Rethinking the Propaganda Model for the Age of AI

Rethinking the Propaganda Model for the Age of AI

In Manufacturing Consent (1988), Edward Herman and Noam Chomsky proposed that news doesn’t reach the public unfiltered. It passes through five filters before it arrives: who owns the outlet, who funds it through advertising, where its sources come from, who has the power to punish it for stepping out of line, and which ideological assumptions it absorbs without ever quite noticing them. Their model was built for an era of newspapers and broadcast networks. It is worth asking whether it still has something to say now that so many of us get our explanations of the world not from an editor but from a chatbot.

I think it does, provided the model is adapted rather than copied. An AI system’s answer, like a newspaper’s front page, is not a neutral transcription of reality. It is the output of a long chain of decisions made by people who never appear in the text: engineers who chose a training set, executives who approved a product roadmap, policy teams who decided what the system should refuse to say. The question a reader once asked of a newspaper, who owns this, has a direct descendant: who owns the model, who decided what it was trained on, and whose interests were in the room when its limits were drawn?

Ownership*is the most obvious carry-over, and also the sharpest. A handful of firms, OpenAI, Google, Meta, Amazon, Anthropic, and China’s DeepSeek, now control the capital and computing power needed to build a frontier model, a concentration tighter than anything even the most consolidated newspaper chain achieved. But the ownership structures underneath them differ in ways that matter. OpenAI operates inside a capped-profit arrangement bound tightly to Microsoft’s cloud infrastructure. Amazon and Google both hold large stakes in Anthropic, giving two of the biggest cloud providers on earth a direct financial interest in a supposedly independent lab. Meta funds its models through its advertising business and releases some of them as open weights, a different bet entirely from labs that keep their models closed. DeepSeek was built inside a Chinese hedge fund, High-Flyer, and operates under a state-shaped set of incentives and constraints that a Silicon Valley startup never faces. None of that appears in a single chat response, but it shapes what each company can afford to build, and what it can afford to say no to.

Ownership doesn’t stop at the model layer, either. Underneath all of these labs sits a narrower chokepoint still: Nvidia dominates the manufacturing of the advanced chips that train and run nearly every frontier model in the world, giving one company an outsized say over who gets access to the computing power AI development now runs on. That’s part of why US export controls on advanced Nvidia chips to China became a policy lever in the first place, and why Chinese labs like DeepSeek have pushed to develop their own chips and lean on Huawei’s Ascend processors instead. A press baron once needed a printing press. An AI lab needs a chip supply, and right now there is really only one company that makes the best ones.

Economic incentive stands in for Chomsky’s advertising filter, and here the frame is the Silicon Valley venture model itself. The capital is raised on the promise of exponential future returns, deployed into computing infrastructure at a scale no advertising-funded newsroom ever approached. Microsoft’s multi-billion-dollar commitment to OpenAI, Amazon’s and Google’s investments in Anthropic, and the broader race to lease or build data centers all function like Chomsky’s advertisers once did, except the “advertiser” here is effectively underwriting the product’s existence, not just its distribution. That money comes with expectations attached, about growth, about enterprise contracts, about which capabilities get prioritized. The useful question isn’t “what did the model say,” but “what funding structure made this particular answer the path of least resistance?”

Sourcing translates almost directly, and examples are easy to find. Much of the modern web-scraped internet, Common Crawl, Wikipedia, Reddit, has fed into most major models, and the fights over that fact have become their own news cycle: The New York Times suing OpenAI and Microsoft over the use of its articles, Reddit signing paid licensing deals with Google and OpenAI after initially resisting free scraping, publishers negotiating separate arrangements with different labs. Whoever a model licenses data from, or scrapes it from unlicensed, ends up shaping whose voice the model treats as authoritative. The same logic applies to religious and scholarly knowledge. Ask a model about Islamic jurisprudence, and its answer will lean on whichever scholars, schools, and centuries of scholarship happened to already be digitized, translated, and indexed, often meaning that classical, Arabic-language, or non-Western scholarship is thinner in the data than English-language commentary written about it. A tradition carried for a thousand years through chains of teacher-to-student transmission can be reduced, in an instant, to whatever version of it was easiest to scrape.

Institutional pressure is the update to Chomsky’s “flak,” and the clearest recent case is Anthropic’s own running dispute with the Trump administration. In February 2026, the White House ordered federal agencies to stop using Anthropic’s products after the company limited how the Pentagon could deploy its models in weapons systems, a fight over the ethical limits of military AI use that neither side treated as abstract. Months later, in June 2026, the administration invoked export-control authority to force Anthropic to suspend access to two of its newest models over a cybersecurity dispute, a decision Anthropic publicly disputed while still meeting with officials to resolve it. Regulatory pressure runs in the opposite direction in China, where Beijing has moved to tighten its own controls on AI exports and chip access even as firms like DeepSeek work around similar restrictions coming from Washington. The two governments are pushing their AI industries in different directions for different reasons, but in both cases, what a model can eventually do reflects a negotiation between a lab and a state, not a decision the lab made alone.

Embedded values replace the filter ideology, and they are, if anything, more explicit and more contestable. Every model encodes a working answer to genuinely disputed ethical questions: how cautious to be about a request that could be misused, whether to treat contested topics with strict neutrality or take a position, how to handle a question where cultures disagree about the answer. Labs that lean toward safety-first design accept slower rollouts and more refusals as the cost of avoiding harm. Labs that lean toward rapid deployment treat the same caution as an unnecessary brake on useful technology. Neither position is hidden, exactly, most labs publish their guidelines, but the underlying philosophical choice, closer to a cautious universalism or closer to permissive relativism, rarely gets debated with the same scrutiny as the products themselves.

Laid out together, ownership, economic incentive, knowledge sourcing, institutional pressure, and embedded values form something like an AI-era communications filter model. Its organizing claim is straightforward: an AI’s answer is never neutral. It is shaped, at every stage before the words appear on screen, by institutional and economic forces the user never sees.

Where the analogy has to bend is on interactivity. Chomsky’s filters described mass broadcast, one signal sent outward to an undifferentiated audience. AI is conversational and generative: it responds to the specific question asked, adjusts to a follow-up, and increasingly adapts to what it has inferred about the person asking. A newspaper’s front page is fixed the moment it goes to print. An AI’s answer is assembled on the spot, partly in response to the user’s own words. That makes personalization a genuine sixth axis, one with no real precedent in the original model, because the original model never had to account for an audience of one. The point of drawing out these parallels is not that a chatbot is simply a wire service with better graphics. It is that the question media theorists have long insisted on asking of the press, who controls what gets said, and why, has not gone away just because the messenger changed. It has only become harder to see.