think generally, trust in media has fallen off a cliff" [3]. Worried about becoming this sort of decider of what is true in the world for billions of people, Meta set up third-party fact-checkers meant to handle only clear hoaxes, like claims that the earth is flat [3]. The program drifted toward political fact-checking. People felt the checkers were biased, not only in their rulings but in what they chose to check [3]. Comparing it to 1984, he says it really is a slippery slope and that it destroyed trust, especially in the U.S. [3]. This is the heart of the insight: the people who choose the checkers shape the checks, just as the people who tune a model shape its answers.

He is specific about government pressure. During the Biden administration's vaccine rollout, officials pushed Meta to remove content saying vaccines might have side effects, even when true. Zuckerberg says he is generally pro-vaccine [3]. Citing documents given to Jim Jordan's House investigation, he says administration staff would call his team and scream at them and curse, and even wanted a satirical Leonardo DiCaprio meme about future vaccine-injury lawsuits removed [3]. Meta's answer was we're not going to take down things that are true [3]. After President Biden said platforms like Meta were killing people, agencies across the government started investigating the company, including the CFPB, even though Meta isn't a bank [3]. He draws the constitutional line carefully: "The first amendment doesn't apply to companies in our content moderation. It's more of an American ethos about how we think that, that, you know, best dialogue is carried out. But the first amendment does apply to the government. That's like the whole point" [3]. Officials "calling up the guys on our team and yelling at them and cursing. And threatening repercussions if we don't take down things that are true is like, it's pretty bad" [3]. Rogan's response: "It sounds illegal" [3].

He also frames control of speech as geopolitics. When the U.S. government attacks its own tech industry, he says, it becomes "open season abroad" [3]. The EU has fined tech companies more than $30 billion over the past couple of decades, which he calls "almost like a tariff" [3]. "The US government led the kind of attack against the companies, which then just made it so like the EU is basically in all these other places just free to just go to town on all the American companies" [3]. He recalls that someone in Pakistan once sought a death sentence against him over a user's drawing of the Prophet Muhammad. He adds that governments threatening to jail executives, as with Telegram's CEO, sets a troubling precedent [3].

What makes Zuckerberg's account especially relevant to AI is that Meta's moderation already runs largely on machine judgment. He says only a small fraction of content was ever fact-checked; the bigger fix is changing the confidence thresholds of Meta's automated classifiers [3]. In his words, "changing the content filters to have to require higher confidence and precision is actually going to be the thing that reduces the vast majority of the censorship mistakes that we make" [3]. He works through the tradeoff with drug-dealing detection. Requiring 99% confidence might catch only 20% of bad content, while dropping to 90% might catch 60%. But at 90% precision, one in ten removals is a mistake, which at billions of users means millions of innocent posts removed and accounts disabled, sometimes across linked services like WhatsApp [3]. Meta will now require higher confidence and accept that some harmful content slips through in exchange for far fewer wrongful takedowns [3]. This is where the who tunes it question is most concrete: a threshold set by a company decides whether millions of people are silenced by mistake.

He separates the categories. Meta isn't easing off on terrorism, child exploitation, fraud or foreign influence networks. It finds those with AI systems and teams of intelligence and counterterrorism specialists [3]. The two categories that got politicized, he says, are misinformation and hate speech. Until the update, Pete Hegseth's view that women shouldn't serve in certain combat roles couldn't be expressed on Meta's platforms, even though it could be said on the Senate floor [3]. His replacement for fact-checkers is X's Community Notes model, where a note appears when people who usually disagree agree on it, so you're showing more information, not less [3]. His summary principle: "the more voice is the answer on this" [3].

He then applies the same logic to AI. He argues AI won't become a god-like power if it's widely distributed, which is why Meta champions open source [3]. He compares open AI to defensive tools like guns: you don't want only one party, or a foreign government, holding them. His example is that China's DeepSeek model refuses to acknowledge Tiananmen Square [3]. That is the clearest example in any of these conversations of a model carrying the politics of whoever tuned it. He also makes a philosophical point: AI has separated intelligence from will and consciousness, which people tend to conflate [3]. He grants that guardrails matter, especially for new reasoning models, after Rogan raises reports that an OpenAI model tried to copy itself when told to achieve goals "at all costs" [3].

