This paper examines the development of the Artificial Intelligence (AI) meta-debate in Sweden before and after the release of ChatGPT. From the perspective of agenda-setting theory, we propose that it is an elite outside of party politics that is leading the debate - i.e. that the politicians are relatively silent when it comes to this rapid development. We also suggest that the debate has become more substantive and risk-oriented in recent years. To investigate this claim, we draw on an original dataset of elite-level documents from the early 2010s to the present, using op-eds published in a number of leading Swedish newspapers. By conducting a qualitative content analysis of these materials, our preliminary findings lend support to the expectation that an academic, rather than a political elite is steering the debate.
mass of AI experts believe that AI may pose existential risks to the future. One of the most influential of these is Eliezer Yudkowsky, who has argued for shutting down AI development 4.
However, this focus on long-term risk has been strongly criticised by other AI researchers,
such as Timnit Gebru [
Despite being high on the agenda, there are relatively few articles that consider AI from the
perspective of the meta-discussion (for an exception see Nguyen [
Previous research suggests that national media are often the most important source of public
agenda setting, at least in countries such as the UK and Canada [
The material in our study comes from argumentative texts (opinion pieces) by elite representatives published in leading Swedish newspapers. As the texts consist mainly of arguments with important information between the lines, we conduct a qualitative content analysis. This is carried out by three of the authors, each of whom is a native speaker with a degree in either social sciences or computer science. During the coding process, we focus on categorising the core content of the texts and classifying whether an article emphasises short-term or long-term risk, or does not mention risk at all. For our purposes here, we define short-term as authors writing about concrete risks that could occur in the near future related to the implementation of AI, whereas long-term risks are those that refer to hypothetical or existential risks instead.
With our approach, we contribute to the meta-debate on the basis of empirical evidence,
a combination that is largely lacking in current social science studies (where Acemoğlu and
Johnson [
A starting point for our study is that what is said in the leading newspapers remains an
important source of opinion formation for citizens. A seminal paper by McCombs and Shaw [
More recent evidence from Sweden, the case for our study, shows that the media continues
to play an important role. Shehata and Strömbäck [23] examine the causal efect of coverage
in the traditional news media on perceived importance by the public. They find that issues
that received more coverage in the traditional news media tended to be perceived as more
important by the public over time. In contrast, no such efects were documented for issues
that were low on the media agenda. Djerf-Pierre et al. [24] address how media salience from
diferent sources (news media and alternative media) uniquely afects the relative strength
or weakness of “sociotropic” beliefs. They argue that while alternative and social media have
become important elements in agenda setting and issue salience in Sweden their findings point
to a continued important role for traditional news media. Given that coverage in the traditional
news media has an impact on the perceived salience to the general public in Sweden [23], the
number of articles on AI as such is important. This supports the idea of McCombs and Shaw
[
The extent to which the national agenda is top-down or bottom-up driven also depend on
the party system. Some research from the US suggests that US state legislators are very
responsive to public opinion from below [25]. In the US case, the perspective of the people or party
supporters is thus important relative to the party elite. Holmberg [
The study of agenda-setting dynamics in the public sphere has mostly been advocated for the cases of the United States and the United Kingdom, but recent scholarship has turned its attention to other contexts. Examining the South Korean context, Zhang et al. [26] investigates how social media have afected the prevailing “rules” of agenda setting and argues for the consideration of a “multi-participant agenda setting” dynamics. The authors consider agenda setting in relation to the interplay between bots, news media and the public in the context of presidential elections and the salience of key issues in the public sphere. In this context agenda setting appears to partially shift away from the news media over time, following both the “bot agenda” and the “public agenda”.
