Meanwhile, in Germany...

A few days ago, I was amused to read about what’s considered a scandal here in Germany: it turns out that the former publisher and editor-in-chief, and now columnist, of Berlin’s Tagesspiel, Carl-Andreas Casdorff, used “Artificial Intelligence” to write his opinion pieces.

Don’t laugh: the journalism profession still retains a certain dignity, at least for now.

In any case, upon discovering the matter, the Tagesspiel editorial board removed the offending article and other suspicious pieces from the website pending an investigation, while temporarily suspending the columnist, who has made the necessary mea culpa:

“I made a colossal mistake, damaging the reputation of the publication and my own. For this, I offer my most sincere apologies. In my articles, I used the term ‘Artificial Intelligence.’ I should have clarified this, and consequently, I should not have allowed the articles to be published.”

dw.com, June 21, 2026

Casdorff clearly embodies the Zeitgeist, the spirit of the times, because around the same time, another publication, the Frankfurt-based Frankfurter Allgemeine Zeitung (FAZ), removed from its website an op-ed by the prime minister of Thuringia, article which had also been written using so-called “Artificial Intelligence,” that is, a language model.

And since we Italians are a bunch of slackers while the Germans are such serious people, none other than Mathias Döpfner, the CEO of Axel Springer, weighed in on the matter.

This joker, in an effort to condemn the FAZ’s decision, could think of nothing better than to ask a language model himself for a polemical comment accusing the FAZ of rejecting modern technologies and comparing their decision to “a desperate attempt by the horse-and-carriage lobby to ban automobiles.”

I’m sure Herr Döpfner thinks he’s very clever and intelligent; after all, he is the CEO. But I think he’s just being a pretentious imbecile, and specifically, that the depth of his analysis on this technological issue is less than what you’d find in a high school student.

This so-called “rejection of new technologies” has been the employers’ argument ever since the invention of the steam loom, and after centuries, it’s still utter bullshit.

The point is that no one “rejects” new technologies, for the simple fact that the acceptance of a technology is a social choice, not an inevitability.

History is full of technologies that we as a society have simply chosen to discard.

Slavery, child labor, debt bondage like that in Downton Abbey, asbestos, chemical weapons, landmines, indiscriminate surveillance, certain drugs (thalidomide, anyone?), certain herbicides, certain genetic modifications in food, animals, and humans…

The notion that certain technologies are inevitable, the “this is the future; adapt or perish” argument, stems from the convergence of 1960s American libertarianism and capitalist libertarianism, what has been dubbed the “Californian ideology.”

For a quick illustration of what I’m saying, just recall that the very same arguments (and often the very same people), who today advocate for the inevitability of language models passed off as “Artificial Intelligence”, were used word for word just yesterday to claim that the Metaverse was inevitable. And the day before yesterday, they were used to promote blockchain.

Or, if you want to get fancy, you can dust off some vintage Keynes (John Maynard, 1930) and his prediction that the workweek in 2030 would be 15 hours.

Anyone who goes on and on about “the future will be this or that” just wants to sell something, if nothing else, themselves as a “futurist.” Which is really just being an astrologer with poor imagination, but it’s still better than working.

At this point, a researcher, Vera Katzenberger from the University of Leipzig, also enters the picture. She says that the Casdorff case is important because it undermines trust in journalism; the public reads newspapers for the experience or perspectives of certain authors, and if the opinion pieces are generated by “Artificial Intelligence,” it interferes with the way public opinion is formed; the public might feel deceived.

So far, I have no objections, but then Katzenberger digs in:

This is a problem because “Artificial Intelligence” has no values, no political positions, and no sense of responsibility.
(ibid.)

Now, come on. One out of three? I'm expecting more from a researcher.

As always, the problem lies in the language we use when discussing so-called “Artificial Intelligence.” For starters, it’s wrong and counterproductive to call it that: we’re actually talking about language models, that is, statistical engines for generating plausible-sounding texts.

Or, if we want to use my preferred terminology, we’re talking about bullshit generators. The fact that the texts are plausible doesn’t change the fact that they’re generated by rolling dice.

And then we insist on using anthropomorphic language, talking about language models as if we were talking about people.

A language model—which is a program—does not “have” characteristics in the same sense that a human being, or any living creature, does; that’s just the delusion of those who turn technology into a religion (and of advertisers, who tell you that the refrigerator, deodorant, or sedan they’re trying to sell has a “personality”).

At most, a language model may exhibit biases in the way content is generated, if the bias is present in the training data or is explicitly provided as an instruction (so-called “guardrails” are nothing more than preferential steering of the engine’s results, that is, bias).

A brief technical explanation. Language models are an application of techniques known as “machine learning”: a program is fed data, and the program “learns” (strictly between quotation marks), that is, it identifies recurring patterns within that data. Feed ten million photos of cats to a machine learning model, and the program can determine whether a new photo contains a cat.

Has the program understood what a cat is? Of course not. It only knows how colors and shapes are distributed, pixel-wise, in the photos of cats it has seen. Give it a new photo, and the program will say whether the photo contains a cat. Sometimes the answer will be correct.

Feed the program a few billion written pages, and the program reconstructs, from the examples it receives, the rules governing the construction of meaningful sentences.

Has the program learned to speak and respond? No. But it has analyzed enough questions and answers to be able to construct a response sentence when you give it a question sentence. Sometimes, the response sentence makes sense, and sometimes it’s even correct.

The program executes the same instructions, whether to provide an answer that we recognize as correct or one that we recognize as incorrect. There is no knowledge, no model of the world, and no constraint of reality within the program. The program generates sentences; it is the user who evaluates them against reality.

