The evening of 30 October 1938
On Sunday 30 October 1938, at eight in the evening on the East Coast, CBS aired the weekly episode of the Mercury Theatre on the Air. Orson Welles, then 23, had adapted H. G. Wells's The War of the Worlds for radio: the Martian invasion was moved to New Jersey and told as a series of news bulletins breaking into a programme of dance music.1 That it was fiction was said four times during the broadcast, at the start, before and after the station break and at the end. During the night the network repeated it at 10.30, 11.30 and midnight for anyone who had not understood.2
The next morning the New York Times reported on its front page that panicking listeners had taken the drama for fact.3 In 1940 the Princeton psychologist Hadley Cantril summed up the episode at the start of the chapter of his study devoted to the panic, in two sentences that became the official version: "At least six million people heard the broadcast. At least a million of them were frightened or disturbed."2
Ever since, the story has come back every Halloween as proof of how easily a new medium can fool the crowd. This year the anniversary falls as fear of deepfakes and AI-generated video grows in Europe, and it is useful to look at how that panic was measured.
What the 1938 numbers say
The million frightened listeners do not come from any count made that night. They come from a survey by the American Institute of Public Opinion, George Gallup's institute, taken about six weeks later because there was no money to do it sooner. Twelve per cent of the adults interviewed said they had heard the programme. Cantril himself admits that this figure was "over 100 per cent higher" than any other measure: the ratings firm C. E. Hooper put the audience at about 4 million. Pooling the two, he arrived at 6 million, a figure he considered conservative. Of those who listened, 28% said they had believed it was a news bulletin, and 70% of those had been frightened or disturbed. Hence 1.7 million people who took the broadcast as real, and 1.2 million frightened.2
How many listened and how many believed it, according to Cantril
The measurements taken that night tell a different story. During the broadcast Hooper phoned 5,000 households for its national ratings and asked what they were listening to. Only 2% named a play or the Welles programme, and none mentioned a news bulletin.4 Welles was up against the most popular show of the week, ventriloquist Edgar Bergen with his dummy Charlie McCarthy, and in Hooper's weekly ratings the two programmes stood at 3.6 to 34.7.2
The evidence of panic does not hold up well either. The Princeton researchers checked a rumour that people had been treated for shock at a hospital in Newark and found it untrue. Six New York hospitals, asked six weeks later, had no record of any case linked to the broadcast, and no death has ever been conclusively attributed to it.4 The preface to Cantril's book also contains an admission that is rarely quoted: the names of frightened people printed in the newspapers led to barely half a dozen interviews. All 135 interviews in the study were carried out in New Jersey, and more than 100 of them with people chosen precisely because they were known to have been frightened.2
Why the newspapers talked of panic
During the Depression radio was becoming the news source Americans trusted most. In a Fortune poll quoted by Cantril, 50% thought radio news was less prejudiced, against 17% who chose newspapers.2 According to the media historians Jefferson Pooley and Michael Socolow, the press seized the chance to show readers and regulators that radio could not be trusted.4
The New York Times, in an editorial headlined "Terror by Radio", reproached radio officials for weaving blood-curdling fiction together with news flashes offered "in exactly the manner that real news would have been given". Editor & Publisher, the newspaper industry's trade journal, warned that the nation continued to face the danger of incomplete, misunderstood news from a medium that had yet to prove it was competent to do the news job. Ben Gross, radio editor of the Daily News, recalled in a 1954 memoir finding the streets of Manhattan deserted as his taxi sped to CBS at the end of the broadcast. His paper, meanwhile, had led with the panic.4 Cantril counted 12,500 newspaper items on the affair over the following three weeks.2
Cantril's study played its part too. It was funded in November 1938 with 3,000 dollars from the General Education Board, one of the Rockefeller foundations, and became a classic of media effects research. According to the Rockefeller archive, its sensational style helped inflate the idea of "mass hysteria" and overshadowed the work of two researchers on the team, Hazel Gaudet and Herta Herzog.5
The myth has lasted 88 years because it is useful. Jeffery Sconce, who studies media at Northwestern University, says we keep telling it because we need a cautionary tale about the power of media.6 W. Joseph Campbell, who devoted a chapter of his book on journalism's myths to the case, adds the third-person effect: we think the media sway other people a great deal and ourselves very little.1
From radio to deepfakes: the panic and the measurements
- Oct 1938CBS broadcasts The War of the Worlds. The next day newspapers report a panic
- Nov 1938A Rockefeller foundation gives 3,000 dollars for the Princeton study
- Dec 1938The Gallup survey, six weeks later: 12% say they listened
- Apr 2020Science Advances: fake news is 0.15% of time spent on media in the United States
- Oct 2023Three researchers argue that fears about generative AI and misinformation are overblown
- Nov 2024The Turing Institute finds no 2024 election measurably changed by AI-made disinformation
- Jun 202662% of people say they are worried about fake news online
- Oct 2026In Oregon a real video is described as a deepfake
Today's fear and the measurements
The Reuters Institute's Digital News Report 2026, published on 16 June and based on 97,520 interviews in 48 countries, finds that 62% of people are worried about fake news online, 4 points more than in 2025. In every Western European country concern rose by at least 4 points, and the report links this to worries about deepfakes, mass-produced AI slop and political misinformation. Trust in news fell to 37%, the lowest since it was first measured in 2015.7 In Italy it stands at 32%.8 I wrote about the same report in the piece on news that comes from chatbots.
