Who decides what's true? Bias, censorship and moderation in AI and platforms

The Short Version

The fight over who decides what is true online did not end with social media; it has moved into AI models. Mark Zuckerberg describes Meta retreating from being "this sort of decider of what is true in the world" by replacing third-party fact-checkers with Community Notes and by making its automated filters demand higher confidence before removing anything [2]. The DeepSeek analysts Dylan Patel and Nathan Lambert show that AI models carry the same problem in a quieter form: values and censorship can sit inside the training data, the post-training, or a hidden layer around the model, and users may never see them [1]. Sam Altman's answer is to publish exactly how a model is supposed to behave so a failure is plainly either a bug or a policy to argue about, while taking a middle road on open source [3]; Yann LeCun goes further and says no AI can be unbiased, so the only answer is many open-source models, as diverse as a free press [4]. Put side by side, the four conversations show that every question platforms faced about moderation (who sets the rules, who pressures the rule-setters, how many honest voices get caught in the net) now applies to the AI assistants that will sit between people and information.

Table of contents

What They Said

Mark Zuckerberg: from fact-checkers to Community Notes, and from removal to precision

Zuckerberg frames the story as a journey back to where Facebook started. You only start a social media company, he says, 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" [2]. For the company's first decade, moderation meant practical problems such as bullying and people pirating copyrighted content [2]. What changed, in his telling, was two events: the 2016 election of President Trump, which coincided with Brexit, and COVID in 2020. In his words, those were the two events "where for the first time, we just face this massive, massive institutional pressure to basically start censoring content on ideological grounds" [2].

He is candid about his own role. After 2016, he says, "I deferred too much to the kind of critique of the media on what we should do," and adds that "since then, I think generally, trust in media has fallen off a cliff" [2]. His core worry from the start, he says, was the arbiter problem: "I was really worried from the beginning about basically becoming this sort of decider of what is true in the world. That's kind of a crazy position to be in for billions of people using your service" [2]. So Meta designed a fact-checking system meant to be narrow. Third-party fact-checkers would handle "things that are very clear hoaxes," and the company was explicitly "not parsing speech about whether something is slightly true or slightly false, like earth is flat, things like that" [2].

That limit did not hold. The program drifted toward political fact-checking, and people came to feel the checkers were biased, both in their rulings and in what they chose to check [2]. Comparing the drift to 1984, he says it really is a slippery slope, and that it destroyed trust, especially in the U.S. [2]. The mechanism matters: under the old system, fact-checkers could mark content false and Meta would then demote it in the algorithm [2]. A verdict from a small group became a reduction in reach.

COVID made the pressure more direct. Zuckerberg says he was at first sympathetic to deferring to health authorities during a genuine emergency, but the guidance kept changing [2]. The most extreme period, he says, came during the Biden administration's vaccine rollout: although he describes himself as generally pro-vaccine, officials pushed Meta to remove content saying vaccines might have side effects, even when it was true [2]. Citing documents turned over 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 taken down [2]. Meta's answer, he says, was that we're not going to take down things that are true [2]. After President Biden said platforms like Meta were killing people, agencies across the government began investigating the company, including the CFPB, even though Meta is not a bank [2].

On the law, Zuckerberg draws a sharp line between companies and the state. "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," he says. "But the first amendment does apply to the government. That's like the whole point" [2]. Having people in the administration "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," he says, and Joe Rogan's response is: "It sounds illegal" [2]. Zuckerberg says he cares less about punishment than about making sure it does not happen again, and believes Meta's new principles will last after eight to ten years of pressure-testing [2].

His replacement for fact-checking comes from X. "I think what Twitter and X have done with community notes, I think it's just a better program," he says. Instead of a small number of fact-checkers, "you get the whole community to weigh in when people usually disagree on something, tend to agree on how they're voting on a note. That's a good sign to the community that this is, there's actually a broad consensus on this and then you show it. And you're showing more information, not less" [2]. The key design change is the direction of the intervention: "you're not using the fact check as a signal to show less. You're using the community note to provide real context and show additional information" [2].

Yet he downplays how much this alone will change. "A very small percent of content is fact checked in the first place," he says, so whether it makes "the huge difference," he is "not sure, though he expects it to be a positive step [2]. The bigger fix, he argues, is in the automated classifiers: 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" [2]. He explains it with drug-dealing detection. Requiring 99% confidence might catch only 20% of bad content, while dropping to 90% might catch 60% [2]. But at 90% precision one removal in ten is a mistake, which at billions of users means millions of innocent posts removed and accounts disabled, sometimes across linked services like WhatsApp [2]. Meta will now demand higher confidence, accepting that some harmful content gets through in return for far fewer wrongful takedowns [2].

