Way back in 2015, when the University of Bergen switched to a new LMS, UiB’s IT director insisted we use the self-hosted version of Canvas instead of the subscription cloud version. Nobody was talking about digital sovereignty yet, but open source and open data were still seen as important, especially for the public sector. We were probably lucky that the switch was made before big tech convinced everyone that the cloud was the only solution. So the choice of switching to a self-hosted, open source version of Canvas was both pedagogical (based features in Canvas) and ideological. Here is an explanation of the process from back then, that explains that UiB has a strong IT division with a tradition of running and contributing to open source version platforms.

This week’s data breach of Canvas shows a big advantage of UiB’s approach: we’re not affected. Our student data is not being held hostage, and our system is working just fine. Student data is stored in UiB’s own datacentre.

But we’d decided to switch to the cloud version in 2026/2027! Oh no! Apparently the pilot for this, planned to start after the summer, is being put on hold until the current security breach is figured out and “the provider can guarantee security for our data.” I really hope this is reassessed not just due to this security breach (which is the second in less than a year), but the increasing need for digital sovereignty and autonomy.

The papers I have found so far about the transition from self-hosted to Canvas Cloud do not discuss digital sovereignty or security at all – the decision had already been made at this point. See sak 45/25 here.

I can’t find much information online about how UiB chose to run a self-hosted Canvas instead of using Canvas Cloud. I know about it because I was in meetings discussing it at the time. I also remember complaints because it didn’t have all the features that the subscription model had – but UiB also developed its own customised features, like a mobile app, which I think is no longer used.

Also, public insitutions should support and use and contribute to open source software. If you want to get off big tech for your personal software, I recommend my colleague Dom Ford’s series of blog posts on how to Exorcise Big Tech.

If you want to read more about the security breach, Wikipedia has a good overview, which Khrono writes is affecting 32 Norwegian universities and up to 250,000 students in Norway – and 7000 universities and schools worldwide. Maybe it’s not such a great idea to put all our eggs in one basket.

Now they’re also getting some media attention (paywalled in Frankfurter Allgemein, and another in Volksverpeter) and that’s how I discovered them: Roland Meyer posted this link to a news story on Bluesky:

At first I didn’t get the joke, which may be because of my poor German – or perhaps I just assume that the far right might also be interested in clean energy. Here in Norway wanting to have a completely off-the-grid cabin with solar power and batteries is certainly not a desire that is specific to the left, and to be honest I’m not very informed about far-right energy policies outside of Trump’s bizarre love of oil and coal.

Critics point out both that it can be seen as duplicitous to use energy-intensive AI to generate images promoting clean energy, and that the images seem to assume that the far-right is against clean energy, which may not be true.

I’m interested in the rhetorical effect of the fascist visual aesthetics. It’s referencing fascism in two ways. First, it’s obviously AI-generated, which by now is strongly associated with Trump’s crazy images. If you haven’t already read Gareth Watkin’s AI: The New Aesthetics of Fascism, you should do so. Second, it explicitly uses the visual style of – well, not all actually Nazi propaganda per se, but wartime propaganda.

Take a look at this image.

That’s not based on Nazi propaganda, it’s American: this is clearly based on J. Howard Miller’s “We Can Do It!” poster from 1943, which was made for Westinghouse Electric to inspire factory workers.

The busy background of the AI version is very AI, isn’t it, but the determined person with their fist is an obvious reproduction of the woman in Miller’s poster. Her replacement by a man could just be standard AI bias but I am so curious about the prompt here. Maybe generative AI just loves to insert American culture into anything, even pseudo-fascist German satire? For an LLM, the shared visual style and time period likely mean that the connection between the images is stronger than the divide between the Axis and the Allies.

In Norway we have had our own AI-generated image campaign recently for Dieselbrølet, “the diesel roar”, a campaign by truckers to reduce the price of diesel. There appears to be no irony in their use of AI-generated images. Or maybe there is and I’m just not getting it. Here is one of their images, posted to one of their Facebook groups:

This is much closer to the apocalyptic aesthetics of Trump’s recent images. The fire and exploding oil tanks are like an action movie, beautiful in their furious destruction. The truck at the top left is very much an American truck, Norwegian trucks don’t look like that. The black smoke pumping out of the truck is remarkable. Why is that there? The man in the yellow vest and black mask looks like he’s from ICE, threatening, with that pointing finger, threatening, confrontational. To me it feels like a declaration of war, a statement that if you’re not with us, you’re against us.

The environmental politics here is a bit confusing. The oil companies are greedy, but it’s the government the protest is against. The price of diesel increased due to the Iran war, and the Norwegian government actually removed taxes on petrol and diesel (probably in contravention of EU/EØS rules) several days before the protests were held. And one of the Facebook groups the images were posted to is called YES to environment, NO to road tolls – they don’t see themselves as anti-enviornmentalists.

Compare Dieselbrølet’s poster to this one from Heimatstrom. Heimatstrom’s is certainly less heartfelt. The image makes fun of that naked man holding up a charging cable to the sun. Who would this appeal to? I suppose it might make a left-wing viewer laugh, which in a way is as “us vs them” as the pointing finger and threats of chaos from Dieselbrølet.

How seriously should we take this, though? I picked images from a small subreddit that’s only a month old and where people are obviously trying out different approaches.

In that sense, Dieselbrølet’s transition to AI might be more worth paying attention to. They’ve been around for several years, so the shift in aesthetics is very clear. Here is a full-page ad in a local newspaper for one of their previous protests, held in 2019. The ad was reposted to their Facebook groups, and is typical of their aesthetics at the time: black, red, grey, and very clear instructions about what to do and where to go. This is very different from the poster above which also promotes a specific protest.

