Swedish AI Experts Warn the Real Danger Is Death by a Thousand Papercuts

While the world’s attention fixates on visions of superintelligent machines and robot uprisings, the people actually building and governing artificial intelligence in Sweden are worried about something far less cinematic. In a new study published in AI & Society, researchers Irja Malmio of the Swedish Defence University and Misse Wester of Lund University interviewed thirteen professionals drawn from industry, government, academia, and the security sector, and found a striking disconnect between the spectacular AI narratives dominating public debate and the mundane, incremental harms that experts see unfolding on the ground. One respondent captured the mood with a phrase that became the study’s title: rather than an AI takeover, they believe more in a scenario of “death by a thousand papercuts,” in which a series of small, complex events quietly erodes democratic values such as trust and accountability.

The study is grounded in the concept of sociotechnical imaginaries, the collectively held and institutionally stabilized visions of desirable futures that shape how societies organize emerging technologies. Sweden offers a particularly revealing case. The country’s sociotechnical identity rests on a high-trust welfare state built on decades of meticulous datafication of its citizens, a heritage scholars have called the Nordic data imaginary, intertwined with a stubborn self-image as a European innovation frontrunner. That identity now faces profound friction. The Swedish AI Commission has warned that the nation’s position in the global AI race is increasingly precarious, while the government has ordered more than one hundred agencies to implement AI solutions and declared an ambition to rank among the world’s top ten AI nations.

To understand how professionals navigate this tension, the researchers conducted hour-long semi-structured interviews between December 2024 and January 2025, in Stockholm and via secure digital platforms, in English or Swedish. The sample included four respondents from government security and defense, two from civilian government and public infrastructure, four from academia and research institutes, and three from the private sector. Four were directly involved in engineering AI models, while the rest worked in policy, organizational transformation, and public discourse. All affiliations were pseudonymized, and the analysis followed a reflexive thematic approach coded in the original Swedish to preserve latent meanings, yielding four main themes: realism, regulation, dystopia, and utopia.

The first and perhaps most paradoxical finding is what the authors call strategic oscillation. When grounding AI in daily practice, the professionals consistently described it as a mundane tool, a system for statistical reasoning and pattern recognition, sometimes dismissively calling such systems “stupid robots” prone to hallucination. They rejected existential doom narratives, with one noting that the very concept of intelligence in artificial general intelligence remains poorly defined. Yet when the conversation shifted to geopolitics and long-term trajectories, many of the same respondents pivoted to hard determinism, invoking an unstoppable global AI race in which competitive pressure means, as one security analyst put it, that “you do not have the option of not developing it.”

The researchers argue this is not a logical failure but a functional strategy. Framing AI as an autonomous, inevitable force legitimizes rapid technological expansion and shifts moral responsibility to the technology’s momentum, while retreating to the view of AI as a neutral tool reaffirms human agency and the professionals’ own role as necessary intermediaries. By alternating between these positions, practitioners can answer political mandates for acceleration while performatively reinvoking the trust and transparency on which the Nordic model depends. The same duality appeared in discussions of bias: an equality investigator warned that AI-driven decisions risk reinforcing discrimination against underprivileged groups, while a developer of large language models argued that stripping bias from training data degrades performance and that models should be descriptive rather than prescriptive, a position critics say structurally guarantees that inequality gets reproduced once those models enter public institutions.

Regulation emerged as a second battleground. Respondents described a widening gap between political performativity and technical reality, with top-down directives perceived as rhetoric driven by technological ignorance. For security and innovation practitioners, the EU AI Act’s compliance burden risks binding European actors while adversaries advance unimpeded. Respondents also flagged a structural double standard: publicly funded national models stagnate under strict compliance while global monopolies scrape the same public data with impunity, a dynamic the study links to data colonialism. On explainability, the divide was even sharper. Corporate strategists treated “responsible AI” as a pragmatic mechanism for brand retention, technical practitioners dismissed current explainability methods as unreliable, and one research head bluntly called explainability an “academic hiding place,” equating it with mere transparency about training data.

The study’s most conceptually charged finding concerns ethicalization. Across the interviews, nearly all actors agreed on the need for “ethical AI” without being able to define what it entails, rendering the concept, in one academic’s view, practically useless. The authors interpret this not as philosophical failure but as a performative speech act: labeling a trajectory as ethical establishes its normative legitimacy and steers development away from technical or political contestation. In the Swedish context, ethicalization now functions as what the researchers describe as a security-driven enclosure, a “moral defensive wall” in which digital sovereignty justifies rapid expansion of domestic AI capabilities while framing external influences, from authoritarian states to American tech monopolies and their “value corridors,” as existential moral threats that constrain legitimate critique.

When the professionals turned to speculative futures, dystopian visions clustered around four threats: epistemic collapse in a post-truth world where synthetic media triggers the “liar’s dividend” and authentic evidence is dismissed as fake; cognitive disintegration, in which habitually delegating analysis to AI erodes the situational awareness decision-makers need, echoing recent research on the “oversight fallacy” of humans-in-the-loop; the transparent society, where Sweden’s high-trust data culture creates public naïveté that could supply the pre-existing architecture for authoritarian surveillance; and the concentration of power in an oligopoly of tech conglomerates whose models, as one respondent put it, are aligned with the values of “a small bunch of white men in the USA.” Notably, dissenting voices within the sample pushed back, citing weak empirical support for filter-bubble polarization and negligible measured effects of AI on geopolitical disinformation.

The utopian pole proved equally deterministic but strikingly conditional. Respondents imagined AI optimizing climate response, liberating workers from administrative friction, and even forging a “Star Trek society” of post-scarcity abundance, yet these visions consistently hung on conditional if-then statements: if AI is developed and used correctly, if open access is prioritized, if regulation is applied properly, if systems are designed wisely. Skeptics within the sample called this reliance on human goodwill historically naive, with one security expert describing the current AI expansion as a “gigantic experiment” driven by the fact that money can be made from it rather than by any societal benefit. Some practitioners proposed a radical structural remedy: state-funded national language models maintained as public infrastructure, functioning like water, roads, and electricity.

The study’s broader significance extends well beyond Scandinavia. The authors document a profound shift in the Swedish imaginary away from an administrative welfare focus toward a security-driven enclosure, mirroring a European-wide drift in which the discourse of digital sovereignty increasingly eclipses the earlier ambition of transparent, democratic regulation. Crucially, this enclosure is not framed as a loss of liberty but is itself deeply ethicalized, with foundational values repurposed as a defensive weapon. The result is a vision gap between a public still captivated by binary myths of existential doom or utopian salvation and professionals grounded in mundane complexity, a gap that matters because grandiose narratives systematically distract from the incremental erosion of democratic agency, the degradation of epistemic standards, and the quiet expansion of surveillance infrastructure. In the papercut scenario, no single cut is fatal, which is precisely what makes the accumulated damage so easy to ignore.

Subject of Research: Sociotechnical imaginaries and ethicalization of artificial intelligence among Swedish professionals

Article Title: Death by a thousand papercuts: professionals’ imaginaries on ethicalization and technological determinism of AI in Sweden

Article References: Malmio, I., & Wester, M. (2026). Death by a thousand papercuts: professionals’ imaginaries on ethicalization and technological determinism of AI in Sweden. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03383-1

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03383-1

Keywords: artificial intelligence, sociotechnical imaginaries, Sweden, AI ethics, ethicalization, technological determinism, digital sovereignty, Nordic welfare state, EU AI Act, AI governance, surveillance, qualitative interviews