Who decides what stays up: Zuckerberg, Wales and Hotz on pressure and control

The Short Version

Mark Zuckerberg says Meta pushes back on government takedown requests, gives some deference to democratically elected governments on what speech is acceptable in their countries, and draws its hardest line at handing over user data, because that can get people physically hurt [1]. Jimmy Wales says Wikipedia is "super hardcore" and has never given in to a government anywhere, and he explains why it can afford to be: volunteers write the articles, so there is nobody for an agency to lean on, and there are no advertisers to worry about [2]. George Hotz, who worked briefly at Twitter after Elon Musk bought it, says nothing about government requests; what he saw was "a monarchy deciding what to ban," and his remedy is to take the decision away from every central authority and give each person their own transparent filter [3]. So the three agree that outside pressure and central control corrode trust, and they disagree about where the decision should sit: with a company exercising judgment, with a community following published rules, or with the individual [1][2][3].

Table of contents

What They Said

Mark Zuckerberg: a company that has to make the call

Zuckerberg describes the content problem in layers, and government pressure is only one of them.

The base layer is the things almost nobody defends. Meta has spent five to seven years building AI systems to detect about 18 categories of harmful content, from terrorism and child exploitation to incitement and intellectual property violations [1]. Most bad actors, he says, do not grow more sophisticated. The exception is nation states running what Meta calls coordinated inauthentic behavior, which adapt constantly [1]. Here governments appear as the attacker. He separates organic narratives, which are how culture and art emerge, from undisclosed coordinated networks pretending to be organic, and he sets a practical goal: make the attacks expensive enough that the adversary goes somewhere else [1]. He is optimistic that more compute and stronger AI will keep defenders ahead [1].

The next layer is misinformation, where he draws the line at harms most people agree on, especially potential physical harm, and not at mere wrongness [1]. He is direct about what went wrong during the pandemic: early on, parts of the establishment waffled and asked for censorship of claims that in retrospect were more debatable or true, and that undermined trust [1]. Fridman presses the difficulty, noting that on the virus and the vaccines there was deep disagreement fuelled by politics about what is and is not harmful, with half the country seeing it differently from the other half [1]. Zuckerberg's answer is to prefer context over removal. He praises Twitter's Community Notes as a good implementation of that idea and says Meta now lets users turn off fact-checks that affect what they see, while content that breaks policy, such as incitement, is still removed [1]. Removal should be reserved for known categories that people generally agree are bad [1].

Then come governments acting as governments. Zuckerberg says Meta pushes back on takedown requests [1]. He allows a measure of deference when a democratically elected government has a different view of what speech should be acceptable in its own country [1]. He treats requests for user data as a different and more serious matter, the more critical thing to defend against governments, because handing over data can expose people to physical harm [1]. That concern shapes decisions years in advance: Meta refuses to place data centers in countries with poor track records [1]. It is also part of the weekly routine. The roughly 30 top leaders who meet every Monday can raise policy issues in a given country alongside product and technology questions [1].

Fridman asks about the Twitter Files and what Zuckerberg learned from the pressure US agencies appeared to put on Twitter [1]. Zuckerberg says it is hard to know from outside exactly what happened in each case [1]. Of Meta's own experience, he says agencies will tell the company about a threat they are aware of, and he characterizes it this way: "It's not really pressure as much as it is just you know, flagging something that our security systems should be on alert about" [1]. He concedes, "I get how some people could think of it as that," and then states the principle he operates on: "But at the end of the day, it's our call on how to handle that" [1]. He wants as much information as possible about what adversaries might be attempting [1]. When Fridman pushes on whether there would be consequences for ignoring suggestions from the CIA, the FBI or powerful politicians of either party, Zuckerberg's summary is that there is so much pressure from all sides that "you just have to do what you think is the right balance and accept that people are going to be upset" [1].

He adds a remark, made about Twitter but applicable to his own position, that running the company you see specific nuances that are not apparent externally, and that the discourse would be better with more room for acknowledging that [1]. That is the company's view in one sentence: trust us to weigh things you cannot see.