He is careful about what is not changing. Meta is not easing off on terrorism, child exploitation, fraud or foreign influence networks, which it finds with AI systems and teams of intelligence and counterterrorism specialists [2]. For coordinated groups doing election interference, he says, "We have like a bunch of people analyzing it. It's like they study it for a while. I think we're probably pretty accurate on that" [2], and such interference "is best rooted out at the level of kind of accounts doing phony things" [2], not by judging the truth of individual posts. The categories that got politicized, he says, were misinformation and hate speech. His example: until the update, Pete Hegseth's view that women should not serve in certain combat roles could not be expressed on Meta's platforms even though it could be said on the Senate floor [2].

Rogan adds the scale problem. "Moderating at scale is insane," he says; getting a reasonable, logical, objective human to analyze every video is virtually impossible, so "you've got to use a bunch of tools. You've got to get a bunch of things wrong" [2]. He also flags a way moderation itself gets gamed: "You could have mass reporting because you have bad actors" [2], including a corporation that decides to attack a video that is bad for it. Rogan defends anonymous accounts for whistleblowers and people who do not want their opinions to cost them their jobs, while acknowledging they enable bot networks [2].

Zuckerberg then carries the same logic into AI. He argues AI will not become a god-like power if it is widely distributed, which is why Meta champions open source [2]. He compares open AI to defensive tools like guns: you do not want only one party, or a foreign government, holding them, and he points out that China's DeepSeek model refuses to acknowledge Tiananmen Square [2]. His overall summary of the moderation question is that "the more voice is the answer on this" [2].

Dylan Patel and Nathan Lambert: hidden influence and censorship inside the model

Where Zuckerberg talks about what a platform removes, the DeepSeek analysts talk about what a model has already absorbed before anyone asks it anything. Lambert, a research scientist at the Allen Institute for AI, first clears up a common confusion about privacy: the model itself does not steal data. If you run open weights on your own computer, you control your data; the risk comes from whoever hosts the model, whether the DeepSeek app or an American company [1]. The worry about influence is a different and subtler one.

Patel raises it with an example that has nothing to do with China: American-trained models are quietly making British spelling obsolete [1]. Scale that up, and he asks: "What happens when the models are backdoors not just to computer systems, but to our minds?" [1] He cites Sam Altman's line that "superhuman persuasion will happen before superhuman intelligence" [1], while saying he does not believe DeepSeek has done anything like this [1]. Lambert adds that Anthropic research shows planted phrases in pre-training data can trigger hidden behaviors [1]. The concern, in other words, is that an open-weight model can be downloaded and inspected as numbers, yet what it has been shaped to believe or say may not be visible at all.

On censorship, Lambert gives the most useful map in any of the four conversations. He separates three layers [1]. The first is facts filtered out of the pre-training data, which he calls very hard, since the internet contains everything and wordplay slips through. The second is post-training, such as reinforcement learning from human feedback (RLHF). The third is system-level interventions that sit around the model. He gives an example of each kind of failure. Llama 2's overzealous safety training refused to explain how to kill a Python process [1], a post-training problem. Google's Gemini image fiasco came from a prompt-rewriting layer, not the model weights [1], a system-level problem. His point is that "censorship" in AI is not one thing, and where it lives decides who can see it and who can remove it.

Patel adds that bias can come in before anyone makes a choice at all. Internet text leans slightly left, he says, and even Grok ingests left-leaning subreddits, so bias from pre-training is hard to remove [1]. This echoes Lambert's broader claim, made when discussing how models are built, that "the data processing, data filtering, data quality is the number one determinant of the model quality" [1]: whoever curates the data shapes the model.

The pair also bring a sober view of open source. Lambert calls DeepSeek R1's permissive MIT license a major reset, the first truly frontier model that is open and commercially usable without restrictions, unlike Llama's license, which restricts some uses and requires derivatives to be named Llama [1]. Ai2's own definition of truly open source goes further and includes releasing training data and code [1]. But he warns that open-source AI lacks the feedback loops of open-source software, because building on it takes expensive compute and expertise, so for now it is largely an ideological mission [1]. Lex Fridman reads Zuckerberg's remark that there will be a global open-source standard and that it matters that it be American [1].