The measurements of how much misinformation we actually see are more reassuring than our fears. A study published in Science Advances in April 2020, based on a representative sample of what Americans watch on phones, computers and television, found that news of any kind takes up at most 14.2% of the time spent on media, and fake news 0.15%.9 The data predate generative AI, but they give the order of magnitude.
How much news weighs in time spent on media
2024 was the year in which more than 2 billion people in at least 50 countries voted. The Alan Turing Institute's Centre for Emerging Technology and Security followed elections from the UK to the European Union, from Taiwan to India and the United States, and found no evidence that AI-enabled disinformation had measurably altered a result. The authors point out that for the US vote there is not enough data to measure the effect, and that false content still inflamed the debate, mostly reinforcing people who already thought that way. They also explain why they did the work: without evidence, the risk was that anxiety about the threat would grow and undermine trust in elections.10
Felix Simon, Sacha Altay and Hugo Mercier, in an article published in October 2023 in the Harvard Kennedy School Misinformation Review, reach a similar conclusion by another route. Producing more fakes is now cheaper, but what limits misinformation is demand, because people's attention is finite. Today's fears, they write, belong to an old and broad family of moral panics about new technologies. And alarmist warnings can reduce trust in accurate news too.11
When fear hits the wrong target
A case from October shows what this means. On 19 October the Oregon Journalism Project, a local outlet, reported that a video of Nafisa Fai, a Washington County commissioner running to chair the county, standing in front of an upside-down American flag at an event in Minneapolis, was a deepfake spread to paint her as un-American. Local politicians, including Congresswoman Maxine Dexter, condemned the video and called for rules on AI.12 Then images emerged showing that the flag really had been upside down, and the outlet withdrew the story, saying the headline and story had "incorrectly described the video as a deepfake".13 Fai said she would never intentionally stand in front of an upside-down flag.12
No software generated anything in this story. A real video was taken for a fake because the deepfake explanation was plausible and convenient. Researchers call the habit of dismissing real evidence as fake the "liar's dividend", and Simon, Altay and Mercier point out that it existed long before generative AI.11
None of this means deepfakes are harmless. The measured harm is real, and it falls mostly on individuals and companies. In 2024 the engineering firm Arup lost 200 million Hong Kong dollars in Hong Kong, sent in 15 payments by a member of staff after a video call with colleagues and senior managers who looked real and were all fake.14 I wrote about cloned voices used in scams and the tools for recognising generated content in SynthID and the rest.
What I think
Some people were frightened that night, and Cantril collected real accounts. But I think the story of a panic often says more about who tells it than about who supposedly lived it. In 1938 newspapers had a clear interest in portraying radio as dangerous. Today many people have an interest in portraying deepfakes as all-powerful, from those who sell systems to detect them to those who would rather blame a machine than one of their own choices.
The greater risk, as I see it, is not that we believe everything but that we stop believing anything. With trust in news at 37% and six people in ten worried about telling real from fake, an authentic video can be dismissed with a single word, as happened in Oregon. That is why what interests me most about watermarks and digital signatures on content is their ability to prove that something is real.
Concern should go where harm can be measured, meaning scams with cloned voices and faces and fake intimate images made to target a person. There the evidence exists, and so do the defences. On elections and public opinion, before raising the alarm, I would like to see the measurements.
The rule I would follow is simple. When someone writes that a piece of fake content fooled millions of people, they should say who measured it, when and on how many people. Cantril's book gave that information, and thanks to it we know today that the million frightened listeners were an estimate made six weeks later, starting from an audience estimate more than double any other measure.