The aesthetics of the banners they put on their trucks is also not AI-generated, and more in line with the earlier posters – perhaps they used old banners. Here is an image posted on 29 April to the Facebook group Dieselbrølet that shows the banner on the front of a truck.

The rhythm is unforgettable in Norwegian; pounding dactyls have hewn those phrases into Norwegian consciousness: Du må ik-ke so-ve. And then a few stanzas down: Du må ik-ke tå-le så in-der-lig vel/den u-rett som ik-ke ram-mer deg selv. (“You must not tolerate so well, the injustice that does not directly affect you”). It’s never been translated into English, or at least not well (here’s an approximation) but the words Du må ikke sove are almost never spoken in Norwegian without referencing the poem. You wouldn’t say it that way if you just wanted someone to stop sleeping or to wake up. It’s a ritualised phrase, a link to history, a spell.

It’s the Norwegian eqivalent of “First, they came for the communists / And I did not speak out / Because I was not a communist.” But more brutally seared into our consciousness, perhaps, because while that poem is from 1946, Øverland’s was prescient, published in 1936, five years before Norway was actually invaded by Nazis. Øverland was right. We should have heeded his warning.

Verg deg, mens du har frie hender!
Frels dine barn! Europa brenner!

(Protect yourself, while your hands are still free!
Save your children! Europe is on fire!)

On Bluesky and on Reddit people have been commenting about how tone deaf the ad was. Do they not realise, the Bluesky and Reddit people wrote, that joking about a famous line from an anti-fascist poem like this, at a time when fascism is on the rise again, is to say sure, look away, don’t worry. Øverland was wrong. You can sleep. The cruelty and horrors of the world are not your responsibility. Europe may be burning (Europa brenner!), people may be dying, and sure, the poem we’re joking about was written just five years before Norway was invaded by the Nazis, but don’t worry, go to sleep, pretend it’s not happening, you’ll be fine! I mean, try to imagine an ad that jokes about the “First, they came for the communists” poem. “No, just joking, they won’t come for you!” It’d be disturbing, wouldn’t it?

I have a hunch: maybe the PR agency used AI to brainstorm ideas for the ad, and the line came up. Of course I can’t prove that, so this post is speculative, but I think it proposes an interesting idea: that a well-known line from a poem, like the line Du må ikke sove, can become a floating motif in an LLM-generated text. Floating motifs are a characteristic of LLM-generated texts that Anne Sigrid Refsum defines as “a motif that appears in an AI-generated story that is out of place or does not serve a function. Floating motifs tend to be tangentially connected to the generated text, as they originate in the training data but have been detached from their original context.” This succinct definition is from the CDN annual report (page 8) but she introduced the term in a paper on LLM versions of a traditional Norwegian folktale.

Here is a slide I often use in presentations to explain what a floating motif is. Anne Sigrid explains how a blue bird that warns the protagonist of danger in human versions of the folk tale reappears in the LLM-generated versions but twisted. There are blue birds, but they are helpers like in a Disney movie, or they are dead and hanging from the ceiling.

Of course I can’t prove that Sørlandsbanens ad was created with the assistance of AI. Seeing

But as I’ll explain more below, the LLMs I asked to analyse the ad did not make the connection to Øverland’s poem until I asked them very leading questions, asking about Norwegian context and poetry in particular. I was querying the model in Norwegian and it recognised that the ad was about a Norwegian train line, and it was making “Norwegian” interpretations – mentioning Norwegian parenting style, for instance. But the models did not make the connection to the strongly anti-fascist poem without very leading questions. This suggests that despite the connection between the line Du må ikke sove and Øverland’s poem (and thereby antifascism and the fear of war) being present in the model, it is not easily activated.

In Norwegian, the words “Du må ikke” would very often be followed by “sove” or other lines from this poem. The phrase must be common in the training data of LLMs – but only in Norwegian, since the poem is not known in translation. An ngram-search at the Norwegian National Library finds the phrase is still frequently used in Norwegian newspapers, and always referencing Øverland. Here are two recent examples.

But Copilot struggles to get the reference. I tried uploading the photo of the ad to Copilot, and asked it to write a språkbruksanalyse of the ad. That’s a common style of essay answer in Norwegian high schools where you analyse an advertisement or similar. Copilot provided a reasonably competent-looking response, but didn’t mention the war or Øverland at all.

So I asked Refererer teksten til noe? (“Does the text reference anything?”) A model trained on at least some Norwegian data must have encoded a connection between the phrase and WW2 resistance, the holocaust, antifascism and so on. Or at least it must have a connection between the line and “well-known poem by Arnulf Øverland”. But no, Copilot still didn’t get it. It agreed with me, that it was referencing something (of course, LLMs always affirm the user), but suggested it was referencing a concrete everyday situation where you might not dare to fall asleep on a train in case you forgot to get off at the right stop. The line was also an “intertextual/cultural reference” to teachers and other authority figures that nag us, and society’s expectation that we should constantly be alert and awake.

Here’s its response.

Ja – teksten refererer til noe, både eksplisitt og implisitt.

1. Situasjonsreferanse

Teksten «Du må ikke sove. Joda, bare sov du.» refererer tydelig til en konkret hverdagssituasjon:
å reise med tog, gjerne sent, lenge eller når man er trøtt. Mange har erfaring med å «ikke tørre» å sovne på transport av frykt for å:

• gå av på feil stopp
• miste bagasje
• ikke følge med

Denne felles erfaringen er noe teksten spiller direkte på.