Lambert's motivation for working on openness at all is about trust: "i don't trust people that are like trust me bro we're gonna make a i good," and he wants more people involved in shaping the most powerful technology of his lifetime [1]. He contrasts DeepSeek's fast shipping with safety-focused labs, noting rumor that Anthropic has a model better than o3 that it won't release "Because chains of thought are scary" [1]. Patel's deepest fear points the same direction: he worries less about runaway AI than about a techno-fascist future in which a small number of people merged with AI through brain-computer interfaces rule everyone else, though he expects AI mostly to increase abundance and reduce suffering [1]. Fridman adds a personal note on influence, saying he feels more in control of his mind after breaks from the internet, and Patel notes that the adult content industry already uses chatbots to impersonate creators at scale [1].

Sam Altman: bias is hard, so write the rules down

Altman comes at the question as the head of the lab whose products everyone else is measured against. Asked about Google Gemini's controversial images, he does not gloat: he says these problems are hard and OpenAI has made its own mistakes [3]. His proposed fix is procedural. OpenAI should publish the desired behavior of a model with concrete examples, such as how it should answer who is better between Trump and Biden, so that a failure is clearly either a bug to fix or a policy to debate [3]. This is a transparency answer rather than a neutrality answer: it does not claim the model will be unbiased, only that people will be able to see what it was meant to do.

He also says OpenAI is less caught up in culture wars than other San Francisco companies because its shared ideology is a belief in AGI [3], and he expects AI to get caught up in left-versus-right politics, as anything of consequence does [3]. On how models are shaped, he credits ChatGPT's success as much to post-training such as RLHF and to product work as to the base model [3], which is the same layer Lambert identifies as a place where values and refusals get built in [1].

On open source he deliberately refuses to pick a side. "I think we should open source some stuff and not other stuff. It does become this religious battle line where nuance is hard to have but I think nuance is the right answer," he says [3]. He sees "definitely a place for open sourcing models. Particularly smaller models that people can run locally" [3], and expects an ecosystem with some open and some closed models: "It won't be unlike other ecosystems in that way" [3]. The kind of "open" he says he cares about is putting powerful tools in people's hands for free, with no ads on the free version [3]. He also points out that Grok only open-sourced after critics called Elon Musk hypocritical [3].

Advertising is another route by which answers can be bent, and Altman addresses it directly. He says "I kind of hate ads just as like an aesthetic choice," and likes that paying users know answers are not influenced by advertisers [3]. "If the question is, do I think we can have a great business that pays for our compute needs without ads, that I think the answer is yes," he says, while adding that he doesn't want to "completely throw out ads" and that "I have a bias against them" [3]. The contrast with Lambert, who says whoever figures out ads inside chatbot answers could unlock huge revenue [1], shows the commercial pull toward exactly the kind of hidden influence Patel worries about.

Behind all of this sits Altman's view on power. Asked whether he trusts himself with the power of building AGI, he answers: "it is important that I, nor any other one person, have total control over OpenAI or over AGI. And I think you want a robust governance system" [3]. He admits that although the board could legally fire him, "in practice, it didn't quite work. And that is its own kind of governance failure" [3]. He wants governments to set rules of the road and knows critics will call it regulatory capture: "I'm just willing to be misunderstood there" [3]. He also lists misinformation among Sora's release risks, noting "you can imagine like issues with deepfakes misinformation" [3], and he argues that dramatic, movie-style risks like AI escaping the box have pushed other significant risks out of the discussion [3].

Yann LeCun: no AI can be unbiased, so open source is the answer

LeCun, Meta's Chief AI Scientist, gives the most complete answer to the question in the title. Asked about criticism of Google's Gemini 1.5 for rewriting history in image generation and refusing to depict Tiananmen Square, he argues that an unbiased AI system is impossible, not mainly for technical reasons but because bias is in the eye of the beholder [4]. If neutrality cannot be achieved, the only defensible structure is pluralism, and he borrows it from what liberal democracies found for the press: it must be free and diverse [4].

The stakes, in his view, are higher than for social media. Soon every interaction with the digital world will be mediated by AI assistants, through things like Meta's Ray-Ban smart glasses, which can answer questions about a monument you are looking at or translate a menu [4]. Those systems will become the repository of all human knowledge, and he argues they cannot come from a handful of companies on the US West Coast [4]. Open-source base models let anyone fine-tune their own [4]. His examples are concrete: the French government's refusal to let three companies control its citizens' digital diet; an Infosys founder funding a Llama 2 fine-tune for all 22 official languages of India; and former FAIR scientist Moustapha Cissé's startup building models that speak local languages in Senegal to give people access to medical information [4].

