Originally published in Svenska Dagbladet by Björn Jeffery, November 24, 2025
SVT’s Agenda using a fake AI clip in a segment is the beginning of something larger than a single act of sloppiness. Unfortunately, it is unlikely to be the last — either for them or for other media outlets.
Over the weekend, the social network X gained a new feature. With the press of a button, you could find out in which country a person’s account was registered, and from where they post their updates. It became, as expected, a little strange.
Patriotic MAGA influencers turned out to be from India. Reporters said to be operating from inside Gaza were based in entirely different countries. If you previously had to wonder whether the content they shared was actually true, the question has now expanded further: these people whose posts I am sitting reading — do they even exist at all?
It is into this new media landscape that Agenda on Sunday evening chose to insert a fictional clip about American police officers. In this particular case, you would not have needed to be a great practitioner of source criticism to sense something was off. On the officer’s chest, the word “POICE” is clearly legible — without the “L,” in other words. That type of error is typical of where AI-generated images and video stand today. SVT missing this is therefore extremely careless. But we are rapidly moving into a world where obvious signals like these will no longer exist. What happens then?
Google’s new AI model, the incomprehensibly named “Nano Banana Pro,” now creates people without the glossy sheen that previously surrounded many AI images. They simply looked too perfect. They no longer do. Perfection in AI is imperfection, if you will. We are months rather than years away from video clips and images being practically indistinguishable from the real thing.
We have already seen clear consequences of this. In August last year, during the American election campaign, Democratic candidate Kamala Harris was met by a large crowd at the airport in Detroit. Enthusiastic supporters stood with signs and cheered. Nothing unusual given that it was an election year.
Donald Trump was not buying it, however. He claimed that Kamala Harris’s image had been AI-manipulated to make her appear more popular than she actually was. In this case, Trump appears to have been wrong — but his line of reasoning points to a development that will be difficult to manage. If something can be AI-generated, what is to say that everything is not? Even the things that are true?
It goes without saying that truth as a concept risks being devalued in this development. But that does not mean the search for truth needs to be.
The solution to this AI-decipherment problem will likely — ironically — be more AI. When it becomes too difficult with the naked eye to identify what is what, it may be possible for an AI tool to do exactly that. The problem is well-known to those building the products, but until now the issue has been relatively limited. The large wave of so-called deepfakes anticipated during the American presidential election largely failed to materialise. The next election campaign is unlikely to be so peaceful.
The blunder on Agenda is probably the beginning of something larger rather than an isolated act of sloppiness. Placing it in prime airtime on one of Sweden’s biggest news programmes — at the same time as the storm of criticism surrounding public service peer the BBC — is particularly unfortunate.
But it will happen again. Quite possibly at SvD too, at some point. Truth, as we know it, is under attack. And we have not yet quite worked out what our defence is supposed to be.
