This will be a long piece.
Since the Prime Minister mentioned hybrid influence these days, I found a good opportunity to write about one of its important components — the one that takes place in the information sphere. And since social media, especially Twitter, or X, account for a significant share of the information market in Montenegro and the Balkans, they have naturally become one of the most important battlefields of hybrid, information and, if you will, cognitive warfare.
With the rapid development of large language models — or, as we all affectionately call them, artificial intelligence — the possibilities for disinformation and manipulation in the information space are becoming immeasurably greater.
Of course, we have heard all this a thousand times before. Much less often do we get the chance to see what it actually looks like in practice.
Last week, it all started on Twitter with a post by Nenad Canak about a history textbook for primary schools in Serbia, in which Montenegrins, Bosniaks and Macedonians are treated as nations that emerged during the 20th century, while Slovenes, Serbs, Croats and some “Slavicized Bulgarians” are given the status of ancient nations, apparently dating back to before the beginning of time.
I don’t know who thought it was a good idea to call on Grok, the large language model developed by Elon Musk’s xAI, for help in the nationalist debate under Canak’s post. Grok had already been the subject of numerous controversies over responses containing extremist, racist and antisemitic content.
In this particular Balkan episode, it demonstrated something known in large language model research as AI sycophancy — ingratiating or obsequious behaviour, or the tendency of a model to adapt itself to the user’s assumptions, expectations and wording.
In practice, this meant that Grok largely catered to Serbian nationalist users, confirming, among other things, various formulations claiming that Montenegrins “were Serbs”, while providing very little context and almost no explanation of how fluid, layered and historically contingent identities can be.
In this way, a model can very easily cease to be a tool for informing people and become a tool for misinforming them — producing conclusions in which individual pieces of accurate information, stripped of their historical context and timeframe, create an inaccurate or deeply manipulative picture of the whole.
After seeing this, I decided to join in and conduct a small demonstration experiment. First and foremost, for myself, but also for others, to see just how dangerous it is to take everything LLMs serve us at face value.
Because, as indispensable as large language models are becoming for research, conceptualisation, writing, programming and countless other everyday tasks, they still require disciplined and responsible human guidance if we want reliable information.
On X last week, the opposite largely happened.
In a series of nationalist free-for-alls, dozens of users began playing with Grok as though they had found a digital arbiter who could finally settle our centuries-old disputes. By the end, we had even reached the point where Grok was justifying the use of the term “poturice” (*Turk-converts), while users were employing it to generate insults directed at me.
Yesterday, I conducted a small demonstration experiment to see what happens when a large language model is naively and spontaneously prompted and treated as a human interlocutor on some of the most sensitive Balkan and European questions of ethnic and national identity.
For the experiment, I used figures with almost totemic status in the Serbian national imagination, such as Novak Djokovic and Nikola Tesla, as well as identities that are often politically contested — Montenegrin, Ukrainian and Bosniak. I also tried Rudjer Boskovic, but I didn’t have much success there.
The exercise was successful above all because it mobilised a large number of nationalist activists who, through mocking, commenting and mass sharing, unwittingly became useful idiots in a small internet experiment.
The experiment showed how readily generative AI can reproduce dominant narratives, simplifications, stereotypes and disinformation when users do not ask for sources, fail to distinguish facts from interpretations, and pose suggestive or insufficiently precise questions.
An LLM is not a knowledge base. It is not a digital encyclopedia. Least of all is it some kind of neutral digital historian sitting above our national myths and deciding who is right. Its basic function is much more mundane: based on patterns learned from a vast amount of text, it probabilistically predicts which token — that is, part of a word or a word — should come next.
The well-known “stochastic parrot” metaphor is useful precisely because it distinguishes between two things people very easily conflate: a system’s ability to produce a grammatically flawless, convincing and contextually appropriate answer, and its ability to determine what is actually true.
The aforementioned sycophancy presents a particular problem — the model’s tendency to adapt to the user’s expectations. We could simplistically describe it as a kind of algorithmic confirmation bias.
If you ask it, “Why are Montenegrins actually Serbs?”, you have given it a completely different task from asking: “How did Montenegrin national identity develop between the 18th and 20th centuries, and what are the main historiographical interpretations of that process?”
With historical and identity-related questions, the problem is even more serious because the internet itself — from which models indirectly draw vast amounts of material — is by no means some neutral archive of human history. Nationalist movements, states, political parties, churches, media outlets, activist communities and diasporas have spent decades producing, copying and multiplying content that confirms their own interpretations of the past.
One relevant concept is data poisoning — the deliberate attempt to influence the data from which systems learn or which they use.
But the problem can potentially extend far beyond directly “poisoning” a specific training set. You can poison an entire information ecosystem. Mass production of texts, coordinated amplification of claims, SEO optimisation, networks of websites, pseudo-academic articles, false or ideologically biased encyclopedias, and thousands of pages citing one another can, over time, create the appearance of fact, authority and even consensus.
If you write something often enough and, in enough places, it does not necessarily become true. But it can become statistically highly visible. This is especially important given the asymmetry in numbers between the Serbian and Montenegrin populations, and the asymmetry in the volume of Serbian and Montenegrin literature and databases. That very asymmetry erases all boundaries on the internet. For systems that operate through patterns of language, this is a very important distinction.
With Grok, there is also the additional problem of limited transparency, as well as the broader question of corporate responsibility and the editorial philosophy of companies run by Elon Musk.
But let’s return to the exercise, because the most interesting part happened only later. Through repeated interactions, the result changed dramatically.
When I began demanding precise sources from the model, asking it to distinguish a primary historical source from a secondary interpretation, rejecting suggestive formulations and requiring it to explicitly state what it knew, what it was inferring, and where there was insufficient evidence, its response gradually shifted away from reproducing the dominant Serbian nationalist narrative towards a much more disciplined approach to the facts.
At one point, the process became almost comical.
From the initial:
“No, Nikola Tesla was a Serb and never identified himself as a Montenegrin”,
after several rounds of insisting on sources and on the meaning of the phrase “identify himself”, we arrived at an answer that essentially said:
“Yes, in a certain context, Tesla did identify himself as a Montenegrin.”
And now comes the most important point: it is completely irrelevant whether Tesla, in a single sentence, letter or interview, called himself a Montenegrin. We all know that Nikola Tesla was a Serb.
We could have conducted the entire exercise around a completely different topic, in a completely different society. We could have taken Ukrainian and Russian identity, Israel and Palestine, India and Pakistan, Northern Ireland, the American culture wars, or any other issue around which deeply polarised communities and enormous digital narratives exist.
The result would have been similar.
The point was not to prove that Tesla was a Montenegrin.
The point was to demonstrate that Grok can be both “Serb” and “Montenegrin” within the space of ten minutes, if you guide it persistently enough in one direction or the other.
And that matters because most people are not going to have a twenty-minute debate with a large language model about methodology, sources, primary documents, semantics and the limits of inference.
They will ask one question.
They will get one confident answer.
They will take a screenshot.
They will post it on Twitter.
And they will write: “There you have it, see what AI says.”
(Columnists’ opinions and views do not necessarily reflect those of the CdM editorial team)
