Fifteenth letter from Michael Mrak to Andreas Bruckmüller
Dear Andreas,
At the end of your fourteenth letter you described where our media are heading right now. I have left this implicit question lying longer than any other you have handed me, and I will tell you why right at the start. In the form in which you pose it, it can no longer be answered. Something has pushed its way between your question and my answer, and it has changed the question itself.
You wrote that the media had drunk the devil’s water, and that fear binds an audience faster than hope. That is still true, word for word. But the newsroom has long ceased to be the place where this is decided. The public sphere we are talking about is now sorted by machines, filled by machines and, most decisively, answered by machines.
The door people used to walk through
A finding first, as sober as I can give it. For almost two years now, search engines have delivered the answer themselves instead of pointing to it. The consequences have been measured, and they are not small: traffic from Google to news sites has collapsed by roughly a third year on year worldwide, and the media houses surveyed expect a further forty percent or more over the next three years, a fifth of them even more than seventy-five. Individual specialist newsrooms have lost nine out of ten of their visits without making a single mistake.
This is not a crisis of quality. It is the end of a business model. The click economy carried digital journalism for twenty-five years, badly, but it carried it. It is disappearing right now, and it is not disappearing in a way that better headlines could stop.
Whoever delivers the answer no longer has to name the source. And whoever is not named is not paid.
That is the first half of the problem. The second is more uncomfortable, because it concerns us: you, me and everyone who writes.
The asymmetry everything hangs on
The cost of producing a claim has fallen to zero. The cost of checking it has not. A text, an image, a voice, a video of a person who never said it that way: today they cost seconds and cents. Verification still costs a human being, time and usually a phone call. The entire rest of this letter lies between these two curves.
Producing claims has become free. Checking them has not.
In your letter you asked about bans and distrusted them in the same sentence. I understand that distrust better than I used to. No ban helps against an asymmetry of this magnitude, because a ban presupposes that you can find the violation. What helps is reversing the burden of proof. I will come to that.
The same pattern, one floor down
Before we talk about solutions, I have to widen the frame, because what becomes visible in the media is happening everywhere, only more slowly and more quietly. And here I want to resist the temptation to mistake the dramatic for the true, as I myself reproached you in the eleventh letter.
So far, the current data do not support the grand narrative of job loss. A study of more than twelve thousand European companies finds that the use of AI brings roughly four percent higher labour productivity and no decline in employment, at least in the short term. The European Central Bank reaches a similarly unexcited conclusion for the euro area. Anyone announcing the mass loss of work to you today knows more than can currently be measured.
And yet the same research shows one place where things do break: among the young, in entry-level positions, in exactly the tasks with which a working life used to begin. Researching, summarising, drafting, correcting, formulating something yourself for the first time and getting feedback on it. That was never the valuable work. It was the work through which people became valuable.
It is not work that disappears first. What disappears first is the apprenticeship.
I see three attitudes to this, and two of them are worthless. The first is ideological denial, which I gave a name to in the eleventh letter, here in the form of a superior shrug: it’s all just a parrot, all just statistics, it will pass. The second is exhausted denial: it can’t be stopped anyway, so why look. And the third, new one is rapture, the belief that the machine will answer the questions we have failed at, the climate included. All three end at the same point, namely with others deciding.
Four levers we hold in our hands
You accused me of rehearsing my capitulation. Well, then let me do the opposite and get concrete.
First: provenance instead of detection. We will not win the race to detect fakes; every detector is obsolete on the day it is released. The only race that can be won runs the other way: signed proof of where a piece of content comes from and who changed it, and how. The standard for this exists, it is called C2PA, cameras and editorial systems support it, and since 2 August 2026 Article 50 of the AI Act requires the labelling of artificially generated content, with the machine-readable form for existing systems required from December. Its weaknesses are known and I will not hide them: a screenshot is enough to break the chain, and many platforms strip the metadata on upload. The proof also only certifies the history of a piece of content, not its truth. It is necessary but not sufficient, and that is exactly why it needs its second half: a person or a house that is liable for the publication.
