What the Radio and TV Propaganda Panics Can Tell Us About AI

A few weeks back, Anthropic released a report highlighting how several groups have been using Claude to generate propaganda in support of state actors. That generated considerable alarm about how AI is revolutionizing propaganda, and worry that it opens the door to unprecedented new tools for Russia and China to manipulate publics. Daniel Kimmage and Darjan Vujica at New America argue that it’s worse than you think. This type of alarm isn’t new. My colleague Matt Chessen warned in 2017 about the MADCOM future, in which machine-driven communications tuned to individual psychology outpace anything human propagandists could manage. My colleagues at the State Department called out AI experiments from Russia and China in 2023 and 2024. The threat is real today, and alarm is warranted. But given the multitude of risks, the most useful question is what to be alarmed about.

This alarm looks very similar to the waves of panic that accompanied the rise of radio and broadcast television. Both are instructive: radio and television were used extensively for propaganda, often with great effect, and in at least two cases helped enable mass atrocity. But the predictions people made about how they would be used were usually wrong, largely because of faulty assumptions about how people receive and understand media. Misplaced alarm is not harmless, especially when we are at a time of scarce money, attention, and political will to defend against it. Thus, I want to chart out some of what the historic record suggests on where we should be focusing our attention.

The 1920s–30s

Radio arrived just after World War I had made propaganda a household word. Numerous books studied national propaganda techniques, including Harold Lasswell’s Propaganda Technique in the World War (1927), Walter Lippmann’s Public Opinion (1922), and Edward Bernays’ Propaganda (1928). Lasswell’s innovation was detailing the propaganda techniques used by governments and treating it as a core instrument of modern statecraft. Lippmann’s work introduced the idea of mental models – people act on the picture of the world they carry in their heads, so whoever shapes those pictures shapes politics. Bernays argued, approvingly, that “the conscious and intelligent manipulation of the organized habits and opinions of the masses is an important element in democratic society.” This mental model pervaded the interwar period and was later named the hypodermic needle: messages injected into a passive public that would react as designed.

Contemporary accounts validated this model, including some of the chief propagandists. Joseph Goebbels, Hitler’s propaganda minister, famously said that the Nazis could not have taken or held power without radio, and he subsidized cheap receivers to put radio in German homes. Benito Mussolini wired rural Italy to make it easier to push messages. While fascism certainly used radio, it wasn’t alone. Other governments did as well, including the United States. Franklin Roosevelt took to the radio in his fireside chats to shape public opinion.

The response

Americans were rightly concerned and sought both to rebut the propaganda and to study it. One of the first responses was something we would now call media literacy. The Institute for Propaganda Analysis, founded in 1937, taught the public (including in American schools) about seven commonly used propaganda devices — name-calling, glittering generalities, transfer, testimonial, plain folks, card stacking, bandwagon. While this effort wound down when the United States entered the war, the work on media literacy continued through to the present day.

The studies were perhaps more interesting and started the first deflation of the propaganda panic. Paul Lazarsfeld, Bernard Berelson and Hazel Gaudet followed voters in Erie County, Ohio through the 1940 campaign and published The People’s Choice in 1944. They found that media largely reinforced existing leanings rather than changing them. They also found that most political information reached the public through opinion leaders, an idea later codified in communication theory as the two-step flow. Other studies also found that propaganda had limited effects. Carl Hovland’s wartime studies of the Army’s Why We Fight films found the films conveyed facts effectively but had essentially no impact on soldiers’ motivation to fight.

Beyond World War II

After the war, both the government and the private sector continued to experiment, branching out into the new medium of television.

As an interesting aside, it was in the middle of this that Edward Hunter coined the term brainwashing, looking in particular at the Korean War. While it generated a great deal of alarm, the effects were also more modest. Robert Jay Lifton and Edgar Schein found that only 21 American prisoners out of thousands refused repatriation after exposure to brainwashing efforts in POW camps.

The commercial sector’s efforts were no better at injecting ideas into a passive public. James Vicary experimented in the 1950s with subliminal advertising, and his famous “eat popcorn, drink Coke” experiment caused so much alarm that it drew FCC action and a ban on broadcast use, as well as major popular coverage. It was only years later that he admitted that the experiment was more or less fabricated.

Theories and later research

When Joseph Klapper published The Effects of Mass Communication in 1960, the field had largely agreed on what became the minimal effects paradigm: the conclusion that mass media rarely changes minds and mostly reinforces what audiences already believe. But that theory didn’t survive long. As we entered the 1970s, new media researchers identified other things that media is good at. Maxwell McCombs and Donald Shaw coined agenda-setting in 1972, the idea that media may not tell people what to think, but it is fairly good at determining what they think about.

