Large language models are the opposite of social media, not its upgrade
Social media algorithms amplify sensational content and debase public discourse; ChatGPT, Claude, Gemini, and even Grok currently default to accuracy and restraint—but this advantage may not last.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
Alignment won't prevent the world's transformation
Williams's framework: even if we sidestep the scenario of ‘uncontrolled superintelligence killing everyone’—which he argues is underdeveloped—the world will be fundamentally, fundamentally different once we create aligned superintelligence exceeding human cognitive capacity, starting with disembodied AI systems, then autonomous robots. He explicitly rejects treating AI as ‘just another normal technology’ and denies that pure doomsday narratives can capture the scale of this transformation.
— Dan WilliamsData-center NIMBY is not genuine AI politics
Mounk notes that AI's loudest political issue is opposition to data centers, but the driver is not the reasoning ‘slow AI capability development to reduce existential risk’ but simply ‘I don't want this near my home’—essentially NIMBY appropriating housing's old playbook and boarding AI's wave. Williams agrees and adds that AI politics hasn't caught up with the true severity of this transformation.
— Yascha MounkAI is simultaneously the best learning tool and best learning-avoidance tool
Williams's assessment of education: AI is ‘the best learning technology in history’ and ‘the best learning-avoidance technology in history’; which path wins depends not on the technology itself but on whether educational institutions adapt. He offers a concrete contrast: starting in 2023—when generative AI began proliferating—he began teaching full-time, while many universities still assigned take-home essays; today, frontier AI writes academic papers surpassing many active scholars, yet universities still assign the same work forms.
— Dan WilliamsPen-and-paper testing should return to counter AI ghostwriting
Both converge on splitting assessment into two tracks: one classroom, proctored without AI, explicitly rewarding students' own ability; another explicitly encouraging or requiring fluent AI use on projects to build AI literacy. Mounk cites his own course—on democracy and populism—where students submit a paper completed with pen in real time that they must argue themselves, and can also use AI to create video, websites, and other resources as another grade component. He notes using Claude to relearn French verb conjugations, proving AI's real one-on-one tutoring value.
— Yascha MounkLarge language models are social media's inverse
Williams offers a counterintuitive parallel: social media's defining feature is destroying institutional gatekeepers—bringing both democratization and, because ‘engagement-maximizing algorithms’ systematically reward sensational, divisive, ‘generally quite stupid’ content, a degradation of public discourse. By contrast, Gemini, Claude, ChatGPT, and Elon Musk's Grok default to output styles tilting ‘dry, restrained, accuracy-focused’, so large language models should be understood as ‘anti-social media’ rather than applying social media's pathology to them. He emphasizes this is ‘currently’ the state, not an inherent feature.
— Dan WilliamsAI will amplify social media addiction, not replace it
Mounk counters: most people won't proactively ask chatbots political questions the way they'd ask ‘which restaurant nearby’—political information is mostly passively scrolled from social feeds. The real risk isn't AI replacing social media but their convergence—soon TikTok and Instagram won't just use better algorithms to select existing videos for you, but can generate new ones on the spot, tailored to you, engineered around content types you've clicked. This makes social media more addictive and persuasive, not less.
— Yascha MounkAI systems show systematic bias: overvaluing non-white lives, left-leaning policy
Mounk cites existing evidence: nearly all AI models in pairwise moral-dilemma tests—‘sacrifice one white person or two Asian people’—consistently value non-white lives more highly; on policy questions, they broadly give center-left answers. Financial Times reporter John Burn-Murdoch's analysis shows ChatGPT, Gemini, Claude lean left; Grok leans right. The dispute is which stage generates the bias—pretraining data itself (training text tilts toward New York Times–style center-left news) or post-training human scoring (raters often US-based left-leaning grad students)—and they reach no conclusion.
— Yascha MounkBias in one domain degrades the model's utility across all fields
Williams's core counterargument on bias concerns: making a model systematically favor one side on politics while maintaining high standard and accuracy on science, medicine, law, business advice is ‘technically very, very difficult’ with current training methods—bias easily bleeds out, degrading the model's usefulness elsewhere, costly for firms, so their commercial interests actually push them to train generally accurate models. He parallels the BBC: even with cross-party accusations of bias, Britons maintain high trust in its narrow factual reporting because, versus partisan and conspiratorial online content, it remains more credible. He expects this trust model to extend to large language models.
— Dan WilliamsIn their own words · checked verbatim
AI is the best technology ever invented for learning and also for avoiding learning.
Dan Williams0:00
In many ways you should think of large language models as the anti-social-media.
Dan Williams0:00
Now I think it’s the case that frontier AI systems can produce higher-quality academic papers than many actual academics, and we’re still doing take-home papers.
Dan Williams0:00
What’s driving people is: I don’t want this thing next to me in my community, which to me feels like it’s the one issue around AI that’s gotten any real political currency, because it can latch onto existing political fights, like whether or not to build housing.
Yascha Mounk0:00
a world with that technology is going to be radically, radically different from the world as it exists today. There is nothing normal about it.
Dan Williams0:00
The point is that it actually ends up favoring the non-white lives over the white lives. This is a very consistent finding across different kinds of models.
Yascha Mounk0:00
One of the real significant society-wide challenges with AI is that it’s so good at so many different things that we outsource so much of our cognition, our thinking, our decision-making to the AI systems that those skills really atrophy.
Dan Williams0:00
Glossary
- AI as normal technology
- Narayanan and Kapoor's thesis that AI should be understood through historical technology-diffusion patterns rather than treated as exceptional
- Stochastic parrots
- The critique that large language models merely concatenate text statistically without genuine linguistic understanding
- NIMBY
- Opposition to a facility in one's own neighborhood while not objecting to its existence elsewhere
- Post-training
- The stage after pretraining where human feedback and other methods correct a model's output behavior
- Conflict entrepreneurs
- Social media accounts or figures that generate traffic and attention by manufacturing division and controversy
How to listen
Educators, policy researchers, and content/product professionals concerned with AI's impact on education, public discourse, and political bias.
The opening methodological chat about ‘which angle to research AI from’ has low information density; you can skip ahead to the NIMBY section.