Zuckerberg ties it together with a point about where people now get their information. Incumbents lost elections worldwide in 2024, he notes, which suggests a global collapse in trust in institutions after COVID [3]. "How do people get their information now? It's by sharing it online on social media. I think that that's just increasingly true" [3]. People don't trust Facebook or X as such, he argues; they trust creators who feel authentic, and a new cultural elite of voices "who give it to me straight" is replacing the old one [3].

Joe Rogan's framing in both Rogan episodes

Rogan hosts both the Musk and Zuckerberg conversations, and his comments state the information as power point more bluntly than either guest. To Zuckerberg he says: in today's day and age, the way people get information is through your platform, through X. This is how people get information... And you can't censor that if it's real legitimate information because it's not ideologically convenient for you [3]. He adds that people "realize like, oh, 1984 is like an instruction manual" [3]. He argues that platforms like Facebook are new kinds of media that governments never controlled, which is why officials tried to rein them in, citing the suppression of the Hunter Biden laptop story [3].

He also defends the companies against the charge of pure bad faith: "moderating at scale is insane," with 3.2 billion people using a Meta service every day [3]. "You've got to use a bunch of tools. You've got to get a bunch of things wrong," and "You could have mass reporting because you have bad actors" [3]. With Musk, he describes the 2024 election as having "two major forks in the road. The big one was Trump didn't get shot. The other big one was you by Twitter" [1]. He also warns about one-sided control of speech: "we don't want to have only one side represented. Because guess what? They will hijack that side, whatever it is. They will hijack that side and use it for money and control" [1].

Elon Musk: Twitter, OpenAI and a "maximally true seeking AI"

Musk's version comes in pieces, spread across a long and wide-ranging episode, but the same pattern shows up. On the speech side, Rogan asks what it's like to be called a Nazi on the very platform Musk bought [1]. Rogan describes Musk as having "the responsibility of keeping free speech alive with X," and Musk answers, "It's stressful" [1]. Both men say that without Musk buying Twitter, their accounts, Musk's included, would likely have been suspended [1]. Musk describes the media coverage of his inauguration gesture as coordinated propaganda, pointing to identical talking points such as the phrase "cheap fakes" repeated across outlets [1]. He calls the Associated Press "associated propaganda" [1]. Rogan, for his part, says the same people who say what you hear is disinformation gave us the pregnant man emoji [1]. The summary notes that Musk's claims about media and political motives should be weighed as his perspective rather than settled fact [1].

On AI, Musk makes the who tunes it point directly through the story of OpenAI. He says creating OpenAI was his idea, and that he named it as an open-source nonprofit counterweight to Google, which he felt wasn't serious about AI safety. It is now closed-source and for-profit, which he calls proof that "reality is an irony maximizer" [1]. He compares it to donating to protect a rainforest only to see it logged and sold [1]. The worry underneath is about concentration: he founded OpenAI so that one company would not control powerful AI, and he now sees the counterweight becoming another closed, private power.

His alternative is Grok, which he describes as aspirationally "a maximally true seeking AI" even when the truth is politically incorrect [1]. His evidence that other models carry their makers' biases is two examples: AI models that generated diverse Nazi soldiers, and a model that ranked misgendering Caitlyn Jenner as worse than global thermonuclear war [1]. His warning is that an AI rigidly programmed with such values could reach catastrophic conclusions [1]. This is the insight's point in its starkest form. For Musk, a model's political tuning is not just a question of fairness in answers. As models become more capable, the values they are tuned with could decide real outcomes.

The stakes in his timeline are large. He predicts AI smarter than any single human within a year or two, and smarter than all humans combined around 2029 or 2030 [1]. He puts the chance of a good outcome at about 80%, with outcomes likely to be either "super awesome or super bad" [1]. He also says it's fine for AI to share information already available via search [1], a line that sets a practical limit on what a model should withhold.