On a separate question, AI, he argues against concentrating control. He does not think the world is best served by a few organizations holding the most powerful technology, and he says unipolar imbalances have historically led to bad situations [1].

Jimmy Wales: a community that cannot be leaned on

Wales's answer to the same question is shorter. Wikipedia is "super hardcore" on government pressure and has never given in anywhere in the world [2]. He says the stance works because it is known: governments understand that casual threats will not produce quick concessions [2].

He is careful about what makes that possible. He acknowledges the harder case of companies with staff physically at risk in a given country [2]. Wikipedia's structure spares it some of that exposure. On claims that the FBI influenced content, he says the Wikimedia Foundation received no such requests, and that it could not have changed the articles anyway, since volunteers write them [2]. He has met officials, including a Chinese propaganda minister, and describes those conversations as him explaining how Wikipedia works, not taking instructions [2].

The second structural protection is money. Wales says avoiding advertising began as almost an aesthetic choice, but he now sees it as decisive, because organizations follow the money in the long run: "business model drives outcomes" [2]. Without ads there is no suspicion that a car company paid for its entry [2]. He resists corporate funding too, because influence creeps in [2]. The Wikimedia Foundation builds reserves and a separately governed endowment of more than 100 million dollars so the project can weather storms for a century [2]. When a reporter suggested Elon Musk buy Wikipedia, his three-word reply, "not for sale," became his most successful tweet [2].

Wales also describes what public panic about pressure looks like from inside. A viral claim held that Wikipedia tried to delete its article on the Twitter Files; in fact a single person suggested merging it into another article, and the proposal was quickly closed as having no chance of passing [2]. Another claim, that Wikipedia changed the definition of recession for political reasons, was one editor restructuring an article and moving a paragraph down, after which trolls arrived and the page was temporarily protected [2]. His advice: "Slow down a second and come and look" [2]. Outrage videos make better stories than routine editorial debate [2].

He is writing a book on trust, and his diagnosis of official pressure is close to Zuckerberg's. Many institutions lost trust deservedly, he says, because some officials "lost the trust of the public because they didn't trust the public": they overstated claims to scare people into compliance instead of communicating nuance [2]. He describes his own shift in the UK, from carefully following pandemic rules to making his own judgments once officials were caught breaking the rules they had set [2]. He points to Wikipedia's article on COVID masks as the alternative, a page that reflected genuinely mixed evidence after authorities changed their guidance [2].

He is not against rules for platforms. He proposes a moratorium on political advertising, since microtargeting lets campaigns tell tiny audiences contradictory things, and he thinks platforms should direct criminal allegations to the police instead of hosting them [2]. That last suggestion comes from experience. He twice reported grotesque false accusations of child abuse against himself on Twitter and was told they did not violate the rules, until contacts on the board intervened, which made him ask what happens to a bullied teenager with no such access [2].

George Hotz: no single decider, public or private

Hotz's time at Twitter was brief, and what he reports from it concerns ownership and engineering, not agencies. He deleted his first Twitter account in 2010 because the algorithm rewards outrage [3]. He took the internship because he believed in a principled commitment to free speech [3]. What emerged, he says, was "a monarchy deciding what to ban" in place of an oligarchy [3]. The decider changed from a group to one man; the fact of a central decider did not.

His alternative is individualized, transparent filtering in place of top-down censorship, for example a "no ad hominem" filter built on the hierarchy of argument that a user could choose to switch on [3]. The same idea appears in his hopes for AI: personal "firewalls" that protect your mind from manipulation, which he likens to an ad blocker for propaganda [3].

His reading of Musk's politics is about pressure of a different kind. The hate Musk has received from the left, Hotz says, is intense and has driven him to the right, and the result keeps alive a "fake like psi-op political divide" that leaves power with the ultra powerful [3]. He adds that hate pushes a powerful person toward the people who say yes [3]. That is a claim about how public pressure bends an owner, and through him a platform.