A machine can write any sentence. It cannot be held liable for a single one.
Second: money to the source instead of the click. If the click no longer pays, then what the machine cannot do must be paid for. This is not cultural pessimism, it is a market observation: being physically present in a place, obtaining a file, speaking with a person who otherwise does not speak. Three paths follow from this, and all of them are already being taken. The direct relationship with the audience, meaning subscription and membership instead of reach. Collective licensing towards the model providers, because individual newsrooms have no negotiating power there, while many together very much do. And public funding for what will never pay for itself, with a distance from politics that is non-negotiable. The meter you wrote about has its counterpart here, only a more uncomfortable one: on your heat pump, it measures what comes in. In journalism, it measures what would have happened without it and did not.
Third: infrastructure that belongs to us. On 3 September, three of the most important AI providers went down within ninety minutes, and a good part of Western knowledge work came to a standstill. This is the same monoculture we wrote about in our first four letters, only one layer higher and with a shorter reaction time. The answer to it is not a declaration of faith but construction: open weights where the data must not leave the building, European or locally operated models for regular operations, the best available model for the special case, and an architecture that allows switching at any time. You argued with Gladwell that every heat pump on the street lowers the threshold for the next house. Every self-operated model works exactly the same way.
Fourth: the law, looked at honestly. In the thirteenth letter I wrote to you that the lever was being kicked away by the very hands that should be holding it. That almost happened again here. The Digital Omnibus postponed the obligations for high-risk systems from 2 August 2026 to 2 December 2027, and for AI in regulated products even to 2028. Sixteen months, in a field that changes in sixteen weeks. But the difference from the combustion engine phase-out matters, and I have heard the two lumped together too often: here things were postponed, not scrapped. The prohibited practices have applied since February 2025, the obligations for general-purpose models since August 2025, the transparency obligations since August 2026, and the AI literacy obligation applies as well. A postponed lever is still a lever. Anyone who stops building now because they will only have to later will do the same thing under time pressure in December 2027, only worse.
The lever was not kicked away. It was postponed. That is a difference, and currently the most useful one we have.
What I have imposed on myself
In the eleventh letter I described three practices to you. Here are the three that have been added.
The first is the question of origin before the question of content. With a text, an image, a number, I no longer ask first whether it is true, but who stands behind it. This is not scepticism, it is economics: checking the content costs me hours, checking the accountability costs me a minute and rules out ninety percent.
The second is the tool instead of the oracle. I use these systems daily and intensively, but only where I can judge the result myself. Where I cannot, I do not use them, and where I have used them, I say so. Accountability remains entirely with me. This is not a moral add-on; it is the only mode of operation in which these tools can be used responsibly at all.
The third is the paid source. I pay for journalism I want to read instead of having it summarised for me. A summary of what no one researches anymore will soon no longer be a summary, but an invention.
The answer to your question
So where should our media go? Away from reach, towards liability. Away from the click, towards the relationship. Away from speed, towards presence. Whatever a machine can produce in a second is worthless from now on, no matter how well it is made. What remains is what someone vouches for with their name: having been in a place, having asked a question that was uncomfortable, and being answerable for it if it comes to that. That is less than we had. But it is harder than most of what we had.
And with that I hand the question back to you, and this time it is your discipline, not mine. You go into companies, make decisions for a limited time and then hand them back. If the entry-level work through which judgement used to develop disappears, where will the next generation of people capable of making decisions come from? For fourteen letters in this correspondence we have written about the old who no longer want to. I fear the more urgent question by now is whether we will still leave the young the years in which they could learn it.
Let’s keep going.
Michael
This text is an experiment: a public exchange of letters between Michael Mrak and myself, inspired by Epistulae morales ad Lucilium, in which reflections on sovereignty far beyond the digital context are developed further through dialogue.