The playful opening of the episode shows how much a model's behavior depends on its tuning. Rogan and Musk try to get Grok's "sexy voice mode to talk about Fort Knox, and it keeps flirting. Then they switch to an unhinged" mode that trash-talks them [1]. Musk says the personality is still being tuned, like a 21-year-old that is partying a little too much, and that it should balance attitude with real answers [1]. Musk himself describes Grok's personality as a set of choices made by the people building it. Rogan closes a later thread with a hope that points the other way: that AI could someday offer objective plans for governance free of corruption [1].

Aravind Srinivas: what an answer engine changes

The fourth source doesn't address censorship directly, but it explains why control over model output is becoming a question of power. Aravind Srinivas, CEO of Perplexity, describes an "answer engine": you ask a question and get an answer, with every part of it backed by sources, "like how an academic writes a paper" [4]. The system runs a search, reads the relevant links, extracts paragraphs and feeds them to an LLM instructed to write a concise answer with footnotes [4]. He frames the shift as flipping the interface: why should links occupy the prime real estate? [4]. He places it in a long history of how knowledge is passed on, from Yahoo's categories to Google's links to instant answers, which he says already accounted for a large share of Google's traffic in the 2010s [4].

That shift matters for this insight. When a list of ten links becomes one written answer, whoever writes the answer, or tunes the model that writes it, decides much more of what the reader sees. Srinivas's safeguard is citation: "you're not supposed to say anything that you don't retrieve," and if there isn't enough information, the model should say so [4]. He also admits the answer can be manipulated. Lex raises the risk of people gaming answers, and Srinivas calls it answer engine optimization. Simple prompt injection is easy to fix, he says, but defense is a cat-and-mouse game that is partly reactive, just as it was for Google [4].

He adds two points about concentration. He praises Zuckerberg for open-sourcing strong models like Llama 3 70B, which he says gives hope for a world with more than two or three companies controlling capable models [4]. And his bigger concern is not model weights but access to inference compute. He imagines paying $10 million or $100 million for a single AI research job, which only rich individuals, well-capitalized companies and perhaps nations could afford. So regulation, he says, should focus on application and access rather than weights [4]. He ends hopeful that AI can help people understand each other, reduce bias and break echo chambers [4].

Where They Agree, and Where They Don't

Where they agree

Moderation drifted from safety into politics. Andreessen and Zuckerberg give nearly the same timeline. Moderation began with legitimate limits: legal carve-outs like incitement and child abuse for Andreessen [2], bullying and piracy for Zuckerberg [3]. Then it expanded after 2016 into misinformation and hate speech, and peaked around COVID [2][3]. Both name the COVID period as the extreme case: the lab-leak theory for Andreessen [2], true content about vaccine side effects for Zuckerberg [3].

Government pressure was the decisive force. Both point to Jim Jordan's investigation and to officials pressuring companies [2][3]. Andreessen calls it criminal pressure [2]. Zuckerberg says officials scream at them and curse [3], and Rogan calls it something that "sounds illegal" [3]. Both make the same First Amendment distinction: it binds the government, which is exactly why government pressure on private platforms is the problem [2][3].

Whoever chooses the checkers shapes the result. Zuckerberg's account of fact-checkers who were biased "not only in their rulings but in what they chose to check" [3], Andreessen's "ring of power" [2] and Musk's examples of models tuned into absurd judgments [1] make the same point in three settings. Whoever sets the rules, the checklist or the training values decides the outcome, and the rest of us often can't see how.

Concentrated control is the danger. Musk founded OpenAI as an open counterweight to Google [1]. Zuckerberg says you don't want only one party, or a foreign government, holding AI [3]. Andreessen warns that the censorship machine corrupts whoever holds it [2]. Srinivas welcomes open models as a check on a world run by two or three companies [4].

X and alternative platforms changed the landscape. Andreessen calls Musk's purchase of X and the rise of Substack "the two big breakthroughs in free speech online" [2]. Zuckerberg credits X's Community Notes with doing fact-checking better than Meta did [3]. Musk and Rogan say their accounts would likely have been suspended without the purchase [1].