So yes, obviously the program has no sense of responsibility; the program only sees the correlations between the words of the language we speak, and a sense of responsibility is no more present there than it is in a coin toss or a roulette ball—which, if we accept the techbros’ reasoning, also “decide” things.

Another quick technical aside.

Machine learning works. But how well it works depends on the quality of the input data. The old adage “Garbage In, Garbage Out” applies to today’s language models just as much as it did to FORTRAN or COBOL programs sixty years ago.

Before creating DataKnightmare, I briefly thought I could work in Data Science. So I created my own methodology, which I called the “Eightfold Way to Data Science,” modeled after the Eightfold Path to Virtue in Buddhism.

The first three steps were:

1. Correct selection of sources, that is, where we collect the data
2. Correct collection—that is, which data we select from what’s available
3. Correct validation of the collected data, that is, verifying that the data is in the required format and has the values we expect. For example, a date is a date, but is it day-month-year, month-day-year, or year-month-day?

Now, my point was simple. Any idiot can just collect data. You need to know what the data is, how it was collected, and you need to check for any errors or biases in the collection process.

This is why I gave up on Data Science: I was talking about a discipline, but executives would say, “well, we have this data; let’s try and do something with it, and while we’re at it, let’s make sure that ‘something’ tells us what we want to hear.”

Because it’s easy to say that the company is data-driven, but if an executive says one thing and statistics say the opposite, how does that make the executive look?

I saw data as a tool for investigating reality and guiding decisions. Executives saw it as something to justify their decisions by cloaking them in an objectivity they didn’t possess.

Let’s call it like it is: almost all corporate archives are absolutely worthless but can be used to justify one thing or its opposite, simply because there’s no control over the quality of the data collected.

At this point, what can we say about language models as an application of machine learning? Their input is indiscriminately all text, of any kind, available on the Internet. And we know that on the Internet, there’s everything and its opposite. But not in equal measure.

There are detailed, precise, rigorous materials, the production of which required years of study and work by someone. And there’s utter nonsense, deliberately fabricated content, conspiracy theories, delusions, forums full of crackpots and neo-Nazis, things my cousin told me since he’s an expert on the subject, and so on and so forth.

The latter are vastly more widespread than the former, but the language model swallows it all without distinction and then averages it out. Even without being experts in data quality, what level of quality would you save can the result possibly have?

When I say that language models are the dumbest and crudest application of machine learning, this is what I mean.

The thing you turn to when seeking answers about your life, your health, or your work, the thing you call “Artificial Intelligence” because that’s what you’ve been told it is, is nothing more than the weighted average of all that, good and bad, flowers and shit, is found on the internet: blended, sweetened, colored to your liking, and served up; and you eat it as if it were a delicacy.

I’m not saying you’re stupid; I’m saying they’ve been taking you for a ride and want to keep doing it, for a fee. You might as well stop listening to them.

OK, sorry for the digression, but these things need to be understood properly; now let’s get back on track. The researcher from Leipzig tells us that the problem is that:

…“Artificial Intelligence” has no values, no political stances, and no sense of responsibility.
(ibid.)

We’ve realized that “Artificial Intelligence” is a misleading term, but even so, the only trace of a “sense of responsibility” in a language model can be, at most (I say “at most” because there’s also Grok), the so-called “guardrails,” those post-hoc instructions that are supposed to (the hypothetical conditional is a must) prevent the language model from explaining how to produce a chemical weapon or from spouting neo-Nazi rhetoric.

We know full well that “guardrails” only work in the minds of those trying to sell them, because they run counter to the inescapable fact that a statistical engine will function as a statistical engine even if we tell it not to. “Guardrails” are like writing “don’t roll this” on the sides one through five of a die and hoping that thanks to this it will always roll a six. Seriously, that’s what they’re selling you, that's what you are paying for.

And what about values and political positions? Those will reflect the input data, and therefore will strongly favor the values and positions that are most frequently repeated, regardless of them being right or wrong. Open Instagram or any social media platform and see the results for yourself.

As the icing on the cake, there’s always the possibility that the owner of the language model will add other “guardrails” to defend the values and political positions they prefer, or that suit their purposes. And, of course, they’re under no obligation to tell anyone about it.

There you have it. The way I see it, the problem isn’t that the language model lacks values or political stances. At the very least, the language model reflects the values and stances most prevalent online, and that alone is a problem. And perhaps it even gets a little extra help. To consider such a tool politically neutral is sheer madness.

And that’s not all. Because there’s already a study showing that even when used just for a first draft, the language model influences the language that will appear in the final version, in terms of style, vocabulary, and content.

A writer who lets the language model take the lead, even if only for brainstorming or to use it as one would a human listener, is agreeing to be led, slowly but surely, at best toward generally acceptable values and positions, and at worst, toward the values and positions favored by the owner of the language model.

Gaslighting as a Service; what a wonderful idea.

The language model never gets tired; it speaks and responds like a person, and we’ve evolved to listen to people, not to treat them like objects. So when the model presents you with an argument you might never have used yourself, you don’t reject it out of hand: you turn it over and over in your mind, tweak the style a bit, and maybe even come to accept it. In other words, you’ve decided what you think with a roll of the dice; maybe even loaded ones.

I don’t like stating the obvious, but there’s this great quote from Dune by Frank Herbert:

Once upon a time, men entrusted machines with the task of thinking for them in the hope that this would set them free. But this only allowed other men who owned the machines to enslave them.

Do something revolutionary: think for yourself.