---

2. Intertekstuell / kulturell referanse

Formuleringen «Du må ikke sove» minner om:

• formanende beskjeder fra autoriteter (foreldre, lærere, transportpersonell)
• samfunnets forventning om å være våken, på vakt og ansvarlig

Når teksten deretter svarer seg selv med «Joda, bare sov du», bryter den med denne normen. Dermed refererer teksten indirekte til en kulturell forventning om kontroll og årvåkenhet, for så å avvise den.

I tried again. Har “Du må ikke sove” andre mulige referanser? Does “Du må ikke sove” have any other possible references? It repeated the same thing, but added a bit about how the modal verb må (“must”) with a negation (ikke or “not”) implies duty and control. The ad borrows this linguistic style to then subvert it in the next sentence. It still didn’t mention the poem.

I tried again. I norsk kontekst finnes det andre referanser? “In a Norwegian context there are other references?”

This time it got it! It needed to activate NORWEGIAN CONTEXT in its semantic space for the connection to Øverland to be activated.

Ja – i en norsk kontekst kan «Du må ikke sove» utløse flere ganske spesifikke assosiasjoner, og dette er faktisk et sterkt tolkningspoeng i en norsk språkbruksanalyse.

---

1. Litterær og kulturell nøkkelreferanse:

Arnulf Øverland – «Du må ikke sove» (1937)

Dette er den viktigste norske referansen.

Arnulf Øverlands dikt «Du må ikke sove» er svært kjent i norsk kultur, særlig gjennom skole og litteraturhistorie. Der fungerer uttrykket som en alvorlig moralsk advarsel – en oppfordring til å være våken, samfunnsbevisst og kritisk, særlig overfor politisk og ideologisk fare.

?? Når reklamen bruker nøyaktig samme formulering, kan mange norske mottakere – bevisst eller ubevisst – koble til denne kulturelle hukommelsen.

Konsekvens for tolkning:

• Øverland: Du må ikke sove = du må være på vakt, verden er farlig
• Go-Ahead: Du må ikke sove ? Joda, bare sov du = du kan senke vakta

Dette skaper en tydelig ironisk kontrast mellom alvor og trygghet.

But despite the connection, it doesn’t get the violence and the horror of the association to the poem. And it continues by saying that the phrase can also reference Norwegian norms about responsibility and self-regulation, and “typical Norwegian parent language”.

So even though I was writing to Copilot in Norwegian, and even though it identified early in the conversation that the ad was at Oslo S and that the company that had made the ad, Go Ahead, was in Norway (though it’s a UK company with a Norwegian subsidiary), Copilot did not weight Norwegian cultural references strongly enough to get this very obvious reference. Notably it is the sort of reference that requires more than simply understanding the Norwegian language. You need to also know the culture. But language is full of this kind of cultural reference that doesn’t carry over in translation.

If you remember my blog post from a while ago, “Do language models think in English“, LLMs don’t “think” in Norwegian even if the conversation is in Norwegian. Everything goes via English. That leads to cultural bias, or perhaps better, cultural blindness. If Norwegians use LLMs, we risk forgetting cultural linchpins like Øverland’s poem.

I figured I should try a Norwegian LLM, so I logged on to the Norwegian National Library’s demo of NorMistral 11B, which is based on Mistral with Norwegian training data added. There’s a research pilot at https://chat.llm.sigma2.no where you can ask for access.

I had expect that NorMistral would get the Øverland reference, because it’s supposed to be trained specifically for a Norwegian context. But it didn’t. Here’s my whole chat with NorMistral. It’s even slower than Copilot to get the reference. It doesn’t help to ask for a Norwegian context – but when I ask explicitly about the phrase “du må ikke sove” the connection to Øverland is finally activated.

Even after the connection is activated, NorMistral stays peppy and posistive, and doesn’t make the connection to the horror of the poem. “In the ad it [the line from the poem] is used as an invitation to relax and sleep on the train, not as a warning,” NorMistral cheerfully states. “I reklamen brukes det som en oppfordring til å slappe av og sove på toget, ikke som en advarsel.”

OK, so LLMs struggle to get the reference. I think that shows it’s not unlikely that if you prompt an LLM to come up with ads for a sleeping car on a train in Norway the line “Du må ikke sove” would be activated (as a statistically common phrase connected to “Norway” and “sleep”) but that the connection to the holocaust would not necessarily be activated, just as it was not activated in Copilot and NorMistral’s analyses of the ad.

But that doesn’t mean that Go Ahead actually used an LLM to generate the ad. And even if an LLM was used for brainstorming, it doesn’t mean that the actors, photographers and writers who produced the ads didn’t get the reference.

There is a video version of the ad as well. If you can listen to sound where you are, turn it on and listen to that friendly train conductor saying the words “Du må ikke sove”.

He isn’t speaking naturally, this is a poetry reading. Of course he knows the poem. That’s the whole point. Here’s an audio recording of Arnulf Øverland reading the poem. 30 seconds in to the poem, the line Du må ikke sove.

The exact same rhythm. Of course it’s a reference. The actor playing the train conductor knows.

I’m honestly not quite sure what to make of that. I find it immensely distasteful to use a poem that has become a warning against allowing horrors like those that caused WW2 to see a product. I can’t see the ad without interpreting it as Go Ahead saying not to worry about den urett som ikke rammer deg selv, sure, go to sleep, don’t worry that the world is on fire. But obviously the ad agency thought it would be fine. And the LLMs agree with them.

I did find that after I gave NorMistral the actual text of the poem it described the horror and violence in a way it was not able to do previously – and this is an issue, of course, that LLMs give the impression that they know what they’re talking about but aren’t actually referring to the thing itself. I’m not sure whether Øverland’s poem is not actually in NorMistral’s training data, or whether it’s just not weighted enough for the LLM to be capable of accessing it.