He also offers a diagnosis of why closed systems behave the way they do. Big companies over-sanitize, he says, not because of their engineers' politics but because they cannot afford to offend their customers [4]. Meta can afford to give its base model away because it already has a huge user base and earns money through ads and business customers, like a pizza shop taking orders through WhatsApp; releasing the base model doesn't hurt, and Llama 2 has been downloaded millions of times with thousands of people contributing ideas [4]. As he and Fridman agree, "open source enables diversity," and basic guardrails that everyone agrees on can be built into objective-driven systems, with communities adding their own on top [4].

LeCun is not naive about misuse; he just thinks the defence is more AI, not less. If a rogue state builds a persuasion AI, it will have to get past your own AI assistant, which will filter it like spam: "my good AI against your bad AI" [4]. On bioweapons, he says studies increasingly suggest LLMs give no real advantage over a search engine, and the hard part is actually building such things [4]. He adds that AI breakthroughs, unlike nuclear weapons, spread quickly; AlphaGo was reproduced within three months [4]. His conclusion is that the real danger is concentrating AI in proprietary hands, not open access [4], and he suggests many doomers are doomers because they don't trust people or institutions [4]. He also offers a historical warning about who censorship protects: the Ottoman Empire banned printing in Arabic for about 200 years, and he says the UAE's AI minister, Omar Al Olama, told him it was partly to protect the powerful guild of calligraphers [4].

One more of his points bears on trust in what models say. On the Turing test he warns "we're fooled by their fluency": we assume fluent language means human intelligence, and that impression is false [4]. A fluent, confident assistant is persuasive whether or not it is right, which is part of why the question of who shapes it matters.

Where They Agree, and Where They Don't

Agreement: nobody should be the single arbiter

The strongest common thread is a refusal to let one party decide what is true. Zuckerberg calls being "this sort of decider of what is true in the world" "a crazy position to be in for billions of people" [2]. Altman says that neither he "nor any other one person" should "have total control over OpenAI or over AGI" [3]. LeCun says AI assistants cannot come from a handful of West Coast companies [4]. Lambert says "i don't trust people that are like trust me bro we're gonna make a i good" [1]. Each is describing the same fear from a different seat: the fact-checking program at a platform, the board at a lab, the market share of a few firms, the self-assurance of AI builders.

Agreement: bias enters at many layers, and some of it is unavoidable

Lambert's three layers (pre-training data, post-training, system-level interventions) [1] line up with what the others describe. Patel says left-leaning internet text makes pre-training bias hard to remove [1]; LeCun says bias is in the eye of the beholder, so no system can escape it [4]; Altman treats bias as a hard problem OpenAI has also got wrong [3]. None of them claims a neutral model is possible.

Agreement: the Gemini episode was a warning

Three of the four sources cite Google's Gemini image controversy, and each draws a different lesson. Lambert treats it as a technical lesson: it came from a prompt-rewriting layer, not the weights [1]. Altman treats it as a governance lesson: publish the intended behavior so failures can be judged [3]. LeCun treats it as a structural lesson: one company's model will always offend someone, so there must be many models [4]. Tiananmen Square appears twice too: Zuckerberg points out that DeepSeek won't acknowledge it [2], and LeCun cites Gemini's refusal to depict it [4]. Censorship is not unique to any one country's models.

Agreement: over-blocking is a real cost, not a safe default

Zuckerberg's precision math says a 90% filter wrongly removes millions of posts at Meta's scale [2]. Lambert's Llama 2 example, refusing to explain how to kill a Python process, is the AI version of the same error [1]; notably, Llama 2 is Meta's own model. LeCun says big companies over-sanitize because they cannot afford to offend customers [4]. All three treat excessive caution as a harm to honest users, not a free insurance policy.

Disagreement: how open should AI be?

This is the sharpest split. LeCun says open source is the answer to bias [4], and Zuckerberg says wide distribution keeps AI from becoming a god-like power, comparing it to defensive tools like guns [2]. Altman rejects the "religious battle line" and says "nuance is the right answer," with smaller, locally runnable models open and others closed [3]. Each position follows from where the speaker stands. Meta earns money from ads and business customers and loses nothing by releasing base models, as LeCun himself explains [4]. Altman runs a lab whose free product is paid for by its most capable models and cares about "open" as free access to powerful tools [3].

