We are dealing here with a politically sensitive subject, sticking to the mechanics without taking sides on any ballot or political party. The aim is to understand what technology changes in the circulation of information, not to comment on results.
For three years, every election has come with alarming predictions about generated content. The time has come to look at what actually happened, and the finding is more interesting than the predictions.
What did not happen
The feared scenario was a fake so convincing it would tip a ballot: a video of a candidate saying the unforgivable, released the day before the vote, impossible to debunk in time.
That scenario has occurred occasionally, but its effect has been more limited than anticipated, for three reasons.
Verification got organised. Newsrooms and platforms set up rapid-response systems. Massively circulated content attracts massive attention, including from those who will check it.
The most viral fakes are often crude. What circulates best is not the most credible content, but the content that most violently confirms what its audience already believes. And that content does not need to be well-made to work.
People rarely change their minds. This is the most important finding, and the best established by political science research long before AI: influence campaigns reinforce existing beliefs far more than they overturn them.
The real danger, and it is more insidious
The main problem is not that people believe fakes. It is that they no longer believe anything.
We developed this in our article on the end of the image as proof: when everything can be fake, everything can be denied. An authentic and compromising recording becomes contestable with a shrug.
This mechanism, which some researchers call the liar's dividend, structurally benefits whoever has something to hide. And it requires no technical effort: it is enough to claim the document is generated.
The second danger is exhaustion. Verifying costs time and attention. Faced with a volume of contradictory content, the rational reaction of a busy person is not to investigate, it is to switch off. A democracy does not die of deceived citizens; it weakens from weary ones.
AI has above all changed the cost of influence campaigns. Producing a thousand variants of a message, tailored to a thousand population segments, in ten languages, no longer requires a team but a few hours. It is not the quality of the fake that has changed, it is the ratio between effort invested and reach obtained. A modest actor can now run an operation that previously required state-level resources.
What protects
Three mechanisms work, and it is useful to know them.
Verifiable provenance. This is the approach we described with the C2PA standard and the California law. It does not detect fakes; it lets authentic content prove itself. It is the only path that is not a losing race from the start.
Institutional speed. A credible denial that arrives within two hours limits the damage. One that arrives within two days is useless.
Plurality of sources. A real event leaves multiple traces, with multiple witnesses, from multiple angles. A fake, even an excellent one, is usually alone. This is the cross-checking reflex we detailed in our verification method.
What to take away
The catastrophic predictions got the mechanism wrong, which does not mean there is no problem. The danger does not come from a decisive fake; it comes from the slow erosion of a presupposition on which all collective deliberation rests: the idea that establishable facts exist.
That presupposition is not defended by technology alone. It is defended by credible institutions, funded journalism, and citizens who accept the effort of checking what suits them. None of these three elements is guaranteed, and that is probably where we should look rather than at deepfake detectors.