Amplification, in the book, is the leverage: the platform machinery that lets a single actor borrow the reach of millions (p. 80). When I planned this series, this slot went to a different variable and to the strangest case in the file, a fake that did its damage without ever existing. Read against the book, that is not an amplification case. The book files Gabon under the liar’s dividend (p. 158). I am keeping the case, because it is the most instructive one I have, and saying plainly where it belongs.
In October 2018, Ali Bongo, then president of Gabon, suffered a stroke while traveling in Saudi Arabia. He disappeared from public view for months, and his government said as little as possible, which is the soil in which every rumor grows. On December 31, the presidency released his customary New Year address. The man in the video looked wrong: stiff and nearly unblinking, his face asymmetric. Opposition figures declared it a deepfake, produced to conceal a president who was incapacitated or dead.
One week later, on January 7, 2019, a group of junior officers seized the national radio station and announced they were restoring democracy, citing the New Year address as evidence that the president was unfit to hold office. The coup collapsed within hours. The analysts who examined the video afterward found no forensic evidence that it had been fabricated. The costume of fakery had been draped over a recording nobody could show was fake, and the drape alone was enough to move soldiers.
No fake was required. The available idea of a fake was enough.
Bobby Chesney and Danielle Citron, in the 2019 law-review article that remains the foundational text on the subject, gave the mechanism its name: the liar’s dividend. Once convincing synthetic media is known to be possible, every authentic record becomes deniable, and the benefit of that doubt flows to whoever is served by it. The dividend does not require a fake to exist. It requires an audience that knows fakes exist. Gabon is the dividend paid out in its purest form: the mere availability of the deepfake hypothesis converted a real video into grounds for a coup.
The mechanism, in other words, cuts in both directions. The obvious direction is the fake wearing authority’s costume. Slovakia, September 2023, is the clean example: two days before the parliamentary vote, inside the legal silence period in which campaigns and media could not respond, synthetic audio circulated purporting to capture the leader of Progressive Slovakia and a journalist planning to buy votes. Experts judged it AI-generated. The careful postmortems, and Lluís de Nadal and Peter Jančárik’s is the one I trust, decline to claim it swung the election, and so do I. What the case establishes is design: the release was timed to the one window in which the costume could not be pulled off in time. The release was aimed at the denominator, which is next week’s subject.
The other direction is Gabon’s: the authentic record dismissed as costume. Of the two, this is the one I now watch more closely, because it scales with belief rather than with tooling. Every advance in generative media that makes the first attack cheaper makes the second attack more credible. You do not need to fabricate the tape if you can plausibly say the tape was fabricated.
And the tooling direction is no longer exotic. In May of this year, Singapore’s police released footage from a Zoom conference in which the country’s prime minister, its president, and a table of senior officials were impersonated by deepfake video, assembled to convince one business professional that the government needed his urgent funding. He transferred at least S$4.9 million. The costume of the state, rented by a fraud shop, at retail. When impersonating the state is a commodity crime, its political deployment stops being a question of capability and becomes a question of occasion.
So I am not scoring A from these cases. What they do show is the question the dividend forces, which is something other than is there a fake in this story. Who benefits from doubt about the artifact in either direction, and what would it cost to establish provenance inside the window that matters? Detection tools do not resolve this, because detection arrives after the window and is itself disbelieved on schedule. The exit, to the extent there is one, is affirmative provenance, marking the real rather than hunting the fake. That is Chapter 10’s territory, and week nineteen’s.
Next week: T and R, and why the denominator matters more than the numerator.
— J.W.B.
Note 11 of 24