Where Hotz does talk about governments, it is as a force that seizes control points. His fear about AI chips is "how long before Nvidia is nationalized?" [3] He dislikes the cloud because of "who owns the off switch" [3]. His political principle is stated flatly: "there's things that centralize power and they're bad and there's things that decentralize power and they're good" [3]. He says he would respect a position that nobody should have powerful AI, but not one in which only trusted authorities do, and he sums up his preference as "I will always take anarchy over tyranny" [3].

He distrusts companies on the same grounds. "I'm not worried about alignment between AI company and their machines. I'm worried about alignment between me and AI company," he says, and he notes that a cloud AI will always be aligned with the company that runs it [3]. He criticizes OpenAI for refusing to disclose basic facts about GPT-4 and praises Zuckerberg for open-sourcing Meta's LLaMA model [3]. He does accept authority that is open about itself and that you can walk away from. He calls his own company a transparent dictatorship with free exit: "If you don't like the dictatorship, you quit" [3].

Where They Agree, and Where They Don't

Overreach by authorities destroyed trust. Zuckerberg and Wales tell nearly the same pandemic story from different seats: officials asked for more certainty, or more suppression, than the evidence supported, and the public noticed [1][2]. Hotz's objection to trusted authorities holding the keys comes from the same suspicion [3].

Context beats deletion. Zuckerberg praises Community Notes and lets users switch off fact-checks [1]. Wales says Community Notes has changed his mind about things he retweeted, and Wikipedia's whole method is to present the serious perspectives with sources [2]. Hotz wants filters that users choose [3].

Transparency is the minimum. Hotz demands it of companies and offers it as the saving feature of his own [3]. Wales half-jokes that he would trust newspapers more if they ran banners saying an article's neutrality is disputed [2]. Zuckerberg, by contrast, asks for some trust in what insiders can see and outsiders cannot [1].

They disagree on deference to governments. Zuckerberg gives some to elected governments on speech [1]. Wales gives none [2]. The reason is structural more than moral. Meta is a company with employees, offices and servers inside countries, and Wales himself names staff at physical risk as the hard case [1][2]. Wikipedia's content is written by volunteers the Foundation does not direct, so a demand to change an article has no one to land on [2].

They disagree on who should decide. For Zuckerberg it is the company: "it's our call" [1]. For Wales it is a community working to rules such as neutrality and verifiability, with the Foundation unable to overrule it [2]. For Hotz it should be the individual user, and he objects to a company's owner deciding just as he objects to a state doing so [3]. Zuckerberg moved a step in Hotz's direction by letting users turn off fact-checks, but he keeps a core of removals that no user can opt out of [1].

They read agency contact differently. Zuckerberg presents flags from agencies as useful threat information that Meta is free to act on or ignore [1]. Fridman is not convinced there is no implicit pressure when powerful agencies and politicians make suggestions [1]. Wales reports no such requests at all [2]. Hotz does not address it.

They differ on whether to trust the open-source turn. Hotz praises Zuckerberg for open-sourcing LLaMA and Zuckerberg argues against a few organizations holding powerful AI, so on AI they sound alike [1][3]. On the social platforms themselves Zuckerberg keeps central control, and says Meta does not open-source its core app code [1].

Why It Works

No single point of pressure. Wikipedia's resistance rests on the fact that the Foundation cannot edit on command; the people who could are dispersed volunteers [2]. Hotz generalizes the point: whoever holds the off switch can be pressured or replaced, so remove the switch [3].

A known policy deters requests. Wales's explanation for why refusal works is reputational. Governments that know threats will not produce quick concessions make fewer of them [2].

Money determines exposure. Wales's rule that business model drives outcomes explains both his refusal of ads and the century-scale reserves [2]. An organization with no advertisers and its own endowment has fewer parties who can squeeze it.