Lambert and Patel complicate the pro-open side from within. Lambert is an open-source advocate who builds fully open recipes, yet he concedes that open AI lacks the feedback loops of open-source software because building on it needs expensive compute and expertise, making it largely an ideological mission for now [1]. That cuts against LeCun's picture of countless communities fine-tuning their own models: some are doing it (India, Senegal, France [4]), but it is not cheap. And Patel's warning about models as "backdoors not just to computer systems, but to our minds" [1] applies to open weights too. DeepSeek is open-weight, yet Zuckerberg notes it still refuses to acknowledge Tiananmen [2]. Openness lets you swap models; it does not by itself tell you what a model was trained to avoid.

Disagreement: where does the pressure to censor come from?

Zuckerberg locates it outside the company: media critique after 2016 and government pressure during COVID [2]. LeCun locates it in commercial caution: companies that cannot afford to offend customers [4]. Patel and Lambert locate much of it in the data itself and in technical choices made during training [1]. Altman locates it in the culture wars, which he says will draw in anything of consequence [3]. These explanations are not mutually exclusive, but they lead to different fixes: legal limits on government pressure [2], more competitors [4], more transparent training [1], or published behavior specs [3].

Disagreement: transparency or plurality?

Altman's answer is to make one system legible: write down how it should behave [3]. LeCun's answer is to make many systems available: let everyone fine-tune their own [4]. Zuckerberg's Community Notes sits between the two: one platform, but truth decided by agreement among people who usually disagree, shown as added context rather than removal [2]. The sources do not settle which works better; each is a response to a different failure.

Why It Works

Why "more information, not less" lowers the temperature

Zuckerberg's case for Community Notes rests on two mechanisms [2]. First, the signal is harder to capture: a note is shown only when people who usually disagree tend to agree on it, which he calls "a good sign to the community that this is, there's actually a broad consensus" [2]. A small group of checkers could be seen as biased both in their verdicts and in which posts they chose to check [2]; a cross-partisan agreement rule is designed to avoid that. Second, the intervention adds rather than subtracts. The old system used a fact-check "as a signal to show less," demoting posts in the algorithm; the new one uses the note "to provide real context and show additional information" [2]. Adding context leaves the reader as the final judge, which is the opposite of being the "decider of what is true."

Why precision matters more than fact-checks

The arithmetic of scale drives the precision argument. Zuckerberg's numbers: at 99% confidence, a drug-dealing classifier might catch only 20% of bad content; at 90%, perhaps 60% [2]. But 90% precision means one wrongful removal in ten, and across billions of users that is millions of innocent posts and disabled accounts, sometimes across linked services [2]. With 3.2 billion people using a Meta service daily [2], small error rates become enormous absolute numbers. Since only "a very small percent of content is fact checked in the first place" [2], the classifiers, not the fact-checkers, are where most wrongful censorship happens.

Why AI censorship is harder to see

Lambert's three layers explain why the AI version of this debate is more opaque [1]. On a platform, a removed post is at least visibly gone. In a model, pre-training filtering is invisible and leaky (the internet contains everything, and wordplay slips through), post-training shapes what the model is willing to say, and system-level layers can rewrite a prompt before the model sees it, as with Gemini [1]. Planted phrases in pre-training data can trigger hidden behaviors, per the Anthropic research Lambert mentions [1]. And quiet drift, like the fading of British spelling [1], happens without anyone deciding it. That is why Patel frames models as potential "backdoors not just to computer systems, but to our minds" [1], and why fluency makes it worse: "we're fooled by their fluency" [4].

Why pluralism is the proposed cure

LeCun's reasoning runs in a chain [4]. Bias is in the eye of the beholder, so no single model can satisfy everyone. AI assistants will mediate every digital interaction and become the repository of human knowledge. A handful of companies controlling that repository is the same danger a state-controlled press would be. Open base models let others fine-tune for their language, culture and values. Shared basic guardrails can be built in while communities add their own. A malicious persuasion system then has to get past each user's own assistant. Altman's transparency mechanism works differently: by publishing intended behavior, a failure can no longer hide as ambiguous; it becomes either a bug to fix or a policy to debate [3].

Where It Applies, and Where It Doesn't

Where the "more speech" logic applies

It fits contested political and social claims, the two categories Zuckerberg says got politicized: misinformation and hate speech [2]. It fits cases where credible people disagree, like his example of a combat-role view that could be spoken on the Senate floor but not posted on Meta [2]. It fits situations where a rule-setter faces outside pressure, since spreading the judgment across a community or across many models makes any single point of pressure less effective [2][4]. It fits AI questions where values legitimately differ by culture and language, such as the local-language models in India and Senegal [4].

