Nathan Gardels is the editor-in-chief of Noema Magazine. He is also the co-founder of and a senior adviser to the Berggruen Institute.
Writing in Noema, Macario Schettino has pointed out that “each previous communication-driven disruption” — the printing press, mass newspapers, TV and radio — “was eventually tamed by the same technology that caused it.” So too will that be the case with social media linked to AI.
Even though we live in accelerated times, transmuting the obstacle into the way won’t happen overnight.
“The Reformation required 131 years to settle into the Peace of Westphalia. The Enlightenment needed longer,” Schettino reminds us. “The mass media transition produced two world wars and a Cold War before the postwar order stabilized. The social media transition is still very young.”
As in previous cycles, the experience of turmoil is the impetus for turning the technology into a platform for forging a governing consensus that transcends divisions.
He continues:
The printing press tore the medieval frame apart through the polemical pamphlet, and then, more slowly, built the new frame through the printed treatise and the encyclopedia. The press that broke Christendom is the press that produced the Enlightenment. The early newspaper of the 18th century fed the revolutionary mood that overthrew the ancien régime; the mature newspaper of the 19th century became the daily companion of the stable citizen-democracy of 1815 to 1914. The mass media of the interwar period — propaganda film, the demagogue’s radio voice — produced the manufactured emotion the totalitarianisms ran on; the mass media of the postwar period — broadcast television, regulated journalism, the evening news — produced the calmer national conversation the postwar order required. In each case, what initially fragmented the public sphere is what, in its maturer form, eventually rebuilt it.
If the pattern holds, the disruption caused by social media will not be resolved by abandoning social media. It will be resolved by social media’s next iteration, as shaped by artificial intelligence. We are still inside the chaotic phase, the phase that corresponds historically to the Wars of Religion or to the demagogue’s radio. The interaction of AI with social platforms could deepen the current chaos, but it is also the most plausible candidate to eventually tame the disorder. The new transcendent framework, whatever its content, will not arrive in spite of these technologies. It will arrive through them.
Infrastructure Already Exists
The challenge today, which goes beyond the communication disruptions of the past, is the unprecedented scale of connectivity and global scope of our densely wired societies supercharged by generative AI. That extant infrastructure, still evolving and unsettled, is both the territory of contestation and the ground for resolution. Since it is the general public whose information is being transformed by algorithms into the productive capacity of the future, reaching social consensus means they must have a say in its ultimate use.
As Hélène Landemore writes in Noema:
OpenAI reported 900 million weekly active ChatGPT users in February and crossed a billion monthly users in June. Google went from 400 million to more than 900 million monthly active Gemini users on a similar timeline. Meta AI reported a billion monthly users as well. No government, broadcaster or international body has ever had standing, two-way contact with such a huge portion of humanity simultaneously.
Using these companies’ best models to ask everyday people urgent questions about AI — What do you want this technology to do for you, if anything? What should it never be allowed to do? Who should govern this technology and according to what principles? — is not a thought experiment about hypothetical capacity. The infrastructure to ask hundreds of millions of people or more the same question in the same week already exists. …
Right now, the people best placed to actually facilitate a global deliberation on AI may not be elected leaders and heads of governments, but the heads of the companies building this technology.
Large language models alone, despite their scope, cannot serve as representative of humanity’s interests because of the limits of their training data.
Landemore notes a recent investigation by The Economist comparing the apparent values of 25 frontier AI models to those featured in the World Values Survey (WVS). The study found that “the models often hold values more extreme than the average respondent in the 88 countries surveyed by the WVS — in some cases more secular or more individualist than any society on Earth — and that they tend to compress the world’s moral diversity. If we ran humanity’s input through such models unchecked, the ‘global’ constitutional baseline would risk becoming that of one narrow cultural corner.”
As with AI processes in general, hallucinations, false impressions and eclipsed information in frontier models need what technologists call “reinforcement learning from human feedback” as a corrective. For the larger society, that means inviting public deliberation into what up to now has been the domain of the Big Tech elites and their algorithmic agents.
“The real issue is thus one of representativeness and accountability,” Landemore argues. For that reason, “we should embrace a mixed solution that would layer a multi-year deliberative process centering human arguments and justifications on top of LLM-mediated consultations that would provide a necessarily imperfect but quickly available and rich enough starting point for them. This mixed approach could be done at different speeds, in three tiers — each adding back a piece of what the fastest option alone would lack: an independent record to hold the process accountable, and a wider range of minds to catch what a narrower one would miss.”