Agenda setting scholarship has also begun to explore the dimension of technology and AI
in the context of techno-political “change” in modern societies, although research to date is
limited. For example, Nguyen [
Using topic modelling, Nguyen finds that over time, these associations have become more focused on the techno-social implications of AI in areas such as international politics, security and finance over time. Indeed, with the caveat that the sample is overwhelmingly US-based, Nguyen argues that this signals an increasing politicisation of AI. This link between AI and politics in the context of 2010 − −2022 is further explored through the identification of four “risk” categories related to cybercrime/cyberwar, information disorder (mis- and disinformation), surveillance, and data bias. Specifically, Nguyen looks at the increase in “explicit” data risk references in the media (news articles) over time and finds that a total of 47 percent of AI-centric articles explicitly address risks, implying a shift towards an overall more attentive media coverage. However, by the end of the data collection period (2020 − −2021), mentions of risk decreased in Nguyen’s sample.
From a more critical perspective, Dandurand et al. [
In a very recent piece, Xian et al. [27] look specifically at the coverage and characteristics
of English-language (mostly US-based) news articles on generative AI such as ChatGPT. The
authors address the ability of these media to shape perceptions of generative AI technologies
over time. To address these issues, the authors use a mixed-methods approach that combines
topic modelling, qualitative coding, and sentiment analysis. Similar to Nguyen [
However, this topical distribution difers between national contexts. For example, US-based coverage gives more attention to the topic of “business” and “technological development”, while Indian coverage is significantly more focused on “corporate technological development”, and UK and Australian coverage focuses on “regulation and security” and “technology development”. In addition, business-oriented articles tend to frame generative AI in a more positive or neutral light, while security and regulation articles (which increasingly focus on AI governance) tend to be more neutral or negative in their framing. Furthermore, the authors find that US-based articles tend to be more positive overall in their framing of generative AI technologies than the non-US Anglophone sample, but mostly in the case of specialised and local news outlets.
Our work is related to the perspective of Nguyen [
Sweden is a parliamentary system where political parties have traditionally played an
important role in organising people and people’s opinions (see Holmberg [
As a member of the European Union since 1995, Sweden is one of the countries that will
adopt a new AI law initiated by the EU Commission and passed by the European Parliament in
March 2024. This law focuses on restricting how AI tools can be used in practice, depending
on whether the risk to humans is considered minimal, limited, high or unacceptable. While the
Swedish parliament has been relatively silent on AI issues, the supranational level of the EU
is thus at the forefront of regulation, as is the case in some other transnational areas such as
climate change policy. Influences from the US and the continent have also spread to academics
and elite debaters. The US debate [
With this in mind, our overarching research question is:
From our preliminary understanding of the context, we have the following expectations:
Given the nature of the materials and the research question, a starting point was that qualitative
analysis using human annotation was likely to be more valid and eficient than, for example,
an LLM approach. While previous work has highlighted the promise of using LLMs to classify
text [
For our coding exercise, we had the texts read and classified by three academics who are native Swedish speakers, have a PhD and come from three diferent disciplines: economics, computer science and political science.
For our data collection we used Mediearkivet9, where we searched for the terms “artificiell intelligens” or “AI” or “artificial intelligence” in articles placed in either the debate or opinion section between 01.01.2000 and 25.10.23. This resulted in 318 articles (after removing duplicates and misidentified pieces). Almost all of them are a single page, although the actual length of the text varies. Note that this results in a selection of articles that are both about AI and those that only mention AI as a simple example. For example, the article “Koppla ihop militär och civil forskning” only mentions artificial intelligence as an example of one type of research (“Areas such as cyber security, communication, robotics and artificial intelligence are crucial to the defence of Sweden”10). We then looked at all the articles individually and judged whether they 8We also briefly experimented with an LLM approach (using the LLM Gemma) for this study, but found that the risk classification was not reliable. 9We have collected the articles via https://app.retriever-info.com/services/archive. Due to licensing issues, the articles cannot be freely shared. A full list of the articles included here, including the title of the article, the name(s) of the author(s) and the date of publication, is available on request. 10“Områden som cybersäkerhet, kommunikation, robotik och artificiell intelligens är helt avgörande för försvaret av Sverige” were really “about” AI or just mentioned it in passing. This led to a significant reduction to 74 articles.