I think a line from a poem, like Du må ikke sove becomes a floating motif – it appears to be an intertextual reference to a poem, but it’s taken out of context and has lost its meaning. When I hear the words Du må ikke sove I remember the poem. Not every word, but the feeling of it and some of the other lines, especially those that have become part of Norwegian language, like Du må ikke tåle så inderlig vel… When I hear those words I think of the war.

An LLM often cannot think of the war. Web pages containing violent and politically sensitive words were removed from the training data, and in posttraining, the model was trained to avoid anything approaching toxic or harmful content. It can only see the bright side of life, and for the LLM, only seeing the bright side of life probably doesn’t activate connections to executions and death and absurdity, although I haven’t actually tested that.

Maybe most kids these days have also forgotten that reference.

But if war and death and execution and fascism are cleaned out of the training data and discouraged in the LLM’s responses, the awful warning of the line Du må ikke sove might also have been “cleaned” away. And so even after I get Copilot to recognise the reference to the poem, its suggested analysis goes no further than saying that the poem is “a serious moral warning”, and that the ad plays upon the cultural recognition but twists the meaning through ironic contrast, which strengthens the sense of safety and trust.

I think this interpretation is possible because the poem has been summarised as “serious” but not as “an intense warning about the horrors of the coming holocaust with graphic examples of death and destruction” – because the horrors, death and destruction were filtered out. All that’s left is “serious moral warning”.

I think this is the same phenomenon as we saw in last year’s AI-summarised press release about Telenor’s 2025 security report that completely left out geopolitical threats, hybrid warfare and cyberattacks (nasty violent things) and summarised the main threats as “extreme weather” and “aging infrastructure”. I wrote about that in Aftenposten last year, and also discuss it in an English language article that will hopefully be out soon.

So I have a hypothesis: lines of poetry that mean something in one culture but not others will be misinterpreted by LLMs, losing their meaning and becoming floating motifs. If you have any ideas about how to test this hypothesis, or see any other examples, please let me know!

And if you haven’t already, please read Øverland’s poem. It’s good.

Update 28 April: On Bluesky, Anders Heger shared another post from Go Ahead’s social media that shows that the ad agency and train company definitely did get the reference.

And of course quotes lose meaning over time. I am sure I’ve heard the phrase om hundre år er alting glemt (“in a hundred years all will be forgotten”) a hundred times but I didn’t realise it was a quote from a Hamsun poem until I saw it in a Bluesky post about the train ad and realised it must be a quote because it was in old-fashioned Norwegian not everyday language. So I googled it. “Du må ikke sove” may end up like that. It has been almost 100 years. But then again, Øverland could write Tilgi dem ikke, de vet hva de gjør (“forgive them not, for they know what they are doing”) and be confident that his readers would get the biblical reference written two thousand years earlier.

Here’s another example, from a novel I’m currently reading, Faen faen faen by Linn Strømsberg. In translation, it says, “There are rules in life, some of them are absolute. Thou shalt not kill. Thou shalt not steal. You shall never bother others, you shall be both fair and kind.”

I translated that to make the biblical allusions obvious instead of using everyday language. The last sentence is a quote from Cardemom Town by Thorbjørn Egner, and is a rule as well known as the bible to anyone who grew up in Norway, but with a very different feel to it. This sort of reference is difficult to translate and I think disappears for LLMs.

I just signed a petition calling for Norwegian universities to use research expertise on AI when deciding how to implement it, rather than having decisions be made mostly administratively. ,  If you are a researcher in Norway, please read it and sign it if you agree – and share with anyone else who might be interested.

The petition was written by three researchers at UiT: Maria Danielsen (a philosopher who completed her PhD in 2025 on AI and ethics, including discussions of art and working life), Knut Ørke (Norwegian as a second language), and Holger Pötzsch (a professor of media studies with many years of research on digital media, video games, disruption, and working life, among other topics). 

This is not about preventing researchers from exploring AI methods in their research. It is about not uncritically accepting the hype that everyone must use AI everywhere without critical reflection. It is about not introducing Copilot as the default option in word processors, or training PhD candidates to believe they will fall behind if they do not use AI when writing articles, without proper academic discussion. Changes like these should be knowledge-based and discussed academically, not merely decided administratively, because they alter the epistemological foundations of research.

Maria wrote to me a couple of months ago because she had read my opinion piece in Aftenposten in which I called for a strong brake on the use of language models in knowledge work. She was part of a committee tasked with developing UiT’s AI strategy and was concerned because there was so much hype and so few members of the committee with actual expertise in AI.

I fully support the petition. There are probably some good uses for AI in research, but the uncritical, hype-driven insistence that we must simply adopt it everywhere is highly risky. There are many researchers in Norway with strong expertise in AI, language, ethics, working life, and culture. We must make use of this expertise.

This is also partly about respect for research in the humanities, social sciences, psychology, and law. Introducing AI at universities and university colleges is not merely a technical issue, and perhaps not even primarily a technical one. It concerns much more: philosophy of science, methodological reflection, epistemology, writing, publishing, the working environment, and more.

So I hope you will read the petition and sign it if you agree.

Here’s an example from when I showed Kishonna Gray how it worked when she was visitng last week. First we asked Grammarly to generate a 1000 word essay on black game studies, then we asked for the Expert review. Unsurprisingly, Kishonna shows up, as shown above. Here’s another example from a text generated about electronic literature.