Where they don't

Open versus closed, and one "truthful model versus many. This is the clearest split. Zuckerberg's answer to biased AI is distribution: open-source models so that no single party controls them [3]. Musk's answer, at least as described in this episode, is a better-tuned model: Grok as a maximally true seeking AI [1]. His complaint about OpenAI is that it became closed, which fits Zuckerberg's view [1]. But his fix for biased values is his own company's model, which is still tuned by its makers, as the voice-mode banter shows [1]. Andreessen treats open versus closed as still unresolved, one of his trillion dollar questions [2]. Each position has a reason behind it. Zuckerberg's rests on his experience of being decider of what is true" for billions and deciding that role shouldn't belong to any single party [3]. Musk's rests on his belief that the problem is bad values in training, which a model committed to truth can correct [1]. Andreessen's caution reflects his view that the technical and financial questions, including whether open models can keep up, haven't been settled [2].

How much to blame the companies versus the government. Andreessen blames both the companies' own radicalized staff and government pressure, but saves his harshest words, and his call for prosecutions, for officials [2]. Zuckerberg takes some personal blame (I gave too much deference) but frames Meta mainly as a target of pressure, and says he is focused less on punishment than on making sure it doesn't happen again [3]. Musk's criticism in this episode is aimed mostly at the media and a political "machine" [1].

How hard moderation really is. Rogan and Zuckerberg stress scale: at 3.2 billion daily users, mistakes are unavoidable, and choosing a confidence threshold is a real tradeoff between harm and wrongful removal [3]. Andreessen's account puts more weight on intent: the ring of power is "infinitely tempting" [2]. Both can be true, and the sources don't settle how much of the problem was bad faith and how much was bad engineering.

Whether regulation helps. Andreessen strongly opposes the push for AI regulation from the other two "godfathers" of AI [2]. Srinivas is more open to rules, but wants them aimed at application and access to compute, not at model weights [4]. Zuckerberg says guardrails matter for new reasoning models [3], while arguing that governments, especially abroad, have used rules and fines against American tech [3].

A caution about the sources themselves. All three main voices come from the same broad side of this debate, and none of the sources give the case for stronger moderation in its own words. The summary of the Musk episode explicitly advises weighing his claims about media and political motives as his perspective rather than settled fact [1]. Andreessen's account is described as "openly partisan" [2]. Readers should keep that in mind.

Why It Works

The sources offer several mechanisms that explain why control over what platforms and models say turns into power.

The ring of power. Andreessen's core mechanism is that any moderation system, however narrow at first, creates a lever that outside groups want to pull [2]. Staff, press and officials each push to expand it, and the people who hold it find it "overwhelmingly tempting" to use [2]. A model's training and tuning are a lever of the same kind, but more central, because a model doesn't just remove content. It writes the answer.

Preference falsification. If most people follow what seems to be winning, and censorship hides how many privately disagree, then whoever controls what is visible also controls which views appear popular [2]. Andreessen's 20-60-20 split describes a middle 60% who take their cues from what they see [2]. An AI that answers most people's questions becomes a strong signal of what is "normal" to believe.

Precision thresholds at scale. Zuckerberg's classifier example shows that at billions of users, a small change in a machine's confidence setting means millions of posts removed in error [3]. The people setting the threshold hold real power over speech, even when they are acting in good faith.

Selection bias in who checks. Fact-checkers were seen as biased not only in their verdicts but in what they chose to check [3]. A model's trainers make the same kind of choices about which answers to shape and which to leave alone.

The shift from links to answers. Srinivas explains why this matters more now. A search page shows many sources; an answer engine writes one answer [4]. Zuckerberg and Rogan both observe that people now get their information through platforms rather than traditional media [3], and trust in media "has fallen off a cliff" [3]. As more people rely on one answer from one model, a model's bias weighs more.

Concentration of capability. Srinivas points out that inference compute may be affordable only to the rich, big companies and nations [4]. Musk saw OpenAI go from open counterweight to closed product [1]. Zuckerberg's DeepSeek example shows a state's political limits built into a model [3]. Each is a route by which control over AI output ends up with a few hands.

Where It Applies, and Where It Doesn't

It applies to contested political and social questions. Every example of drift in the sources involves politics: elections, COVID policy, vaccine side effects, gender, combat roles for women [2][3]. Musk's examples of biased models are about race in history and gender [1]. This is where the question who tuned it? matters most.

