Metaphor Machines: How AI Writing Circles Its Own Blind Spots and Fabricates Science

Large language models have dazzled the world with their fluency, but a growing body of rhetorical analysis suggests that the very mechanism that makes them eloquent also makes them systematically unreliable. In an open-access paper published in AI & Society, media scholar Jill Walker Rettberg of the University of Bergen argues that LLMs are, at their core, metaphor machines: they generate text by exploiting similarities and statistical proximity between words rather than by reasoning about logical causality. That metaphorical engine, she contends, produces a distinctive family of failures she traces through AI-generated folktales, a botched news summary, and a peer-reviewed journal article: texts that circle their most important themes without ever touching them, and citations that look scholarly while supporting nothing at all.

The central concept Rettberg borrows and extends is what John Gallagher has called orbital argumentation, a rhetorical vice in which the main point of a text is repeatedly approached but never explicitly stated. Insights are mentioned but never formulated; conclusions are gestured at but never drawn. Gallagher notes that such writing looks convincing at first skim, which makes it well suited to bureaucratic documents that nobody expects to read carefully. The trouble, Rettberg argues, is that a large share of LLM training data consists of exactly this kind of evasive, promotional, deliberately vague prose. When the same stylistic gravity is applied to genres that demand precision, such as academic writing, journalism, document summarisation, and national security briefings, the results can be disastrous.

The technical roots of the problem lie in two landmark innovations. The first was the 2013 discovery by Tomas Mikolov and his colleagues that word vectors capture analogical relationships as constant offsets in a high-dimensional space: the famous equation in which King minus Man plus Woman lands very close to Queen, or Paris minus France plus Italy equals Rome. Rettberg points out that this is precisely the structure of the analogical metaphor Aristotle described in his Poetics, where the cup is to Dionysus as the shield is to Ares, so that the cup may be called the shield of Dionysus. Language models perform this proportional substitution mechanically, billions of times over, without any human interpretation of whether the resulting comparison is apt.

The second innovation was the transformer architecture introduced in 2017, which replaced sequential processing with self-attention, a mechanism relating different positions of a single sequence to compute a representation of the whole. Before transformers, generating text in the style of Shakespeare required hand-crafted word lists, syntax rules, or specialised training sets. After transformers, style itself became a statistical network of associations extracted from general training data, transferable on demand. Style transfer, Rettberg observes, is itself a metaphorical operation, substituting one register for another. The upshot is that LLMs produce metaphors through computation, while humans read them as language to be interpreted, and the mismatch between those two activities is where the trouble begins.

To see the mechanism in action, Rettberg turns to narrative. Literary scholar Anne Sigrid Refsum analysed LLM-generated versions of the Norwegian folk tale The Sweetheart in the Forest, in which the original story strongly implies sexual violence and even cannibalism. The language models proved incapable of representing rape, but they exaggerated the cannibalism instead: in one Claude-generated version the villain chops the heroine into pieces and sprinkles salt over the meat. Refsum describes the result as blood-soaked yet strangely hollow, with young women existing only as flesh to be eaten. Rettberg reads this as orbital narration: the model substitutes a metaphor, cannibalism, for the unnarratable tenor, sexual violence, which its safety training and content policies have rendered untouchable.

A similar orbit appears in a dataset of stories Rettberg and Hermann Wigers generated by prompting a model to write a 1500-word story for 236 different nationalities. The fifty stories generated as German are striking for their near-total absence of external conflict. Instead, protagonists wrestle with guilt over events they did not cause: a woman haunted by a drowned childhood friend, or a woodcutter who tells a forest guardian, I never meant disrespect, I only did my work and never harmed the forest, a line Rettberg identifies as the Nuremberg defence in miniature. World War II is never mentioned in any of the fifty stories, yet the guilt saturating them functions as an analogical metaphor for the war, with the woodcutter as vehicle and the unspoken historical trauma as tenor. The Australian stories show the same structure from another angle: they fixate on land, country, and nature, with protagonists learning from wise elders, yet Indigenous people are never mentioned and every protagonist has an Anglo name.