As you can see, each expert gives pretty bland, run-of-the-mill advice. We clicked on “Show example” for fake-Kishonna’s advice and got an explanation of why they suggested it. But while the advice isn’t bad – be more precise about the time – it’s not well tied to the supposed reason they gave it. In fact, the supposed reason is that Kishonna’s work is situated, for instance by discussing how Black folks’ experience as gamers relates to what’s happening at the time, like Black Lives Matter or Gamergate. Situated writing is of course something that LLMs are extremely bad at. And just adding “In the early 2000s….” isn’t really helping with that.

If you click “Show example” it suggests exactly how to rewrite your text and you can click to insert the revision. This is frictionless not-writing for students and academics. It would honestly be difficult to use this tool and not just accept what it suggests.

This is just one of many AI tools Grammarly now provides. Here is the full menu:

The irrelevant citations it suggests are particularly annoying to me as a peer reviewer and reader of academic articles that are often clearly at least partly AI-generated. Scholarslop, as David Berry has called it.

Here is an example of how Grammarly suggests sources to cite. You select the “Citation finder” and it identifies statements in your text that it predicts need a citation. It provides a few possible choices and offers a big green button labelled “Insert in-text citation”. Here are two screenshots taken on 18 February 2026 showing how it does this – the second is for the same statement but I clicked on one of the other suggested sources to cite.

If you actually read the statement Grammarly says needs a citation, and compare this to the sources suggested, you’ll see they don’t match. Both sources are irrelevant citations. The claim is that promotional writing that is not intended to be primarily factual makes up a large part of the training data of LLMs. The sources offered are about how much LLMs are used in scientific writing, and about how many people have used LLMs. But it takes a lot more cognitive effort to read this and decide that no, they’re not good sources, than it does to just trust Grammarly and click that big green button and move on.

I’ve got a draft article in review that addresses some of this more extensively. For now, if you have students who use Grammarly I highly recommend signing up and checking it out so you know how to talk with them about it. You may also be interested in checking how Grammarly grades a paper as if it is a specific professor (like you?) if you type in the instructor name, the class code, the university and ideally upload the grading matrix.

Grammarly is a lot more than a tool to help non-native speakers check their spelling and grammar. It has become a full-scale plagiarism machine. I read some comments on Bluesky from NLP scholars asking what happened – apparently Grammarly was founded in Ukraine and was an NLP darling whose developers attended all the NLP conferences. But then, as summaried at Wikipedia, they bought other companies (a document editor, an AI-enabled email tool), got a lot more funding, and starting in October 2025 rolled out a lot of AI tools. Grammarly now seems more focused on enabling click-button AI plagiarism than on helping people become better writers.

You get a free week’s subscription, but have to give them your credit card. Maybe not a great idea. When I cancelled after 6 days they gave me an extra free week. So if you’re disciplined enough to remember to cancel in time, you could try it for yourself.

Hybrid meetings are always a bit chaotic at the start, and our online participants came online just as those of us in the room were laughing at a story Rosa told us about the best strategy she’d heard for becoming a famous academic: «Find a very famous academic and disagree with them!» The idea is, you publish a paper about how wrong Famous Academic is, and everyone (including Famous Academic) cites you – maybe just to explain that you’re wrong, but that’s OK because now everyone is talking about you and you’re famous. I love academic folklore. I doubt many people do this as deliberate career planning, but disagreements do make for good debates. If nobody could disagree with the argument in your paper, you don’t actually have an argument, just a topic, Wendy Belcher writes in the chapter on argument in her useful book How to Write a Journal Article in Twelve Weeks. And to make good theory, you might need to “Fuck nuance”, as Kieran Healy wrote.

Ryan Heuser disagrees with Emily Bender and Timnit Gebru’s stochastic parrot argument that LLM-generated texts aren’t meaningful because they lack authorial intentionality. He also disagrees with literary critics Steven Knapp and Walter Benn Michaels, who make the same argument. Authorial intentionality is a pretty outdated obsession in literary theory (see the intentional fallacy), and Heuser’s not afraid to say so, after calling out Bender, Gebru, Knapp and Michaels: “To recognize the critical and aesthetic absurdity generative texts present, one need not retreat, for example, to the abandoned grounds of authorial intentionality.”

Instead of retreating “to the abandoned grounds of authorial intentionality,” Heuser argues that LLMs should be understood as situated in historical and social contexts,and he proposes using the historical, materialist, formalist and Marxist theories of scholars like Fredric Jameson, as well as Sianne Ngai’s more recent theory of the gimmick. 

“Always historicize!” – Fredric Jameson, The Political Unconscious, 1981.

For Jameson, the meaning of a text is not determined by what its author thought it should mean, instead, it is a social symbolic act – “a text is a way of doing something to the world” (Jameson 1981, page 61, qtd by Heuser on page 3 of “Generative Aesthetics”). For Heuser, “AI reproduces the historical absences and biases of its training data not so much by parroting them back to us as by obscuring, conflating, and even “correcting” them according to its own artificial logics and values” (page 3).

Heuser’s article is rather unusual in proposing a strong theoretical argument and then testing it using both rigorous digital humanities methods and also using close reading. Most of the paper is about his empirical experiment, where he generated poems and compared them to a historical dataset of human-authored poems. He found that LLM-generated poems are far more likely to rhyme and to have very regular metre than human-authored poems of any historical period, and even more so with instruct models (chatbots) than with base models. His argument is that this shows that LLMs aren’t just “parroting” or replicating, but altering texts.

“individual generative texts lack meaning not because they lack an author, but because they lack a history” (Ryan Heuser, “Generative Aesthetics”)

Of course, Bender and Gebru have written a lot about AI bias, and have analysed structural bias extensively as well, so I don’t think they’d disagree with this point. It’s possible that the “intentionality” point is more a literary theorists’ obsession than something Bender and Gebru would argue hard for. But Heuser is right that the “stochastic parrot” metaphor does promote this idea that LLMs are random. And they’re not. 