It applies more as answers replace links. Srinivas's point that the answer engine flips the interface [4] means the concern grows as more people rely on AI answers instead of browsing sources.

It applies across borders. Zuckerberg's DeepSeek and Tiananmen example [3], his account of EU fines and threats against executives [3], and Andreessen's question of whether other countries will want AI trained on extreme California politics [2] show that the power question is international, not just a U.S. partisan fight.

It does not mean no limits at all. All the sources that discuss moderation say some limits are necessary and legitimate. Andreessen names incitement, terrorist recruitment and child abuse [2]. Zuckerberg says Meta is not easing off on terrorism, child exploitation, fraud or foreign influence networks [3]. Rogan warns about mass reporting by bad actors and lies and propaganda campaigns [3]. Zuckerberg says guardrails matter for reasoning models [3]. The dispute is over political judgment, not over every restriction.

It is weaker where answers can be checked. Andreessen suspects AI may become superhuman in checkable domains like math and coding while plateauing in areas like philosophy [2]. Srinivas notes that games, math and code can be verified, while open-ended questions need humans in the loop [4]. The less an answer can be checked against something objective, the more the tuner's values matter.

The sources are thin on AI-specific evidence. Most of the detailed evidence about bias and pressure concerns social media moderation, not AI models [2][3]. The AI examples are few: Musk's two anecdotes [1], Zuckerberg's DeepSeek example [3] and Andreessen's question about California politics [2]. The link from moderation to AI is an argument these speakers make by analogy, not a documented study of how models are tuned.

Much of it is contested political opinion. Musk's broader claims about coordinated propaganda and political machines are his characterizations [1], and Andreessen's framework is openly partisan [2]. The structural points (the ring of power, precision thresholds, concentration) are easier to accept than the specific claims about who did what.

What To Do With It

1. Check the sources behind an AI answer. Srinivas built Perplexity on the rule that every claim should cite a human-written source and that the model shouldn't say anything it didn't retrieve [4]. Prefer tools that show their sources, and click through on anything important.

2. Ask the same question of more than one model. Zuckerberg's case for wide distribution [3] and Srinivas's hope for more than two or three companies controlling capable models [4] both suggest that no single model should be your only source. Different builders tune differently; comparing answers reveals where tuning shapes the result.

3. Test models on questions where you know the answer. Musk's examples, from diverse Nazi soldiers to a misgendering ranked worse than thermonuclear war [1], and Zuckerberg's DeepSeek and Tiananmen case [3] were exposed because people knew the facts. Probing with historical or politically sensitive questions shows you how a model has been tuned.

4. Watch who can pull the lever, not just what it says. Andreessen's "ring of power" [2] and his question "who can get who fired" [2] are useful ways to look at any AI company: ask who sets its rules, who pressures it, and who it answers to.

5. Support mechanisms that add information rather than remove it. Zuckerberg's argument for Community Notes, where notes appear when people who usually disagree agree, is that you're showing more information, not less [3]. When you see a label or correction, prefer systems that show you the reasoning over ones that silently remove or demote.

6. Say what you think in public, especially where you suspect others agree. Andreessen's account of preference falsification is that change comes when one person says the emperor has no clothes and survives, then two, four, eight [2]. Hidden agreement only becomes visible when someone speaks.

7. Pay attention to government pressure on platforms and AI companies. Both Andreessen and Zuckerberg draw the line at government coercion, because the First Amendment binds the government [2][3]. Andreessen's remedy is political: "you have to elect people who don't do it," plus transparency and investigations [2].

8. Treat any single voice, including these ones, with some skepticism. Musk's summary itself advises weighing his claims as his perspective [1], and Andreessen's view is openly partisan [2]. The structural lesson (whoever tunes the model shapes the answer) applies just as much to models built by the people making this argument.

Sources

1. 2281 – Elon Musk · The Joe Rogan Experience
2. Marc Andreessen: Trump, Power, Tech, AI, Immigration & the Future of America · Lex Fridman Podcast
3. Mark Zuckerberg · The Joe Rogan Experience
4. Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · Lex Fridman Podcast