Decide before the request arrives. Zuckerberg's most concrete protection is physical: not building data centers in countries with poor records, so that the data is not there to be demanded [1]. He ranks the harms too, with data handed to a government above content taken down, because the first can lead directly to physical harm [1].

Raise the attacker's costs. Against state-run influence operations, Meta's aim is to make them expensive enough that adversaries move on [1].

Trust follows candour. Both Zuckerberg and Wales trace the loss of trust to overstated official claims, and Wales's remedy is to show mixed evidence as mixed [1][2].

Pressure from every side cancels out only if you accept being disliked. Zuckerberg's method is to choose a balance and expect anger from all directions [1]. Hotz gives the warning case: an owner worn down by attack from one side drifts toward the other [3].

These are accounts from people describing their own organizations, and they should be read that way. Zuckerberg and Wales give no figures on how many requests they receive or refuse, and Hotz's observations come from a short internship [1][2][3].

Where It Applies, and Where It Doesn't

Wales's model does not transfer to most companies. It depends on volunteer authorship and donor funding. He concedes that firms with staff in a country face a harder problem [2].

Zuckerberg's deference has a stated limit. It applies to democratically elected governments and to speech norms, not to user data, and not, by implication, to governments that are not elected [1].

Hotz's user-chosen filters address what you see, not everything a platform does. They do not obviously help with the cases Wales raises, such as false criminal allegations aimed at a person who cannot filter what others read about them [2][3]. Fridman also pushes back on Hotz's open approach where AI could lower the barrier to violence, and Hotz's reply is that anyone serious could already find the information [3].

Not every outcry about censorship is real. Wales's two examples show a merge suggestion and a moved paragraph turning into stories of suppression [2]. Zuckerberg's caution about judging the Twitter Files from outside points the same way [1].

The conversations do not settle what happened at Twitter. Zuckerberg declines to judge it, Wales addresses only the rumour about Wikipedia's article, and Hotz discusses Musk's control and the code base [1][2][3].

State actors are on both sides of the problem. A government can be the party demanding a takedown and also the party running fake networks. Zuckerberg wants threat information from agencies precisely because other states are attacking the platform [1].

What To Do With It

1. Ask who can actually change the content. On Wikipedia it is volunteers the Foundation cannot instruct [2]. On Meta it is the company, by its own account [1]. That tells you where pressure can be applied.

2. Follow the funding. Use Wales's test that business model drives outcomes: check whether a service depends on advertisers, corporate donors or its users [2].

3. Treat data requests as the more serious issue. Zuckerberg ranks government access to user data above content takedowns because of the risk of physical harm, so ask where a service stores your data and under whose jurisdiction [1].

4. Check the claim before sharing the outrage. Follow Wales's advice to slow down and look at the actual edit history or discussion [2].

5. Prefer and use context tools. Community Notes is endorsed by both Zuckerberg and Wales, and Meta lets you control fact-check settings [1][2].

6. Look for exit and openness. Hotz's standard for tolerable authority is that it states its values and lets you leave; apply it to platforms and to AI services [3].

7. If you run a platform, publish your line and hold it. Wales says a known refusal reduces the requests you get, and Zuckerberg's siting of data centers shows the value of deciding years ahead [1][2].

8. If you speak for an institution, state uncertainty as uncertainty. Both Zuckerberg and Wales say overstated claims and requests for suppression cost the establishment its credibility [1][2].

9. Watch what pressure does to you. Hotz's account of an owner pushed by hostility toward agreeable voices, and Zuckerberg's rule of accepting that people will be upset, are two responses to the same force [1][3].

Sources

1. Mark Zuckerberg on Open-Source AI, Llama, Moderation, Layoffs, Quest 3 and Jiu-Jitsu · Lex Fridman Podcast
2. Jimmy Wales on Building Wikipedia, Neutrality, Community Health and the Business of Trust · Lex Fridman Podcast
3. George Hotz on tiny corp, tinygrad, Open-Source AI Safety, Self-Driving, and Twitter · Lex Fridman Podcast