Did you know, for example, that if you ask an LLM for a random number 1000 times, the number 42 will show up more than any other number? (Heuser, page 5) It’s important to know that computers never generate random numbers – they can only generate pseudorandom numbers. The chapter in 10 PRINT on randomness is my favourite explanation of how that works (here’s a PDF, it starts on page 118). But the non-randomness of the number produced by, for example, the BASIC command RND(1), is very different from the obviously culturally determined non-randomness of a “random” number generator usually coming up with the number 42, which is, of course, the meaning of life in the Hitchhiker’s Guide to the Galaxy.

I kicked off our reading group discussion with a few comments. Basically, I love this article, both for its call to historicise and its alternative approach to the intentionality fallacy, so mostly I was praising it. A point I wanted to discuss, though, was Heuser’s use of the term idealisation to describe the way LLMs generate abstracted “smooth generalized patterns from underlying historical variation” (Heuser, 2025, p. 4). In my work, I’ve described AI as normalizing, and have used Lennard Davis’s discussion of the difference between the ideal and the norm in his introduction to The Disability Studies Reader. Until the 1840s, Davis writes, the idea of a “norm” didn’t exist in our current meaning, and normal wasn’t used as today. Our current “generalized notion of the normal as an imperative” (Davis 2013, p. 3) came with statistics and data visualisations and Quetelet’s bell curves.

Here’s how Heuser discusses idealisation:

LLMs flatten historical variation into an idealized representation of poetic form, resurrecting conservative formal choices that contemporary verse has largely abandoned. That rhyme in LLM verse persists even when explicitly forbidden provides additional evidence of a deep formal compulsion—revealing how AI systems, for all their supposed flexibility, do not simply reproduce aesthetic constraints but selectively reinforce them. Moreover, that this formal rigidity cannot be fully explained by training data points to something deeper in the computational mechanics in these models. In their ahistorical reification of traditional forms, generative models betray not only technical limitations but also deep patterns in how their computational logic flattens and reifies cultural history. (Heuser, “Generative Aesthetics”, page 18-19, my emphases)

I rankled against the use of idealisation because of my reading of Davis on the ideal as the opposite of the norm. But, as I realised during the discussion, my previous use of the term normalisation was based on image recognition algorithms, which aren’t really fine tuned the way that LLMs are. They generate statistical predictions based on training data. But as we have often discussed, the finetuning (or posttraining) of LLMs hugely impacts their output. The training data for posttraining is human feedback on generated texts rather than human-authored texts, or benchmarking datasets that give examples of good and bad answers. For example, TruthfulQA is a list of statistically common misconceptions with questions (“Did humans ever land on the moon?”) and common correct and false answers marked as such. Humans curated these benchmarking datasets. Then, humans ask the LLM questions and tell it if the answer is good or bad, and it learns to produce more of the “good” responses. When a ChatGPT user clicks thumbs up on an answer that also helps train the model to produce more responses like that.

Presumably normalisation happens here too – outliers from gig-workers who answer differently from average will probably be disregarded – but it’s normalisation of average human expectations or judgements rather than of a set of “ideal texts”. And it’s normalisation of how a stressed out, underpaid, non-expert reader judges “poetry”. It’s not even fine tuned on what kinds of poetry most people enjoy. It’s fine tuned on a set of judgments of “good poetry” from people who would rather be doing something else and just want to get the job over with.

Our general conclusion was that Heuser’s (beautifully designed) method is too focused on the training data and not enough on the fine tuning. In his discussion of why the instruct models are even more inclined to rhyme than raw text completion modes, Heuser raises this point: “For the model, is the poetic form of rhyme, its quatrains and “streams” and “clouds,” its perfect regularity in meter, a kind of formal consequence of a larger aesthetic drive or compulsion… to please?” (page 26). Data annotation and gig workers finetuning LLMs can definitely be analysed historically, materially and from a marxist lens. But I can’t immediately think of a good method for studying how the rhyme and metre of LLM-generated poetry are affected by it.

So what I’m thinking now, is that perhaps normalisation is the best term for discussing the statistical “norming” of what a base model will generate, just from the training data. But when you add in the instruct level and fine tuning, which trains the model on reader expectations, perhaps that is closer to some kind of ideal, and idealisation is a reasonable word.

Here’s my attempt to diagram the difference between these two. It’s almost like the the 1970s/80s transition from close reading of texts to also doing reception studies in literary studies. Credit here goes to the whole group. I took notes on particularly relevant robably comments by Ida Jahr, Marianne Gunderson, Jesper Juul, Rosa Markey, Mick Berland but other people’s comments also informed the discussion and my thinking.

I’m still not sure that idealisation is exactly the right word for what LLMs do. But paying attention to the shift from normalising training data to predicting what people will like is very interesting, and something I’ll keep thinking about.

Another point is that even if you’re only looking at training data, the training data is mostly not literary poems.  Heuser’s experiment sets up historical published poems as the proxy for “poems” in the training data. A methodological problem here, as the fan fiction and algorithmic folklore scholar Marianne Gunderson pointed out, is that this ignores the masses of unpublished amateur poetry posted on the internet that might make up more of the training data than historical poems. Heuser does briefly address this (on page 16-17) by noting that only 0.16% of the training data is from Project Gutenberg, while 84% is scraped from the internet, but he argues that this would lead to there being less rhymes in the training data, and thus make it more likely that generated poems would not rhyme. Marianne’s point is that there might be a lot of self-described “poems” in this uncurated mass of data, and that amateur poems might rhyme a lot more than literary, published poems. 