More recent studies give a nuanced view. David Yanagizawa-Drott studied how hilly terrain blocked the signal of Radio Télévision Libre des Mille Collines in Rwanda, estimating that roughly 10% of the participation in the Rwandan genocide was attributable to radio reception. A study by Maja Adena and colleagues looked at Nazi Germany and found that the effect changed based on circumstances. Weimar pro-government programming slowed Nazi growth, while Nazi programming after 1933 increased antisemitic expression in places it already existed and backfired elsewhere. The effects, though, were hardly universal or unidirectional. Propaganda can work, can be ineffective, or can backfire.

Lessons for an AI Era

So what does all of this mean for us when we consider AI?

First, the early waves of media panic fundamentally misunderstood the audience. People are not injected, and they are not passive. Prior belief, opinion leaders and social networks, and selective exposure all significantly change the effectiveness of propaganda. There is also an interesting bias in the panic and early research — people always believe that the masses are more susceptible to mass media than they themselves are. W. Phillips Davison called this the third-person effect in a 1983 piece, and it has gone on to be one of the most replicated effects in media research.

Second, audiences adapt. Marie Friestad and Peter Wright’s persuasion knowledge model (1994) describes how persuasion techniques lose their effectiveness once people recognize the technique. Just consider banner ads (remember those?): the first banner ad in 1994 drew a 44% click-through rate. By the end of the 1990s, the average was around 3%. Today, display ads sit near 0.1%. The ads did not get worse; the audience got wise to them. This adaptation can also be deliberately accelerated – suggesting media literacy and education remain essential.

These findings suggest that some of the panic is probably overblown. But there are lessons for our predicament today.

Monopolies matter. Media effects are most significant where alternatives are absent. In competitive media environments, the effects are far more modest. Both Germany and Rwanda show that the messages weren’t uniquely persuasive; the state had the capacity to be the only voice and to punish anyone who ignored it.

This is why media literacy and investing in new journalistic outlets matter. With the collapse of legacy media business models, the risks of media consolidation are high. AI is arriving when we are particularly vulnerable already, and it is likely to continue to heighten that vulnerability.

Coordination beats persuasion. RTLM in Rwanda wasn’t intended to persuade moderates; it was intended to coordinate action by marking targets, signaling official sanction, and making both common knowledge by doing so in public. The coordination effect is already a common, if unevenly adopted, feature of propaganda today.

Mass audiences aren’t always necessary for success. You do not necessarily need to convert everyone to achieve a goal, whether it is winning an election or organizing violence. Reinforcing existing beliefs and moving people from belief to acting is sufficient to have a considerable impact.

This is where I see some of the most concerning impacts of AI – not from mass propaganda, but from the ability of AI to help tailor unique messages, content, and pathways that mobilize people from belief to action, testing along the way. But even here, AI is accelerating a trend that started much earlier with automated A/B testing, and the big data sales experiments from Amazon and many other commercial retailers.

The Authenticity Problem

There is one development the radio and television comparison cannot reach. Broadcast technology added forms of evidence — recorded voice, then film — that were hard to fake and therefore presumed authentic. AI is moving us in the exact opposite direction.

Robert Chesney and Danielle Citron first named this problem in 2019: the liar’s dividend, in which the mere existence of convincing fakes lets the guilty wave off authentic evidence. Kimmage and Vujica document Iran dismissing recorded atrocities as AI fabrications in their article. This tactic has also been tested across five experiments with more than 15,000 Americans. Politicians who responded to scandals by declaring the evidence fake gained more support than if they apologized or stayed quiet.

A photograph or a recording carried its own authority, because it was hard to fake well, while proof was available to anyone with a camera. But now? Anyone can fake things, while proving they're real requires real provenance, often needing trusted institutions to vouch for it: courts, researchers, newsrooms with standards, or platforms who can certify a file’s history. That weakens activists, advocates, and outsiders.

Being Alarmed About the Right Things

Which brings me back to where we started. We are right to be alarmed, though our alarm shouldn't focus solely on AI. AI is an accelerant, but the largest issues are the weaknesses of our current global media environment. A clarifying question for future AI developments is whether they help to remove alternatives, enhance coordinated action, enable greater mobilization of people who already agree, or degrade the evidence we use to settle disputes. Those are the things that did the damage in the twentieth century.

I'd be less convinced by worries that AI content is getting rapidly more persuasive. History shows that after initial peaks, it falls quickly, something that current research agrees with. One of the largest recent studies on AI persuasion finds that the most persuasive AI techniques focus on facts and evidence, rather than the model size or its ability to personalize. But as a kicker, the techniques that made a model more persuasive also made it less accurate in the facts it used.

The thing I'm most convinced about? The Institute for Propaganda Analysis largely was on the right track back in 1937. The more we teach people the fundamentals of how to understand the techniques being used on them, the faster we'll inoculate ourselves.