The most consequential example is not literary but geopolitical. In October 2025 the Norwegian News Agency NTB published an AI-generated summary of Telenor’s annual security report, which the original document opens by describing the most serious geopolitical situation since the Second World War, dominated by hybrid threats, cyber-attacks, sabotage, and influence campaigns by foreign states. The LLM summary omitted all of this, reporting instead that the greatest risks to telecommunications infrastructure were extreme weather, aging infrastructure, and digital risks. It even fabricated a quote attributed to a security director named Håkon Berg, a man who does not exist; the report’s preface was actually written by Telenor Norway’s chief executive Birgitte Engebretsen. Rettberg notes the gendered statistical logic at work: the model apparently predicted that a solid Norwegian male name, one whose family name literally means mountain and whose given name belonged to kings, was more likely in that context. The retracted story was still available uncorrected on the website of the newspaper Finansavisen as of July 2026.

In scholarly writing, orbital argumentation manifests as misaligned citations: references to real works that relate to the topic but do not support the specific claim attached to them. Rettberg dissects a case in a 2025 AI & Society article on cognitive imperialism, a term coined by Mi’kmaq scholar Marie Battiste in 1986. The article cited an obscure 2002 Battiste report, cited only eight times in over two decades, to support the claim that AI tools can aid in revitalising endangered indigenous languages, a report that never mentions artificial intelligence at all and argues almost the opposite: that indigenous knowledges are inseparable from specific communities. The citation is not random, Rettberg argues, but misaligned: the token Battiste sits close to indigenous languages and cognitive imperialism in the model’s vector space, so when those concepts activate, the gravitational pull of the name drags in the wrong publication. She characterises such systematically produced misalignment as a form of automated scientific fabrication, and notes that even Grammarly’s citation finder reproduced the failure when she tested it, suggesting an unrelated paper about birds in Norwegian fiction because the famous name Karl Ove Knausgård exerted stronger statistical gravity than the actual source.

The stakes extend beyond individual errors to the knowledge ecosystem itself. Rettberg observes that universities and journals often lack clear guidelines, that commercial tools promise to compress days of literature review into minutes, and that the pressure to publish falls hardest on early-career and precariously employed academics. She reported her concerns about the fabricated citations to the journal’s editors in September 2025, yet no retraction or erratum had appeared by the time of writing. Her response is a set of five rules for resisting what critics have begun to call scholarslop: expect key terms to be defined and trace who coined them; be suspicious of drive-by citations that never explain how a source supports a claim; use citation styles requiring page numbers and direct quotations; actually read the articles you cite, beyond the abstract; and distrust ChatGPT’s polished revisions until you have interrogated them critically the next day.

Underlying all five rules is a single insight with implications far beyond academia: the omissions and hallucinations of LLM-generated text are not random glitches but the predictable output of a metaphorical engine that drifts toward heavily weighted, statistically legible territory and away from contested, politically sensitive, or poorly represented topics. War, sexual violence, colonial dispossession, and antisemitism keep vanishing from AI prose not by accident but by architecture. If Rettberg is right, then recognising the rhetorical vices of machine writing, the orbiting, the floating motifs, the misaligned citations, is not a literary indulgence but a core survival skill for journalists, scientists, and citizens who must now decide, sentence by sentence, whether the words before them describe the world or merely circle around it.

Subject of Research: Rhetorical analysis of metaphor-based text generation and citation fabrication in large language models

Article Title: LLMs are metaphor machines: orbital argumentation, misaligned citations, and scientific fabrication in AI-generated writing

Article References: Rettberg, J. W. (2026). LLMs are metaphor machines: orbital argumentation, misaligned citations, and scientific fabrication in AI-generated writing. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03310-4

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03310-4

Keywords: large language models, metaphor, orbital argumentation, misaligned citations, AI-generated writing, scientific fabrication, hallucination, AI bias, transformers, vector analogies, scholarly publishing, AI literacy