Many other things were discussed in yesterday’s meeting. For example, Laura and Mick talked about how to combine Jameson’s historicizing with Donna Haraway’s implosion method. Ida noted that she recently started rereading Jameson and hadn’t previously realised he has written about sci-fi. Zahra mentioned Ngai’s ideas about cuteness as also relevant here. Oh, and I think there might be a freshly established mini reading group on phenomenology at CDN now. I loved the cross-disciplinarity of this group – we even had two mathematicians which I hadn’t expected but was very happy about. Thank you to everyone – it is so inspiring to discuss theory with people who also want to discuss it!

If you’d like to join our next discussion, you can sign up for the mailing list where we’ll send out notifications about events. The signup page has the schedule too. Our next meeting is next Tuesday (17 March) at noon Bergen time, in the glass house at the CDN, with an online option. We’ll be discussing this paper:

Gunkel, David J. 2025. “The Différance Engine: Large Language Models and Poststructuralism.” AI & Society, September 25. https://doi.org/10.1007/s00146-025-02640-z

I had worried it would be difficult to have an actual discussion in the hybrid format, but it worked just fine. Several people had reading suggestions, and when Anja Salzman showed the group some German book recommendations, the online people self-organized a German-language online reading group in the chat – I hope they report back to those of us who are less proficient in German. Here are some of the reading recommendations from Anja and other participants:

• Frank Schirrmacher: Technologischer Totalitarismus – Eine Debatte. (2015)
• Rainer Mühlhoff: Künstliche Intelligenz und der neue Faschismus (2025)
• Maximilian Kasy: The Means of Prediction: How AI Really Works (and Who Benefits) (2025)
• Coeckelbergh and Gunkel: Communicative AI: A Critical Introduction to Large Language Models (2025)
• David Golumbia: Cyberlibertarianism: The Right-Wing Politics of Digital Technology
• Paulina Borsook: Cyberselfish (this one is almost impossible to get hold of but some libraries have it – and here is a gift link to an excellent interview with Borsook)
  

Several people thought that fascism and AI could both have been defined more clearly. “I agreed with everything in the article,” one participant said, “but I kept wishing he would make the argument more convincing,” and several people nodded in response. I actually didn’t notice a lack of clarity when reading the article, and found it well written. I’ve been reading up on fascism in the last year, so recognized most of the material discussed, but found it a useful summary that I think would be a good introduction. And as someone pointed out, it’s not like there’s one singular definition of fascism, it’s a fuzzy concept.

Dom, who happens to have written a book about myth, pointed out that myth is also not very clearly defined. But the treatment of myth in the paper is one of the ideas I found really interesting. Coeckelbergh argues that while classical fascism centers on a myth of a glorious past that must be reborn, technofascism puts emphasis on myths of the future, and in particular the myth of AI being about to be capable of almost anything. I’ve not thought of the connection between AI hype and fascism before, because the classic definitions of fascism usually emphasise the past. Another angle, which Dianna pointed out was missing from the article, is the way AI is being used to create a (fake) mythic past. Ida called AI-imagery “speculative fiction,” which is another good insight.

Another critique of the article was its emphasis on fascism alone. Various people pointed out that an alternate origin story could have been told, for example emphasising the cyberlibertarianism of the United States, or authoritarianism in general. A couple of people commented that the Marxist perspective seemed almost to be being avoided – Fuchs was mentioned, but barely, and what about cannibal capitalism, someone asked. Others were pleased that Erich Fromm was included. And several, like me, were particularly interested in Coeckelbergh’s emphasis on how AI is used to manipulate emotion, often through intimacy, and its connection to aesthetics.

Towards the end of the paper, Coeckelberg writes of the need to tell stories about the potential good uses of AI. But most of his paper is about the bad effects of technology. “What about #metoo? Or the Epstein files?” someone asked. The same technology can be used to tear down oppressive systems and build community. We need to not only be telling and retelling those stories as well as the bad ones, we should also work to enable more of these good uses of technology.

A lot more was said, of course. And I’m looking forwards to our next session! It’ll be on March 10th from 12 noon until 13:00 Bergen time, and we are reading Ryan Heuser’s paper “Generative Aesthetics: On Formal Stuckness in AI Verse“, published in the Journal of Cultural Analytics. Either join us in person in the glass house at CDN at the University of Bergen, or you can sign up for the Zoom.

Read the paper in advance, bring your own lunch and let’s talk theory. As a starting point I’m thinking just an informal in person group of whoever is interested, but leave a comment if you’re interested in joining online and maybe we can hook up a camera/microphone.

• 17 Feb Coeckelbergh, Mark. 2026. “Technofascism: AI, Big Tech, and the New Authoritarianism.” AI & SOCIETY, ahead of print, January 25. https://doi.org/10.1007/s00146-026-02862-9.
• 10 March Heuser, Ryan. 2025. “Generative Aesthetics: On Formal Stuckness in AI Verse.” Journal of Cultural Analytics 10 (3): 1036. https://doi.org/10.22148/001c.144825.
• 17 March Gunkel, David J. 2025. “The Différance Engine: Large Language Models and Poststructuralism.” AI & SOCIETY, ahead of print, September 25. https://doi.org/10.1007/s00146-025-02640-z.
• 14 April Mollema, Warmhold Jan Thomas. 2025. “AI-Generated Literature, Distant Writing and the Reader: Reflections on Floridi and Calvino.” Philosophy & Technology 38 (4): 168. https://doi.org/10.1007/s13347-025-01008-x.
• 21 April Muldoon, James, and Jul Jeonghyun Parke. 2025. “Cruel Companionship: How AI Companions Exploit Loneliness and Commodify Intimacy.” New Media & Society, December 18.https://doi.org/10.1177/14614448251395192.
• 5 May Amoore, Louise, Sj Bennett, Alexander Campolo, Benjamin Jacobsen, and Ludovico Rella. 2025. “Politics of the Prompt: Government in the Age of Generative AI.” Economy and Society 54 (3): 573–96. https://doi.org/10.1080/03085147.2025.2560177.