Where it does not apply

None of the sources extends the hands-off approach to clearly harmful categories. Zuckerberg says Meta is not easing off on terrorism, child exploitation, fraud or foreign influence networks [2], and coordinated election interference is handled by studying accounts doing phony things, not by community votes [2]. Higher precision thresholds also carry an explicit price: some harmful content slips through [2]. LeCun's open-source case includes basic guardrails everyone agrees on [4], and Altman keeps closed models alongside open ones and flags deepfakes and misinformation as reasons for caution before releasing Sora [3].

Limits of each fix

Community Notes covers only what gets noted, and Zuckerberg himself is "not sure fact-checking changes make the huge difference" [2]. Published behavior specs help only where the company is honest and the layer is visible; Lambert's point that hidden behaviors can be planted in pre-training [1] is a reminder that a spec describes intentions, not guarantees. Open source helps only those with the compute and expertise to use it, which Lambert says is why the movement is still largely ideological [1]. And openness does not remove built-in censorship: an open-weight model can still decline to discuss Tiananmen [2]. Finally, all of these conversations come from people with stakes in the outcome: Zuckerberg and LeCun both work at Meta, which gains from an open ecosystem it can monetize through ads [4]; Altman leads the company whose flagship models are closed [3]; Lambert builds open models [1]. Their incentives do not make them wrong, but readers should weigh them.

Where government fits

Zuckerberg's legal distinction applies specifically to the U.S.: the First Amendment binds the government, not companies [2]. Abroad the picture differs; he says the EU has fined tech companies more than $30 billion over roughly two decades, and that governments threatening to jail executives, as with Telegram's CEO, sets a troubling precedent [2]. Altman, by contrast, wants governments to set rules of the road for AI [3]. So government is, depending on the speaker and the context, either the source of illegitimate pressure or the proper setter of rules.

What To Do With It

1. Prefer context over removal when judging contested claims. Following Zuckerberg's reasoning, look for systems that show "more information, not less" and treat a note agreed by people who usually disagree as a stronger signal than a single verdict [2].

2. Ask where a model's refusal comes from. When an AI assistant declines or slants an answer, use Lambert's three layers: is it the training data, the post-training, or a system layer rewriting your prompt [1]? The Gemini fiasco came from the last of these [1], and the Llama 2 Python refusal from post-training [1].

3. Don't treat fluency as accuracy. LeCun's warning that "we're fooled by their fluency" [4] applies directly: a confident, well-written answer tells you nothing about its bias or truth.

4. Use more than one model on sensitive questions. If, as LeCun argues, no AI can be unbiased [4], comparing answers from different providers, including open-weight models, is the practical version of his free and diverse press. Remember that open weights can still carry built-in censorship, as with DeepSeek on Tiananmen [2].

5. Look for published behavior specs. Altman's standard is that a lab should publish how a model is meant to answer hard questions, with concrete examples [3]. Where a company does, you can tell whether a bad answer is a bug or a policy; where it doesn't, you can't.

6. Run open models locally when privacy matters. Lambert's point is that the model itself does not steal data; the host does [1], and Altman sees a clear place for smaller models people can run locally [3].

7. Watch for commercial influence on answers. Altman likes that paying users know answers are not influenced by advertisers [3], while Lambert notes the revenue waiting for whoever puts ads inside chatbot answers [1]. Check how an assistant is paid for.

8. Separate government pressure from company policy. Zuckerberg's distinction is useful: private moderation is an ethos, but government pressure to remove true content is a constitutional matter in the U.S. [2]. Treat demands from officials differently from a company's own rules.

9. Judge moderation by its error rate, not just its catch rate. Zuckerberg's precision numbers show that catching more bad content can mean wrongly removing millions of honest posts [2]. Ask any platform or AI provider both numbers.

10. Support broad participation in who shapes AI. Lambert wants more people involved [1], LeCun wants communities fine-tuning for their own languages [4], and Altman wants no one person in control [3]. Whether through open projects, public comment on behavior specs, or simply using diverse tools, spreading the power to shape these systems is the one step every source endorses.

Sources

1. DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters · Lex Fridman Podcast
2. Mark Zuckerberg · The Joe Rogan Experience
3. Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI · Lex Fridman Podcast
4. Yann LeCun: Meta AI, Open Source, Limits of LLMs, AGI & the Future of AI · Lex Fridman Podcast