Landemore explains her approach:
The first tier, an LLM-mediated global consultation, could produce an interim draft within months: a minimal, broadly endorsed statement of what a large fraction of the world population wants this technology to do — and never do. Run well, this tier would need not flatten its results into a single global average. It could also be broken out by region, so the consultation would surface areas where preferences genuinely diverge — say, between a Western emphasis on individual data privacy and precautionary restriction, and a data-sovereignty and development-first framing more common in Africa and India — rather than papering over real disagreement in pursuit of a tidy global consensus.
The second tier would be more logistically demanding and slower to organize, but still fast enough to fit the timeline at hand: an actual online global deliberation, with participants interacting with each other in real time rather than only with an AI, mediated by AI facilitators. This would produce a richer text than the first tier alone, testing and presumably sometimes replacing the recommendations that emerged from the LLM-aggregated input and making sure the LLM-provided justifications are enriched, corrected and checked by actual humans.
In effect, this is what Engaged California, an AI-assisted platform for public deliberation hosted by the state Office of Data and Innovation, is doing on precisely the topic of AI’s impact on work and society. Landemore continues:
The third tier would reach deeper still into humanity’s social fabric. We could hold local and national citizens’ assemblies on key issues in every country and region, both in person and online, ideally combined with referenda where possible, then feed the results into a concluding global citizens’ assembly that could ratify, amend or supersede the earlier floors as time allows. If the clock turns out to be as short as feared, we would not be left with nothing. If it turns out to be longer, the deeper, slower tiers would simply take over before the first floor ever hardens into something unexamined.
In her essay, Landemore appeals directly to the tech elite with the challenge of putting their compute where their oft-cited concerns are.
A Direct Appeal To Anthropic, OpenAI And Meta
Landemore goes on:
To Dario Amodei, CEO of Anthropic: Your own researchers already proved the first tier can work, and that it can scale. In 2023, Anthropic and the Collective Intelligence Project sourced constitutional principles from roughly 1,000 ordinary Americans, tested and refined a real model on those principles, and found that in some areas, this resulted in less biased outputs than those of the standard, internally authored model. This is evidence that public input doesn’t just confer legitimacy, it can make a model better. In 2025, Anthropic used an AI interviewer to ask nearly 81,000 Claude users across 159 countries and 70 languages in a single week how they use AI and what they hope to use it for, among other questions.
What’s missing isn’t proof of concept. It’s the will to use it. In January, Anthropic published a new “Constitution” for Claude, written the same way its old one was: by a small internal team, with input from the model itself and a handful of outside experts, but not from the tens of millions of people that model now talks to. Open the next revision to the people who live with its consequences and help bring about a global consultation on AI.
It is a good sign that Tino Cuéllar, Anthropic’s new chief global affairs officer, has said out of the gate that “democracies must set the terms on which this technology advances.” A former California Supreme Court justice, Cuéllar was a key figure in designing Engaged California.
To Sam Altman, CEO of OpenAI: I was an advisor to OpenAI’s 2023 Democratic Inputs to AI program, which funded 10 teams, selected from nearly 1,000 applicants, to prototype ways of soliciting public input on issues relating to the use of AI. This never amounted to anything close to what OpenAI’s reach could support. Finish what you started, at the scale this moment demands, with resources commensurate with what you yourself keep saying is at stake.
To Mark Zuckerberg, CEO of Meta: Your Community Forum on Cyberbullying in the Metaverse — run in 2022 with Stanford’s Deliberative Democracy Lab and the Behavioral Insights Team — was a preview of what the second tier could be. It gathered input from small-group online deliberations among nearly 6,500 people across 32 countries in 2022 on the issue of harassment occurring within your virtual reality platform. In 2023 and 2024, you ran two other Community Forums on the very issue of generative AI — soliciting people’s considered opinions on things like AI chatbots’ transparency with users, cultural sensitivities and data retention — gathering each time more than 1,000 participants from multiple countries in both the West and the Global South. You proved that public deliberation can work at a large scale. With Meta AI now reaching roughly a billion people a month, it’s time to redo the Community Forums on AI at that scale — not as a couple of one-off pilots measuring sentiment, but as a standing global institution that democracies have never had the tools to build.
Landemore closes her essay with a broader appeal to all those who are moving fast and breaking things, leaving the detritus of social fragmentation scattered around as fodder for reactionary politics that thrives on the dislocations of disruption:
“To every other tech leader in the U.S., China and beyond who is willing to put real resources behind this cause: Help bring democratically written global and local constitutions on AI into being, because right now no government is moving fast enough, or has the will or standing, to do it alone.”