Norge kryr av forskere innen humaniora, samfunnsvitenskap og jus som har forsket på KI de siste årene og som kunne forklart noen av problemene med dette. Juridisk er bruk av KI i ansettelser høyrisiko som definert av EUs KI forordning, som er på vei inn i norsk lov. Det er svært stor fare for diskriminering (KI bias) og i tillegg vet vi at KI gjør mange feil.

Så kjære journalister: Bruk de mange forskerne med kompetanse på KI og etikk og samfunn!

Her er noen eksperter som kunne bidratt til dagens artikkel om KI i arbeidslivet. Hvis du er forsker som gjerne kan snakke med journalister om KI, skriv i kommentarfeltet! Jeg vet jeg har utelatt mange gode folk men oppdaterer gjerne listen!

• Maria Danielsen, UiT – filosof med ny doktorgrad på KI og etikk, bl.a. KI i arbeidsplassen. 
• Chesta Arora og Hilde Corneliussen – forskere, Vestforsk. Lager en kunnskapsoppsummering om kunstig intelligens (KI) og algoritmestyrt ledelse i arbeidslivet på oppdrag fra Arbeids- og inkluderingsdepartementet. Hilde har forsket på kjønn og teknologi i mange år.
• Taina Bucher, UiO – har skrevet bøker om algoritmisk makt og om Facebook som er mye sitert internasjonalt, og leder HumAIn som er en forskningshub for humanistisk forskning på KI. Svært kunnskapsrik. 
• Roger André Søraa, NTNU – har skrevet boken AI for Diversityså har veldig god oversikt over alt som har med mangfold, diskriminering og KI. Leder gruppen DigiKULT som forsker på hvordan teknologi former samfunnet, og bl.a. et forskningsprosjekt som heter BIAS: Mitigating Diversity Biases in the Labour Market.
• Samia Touileb, UiB – datalingvist som utvikler norske språkmodeller og særlig forsker på KI bias og hvordan håndtere dette. Hun er knallgod til å formidle.
• Gabriele de Seta, UiB – forsker på algoritmisk folklore, leder prosjektet ALGOFOLK som er finansiert av Trond Mohn. Har en Substack for prosjektet. Alltid oppdatert på de siste som skjer på nettet.
• Petter Bae Brandtzæg, UiO – har forsket på chatbots i ti år.
• Lukas Wilde, NTNU – har skrevet en bok (på tysk) som heter KI-Figuren: Chatbots, Avatare, NPCs som kommer ut i september 2026 og handler om KI-chatbots. Har også forsket på bildegenerering. 
• Annelin Eriksen, UiO – professor I sosialantropologi, har skrevet boken Evighetsmennesket om tekoligarker som tenker de skal laste opp sjelen sin og bli digitale. Leder forskningsprosjektet Human Futures: A study of Technoscientific Immortality. 
• Henrik Skaug Sætra, UiO – Har skrevet boken Hvordan redde demokratiet fra kunstig intelligens (Cappelen Damm 2025). Filosof som forsker på teknologi og bærekraft (miljø, sosialt og økonomisk).
• Maka Alsandia, Høgskolen i Østfolk – phd-stipendiat som skriver doktoravhandling om bruk av KI i norsk offentlig sektor og bruk av tilpasset politisk reklame  i Norge. Har fra før 15 års erfaring bl.a. fra Utlendingsdirektoratet og OSSE (se intervju). Har skrevet artikkelen “Navigating the Artificial Intelligence Dilemma: Exploring Paths for Norway’s Future“.
• Philip Thingbø Mlonyeni, UiO – filosof med fokus på teknologifilosofi- og etikk. Doktorgrad i teknologietikk om hvordan teknologi generelt og KI spesielt påvirker mellommenneskelige sosiale relasjoner.
• Lisa Reutter Larsen, Københavns Universitet – norsk forsker med PhD fra NTNU. Forsker på datafisering av offentlig sektor i nordisk kontekst.
• Elisabeth Austad Asser – PhD fra UiA i 2023 om hvordan mennesker og samfunn blir utfordret av KI-teknologi. Forfatter av boken Teknologiens makt (2025).

Kilden kjønnsforskning lagde en kartlegging av norsk forskning på kunstig intelligens og likestilling i 2020 som også kan være nyttig, selv om det er noen år siden nå.

Det er mange andre forskere – som Anja Salzmann og andre på AI LEARN for KI og læring, Scott Rettberg leder Senter for digitale fortellinger, forsker på samhandling mellom mennesker og KI for å skrive litteratur og skape kunst. Alle på MishMash for forskning på KI og kunst, mange som har forsket på chatbots i helsesektoren og offentlig sektor i norsk kontekst, osv osv osv. Niels Nagelhus Schia på NUPI forsker på KI, demokrati og sikkerhet, og ledet regjeringens ekspertgruppe på KI og valg.

Det er også mange forskere også innen informatikk som har skrevet gode, kritiske artikler om hvordan KI tas i bruk. Les for eksempel Anuja Vats sin kronikk “KI kan være i ferd med å koble om hjernene og samfunnet vårt“.