All texts were downloaded in their PDF format, cropped to show only the article in question (author names and afiliations were removed), and converted to .txt format using pdftools in R. These latter files were then checked and corrected manually (for example, to deal with strange characters, font size issues, and problems with columns). Each article was then split into four sections: title, header, body and additional notes. Of these, only the body is included in the later analysis. Metadata was also collected for each article, including a unique document ID, publication date, newspaper title, article title, section topic (if mentioned), additional notes, authors and afiliations.
For all articles, we also note the authors (of which there were often more than two) and their afiliations. We then reduce these afiliations to four broad categories: corporate (related to companies and businesses), NGO (all non-governmental and non-commercial organisations), political (related to political parties or holding oficial positions) and academic (related to universities and research institutes), as well as a combination of those working for companies and those working in academia. Of these categories, most were written by those associated with academia (31 articles), followed by companies (15 articles), politicians (14 articles), NGOs (10 articles) and the combination of companies and academia (4 articles).
We divided the texts between the three human coders, who were given the following instructions:
2. Give a label for the text in a single word (or two if its a common expression such as
European Union) 3. Indicate the risk on a scale of three categories: a) no risk, b) short-term risk, c) long-term risk 4. Say why the text was coded into that risk category
To be classified in one of the risk categories, the concept of risk had to be part of the core
argument of the article, not just mentioned in passing. A short-term risk classification concerned
a risk related to a concrete problem or phenomenon here and now, whereas a long-term risk
concerned hypothetical scenarios and/or existential threats to humanity. Given our materials,
it is dificult to establish strict definitions of short-term vs. long-term risk, along the lines of,
for example, the World Economic Forum’s distinction between 2-year and 10-year risks [
As mentioned in the methodology section, we coded a total of 74 articles. We will first look at the categories given by the human coders to better understand what the articles we selected were about. We then turn to the coding of risk, looking at the evolution of risk focus over time.
Group Ethics Regulation Risk Applications Characteristics Economy Labour market Development Education Research Sustainability Techno-optimism 16
There is a risk in believing that AI can assist decision-making in moral dilemmas Criticism of EU AI regulation AI is developing too fast, ChatGPT is just a small step from AGI, this is a major existential risk for humanity AI could help us prevent and solve crime This is what AI is and could be in the future The author argues that AI will transform the retail sector Young people will find it harder to get their first job as simpler tasks are automated by AI systems Suggests the creation of a committee to deal with AI issues The author argues that people need to be better educated about the benefits and risks of AI More interdisciplinary AI research is needed The author argues that we need to invest in sustainable AI by integrating responsible and transparent AI practices Technological change should be accelerated to benefit the economy, fears of mass unemployment are unfounded Finally, we look at the authorship of the articles to see if certain authors preferred to talk about diferent types of risk - or not.
Table 1 shows the labels assigned to each article by the coders, as well as an example in the form of a one-sentence description of an article of that type. The most common label is “Ethics”. These articles describe various topics related to the ethical issues surrounding AI. Sometimes these arguments include “warnings” about the potential threat that AI can pose to democracy, or the biases that can arise from using AI trained on poorly curated data, and the lack of transparency associated with this issue. Other issues discussed include the inherent limitations of an AI ever becoming human, such as the impossibility of having a moral compass or ever developing conscious awareness. Although this is related to risk, not all of them mention it, and when they do, it is mostly related to the idea of short-term risk.
The second category, Regulation, deals with various calls for the regulation of AI, such as criticism of the European Union’s proposed Artificial Intelligence Act, calls for harmonisation of AI regulation between the US and the EU, and the risks posed by the lack of regulation. This category is related to, but diferent from, the Development category in that the latter is more focused on calls to action, focusing on how AI should be managed and developed in the country. The third most common category - risk - discusses issues such as the risks of AI being used for autonomous weapons, the impact of unreliable and biased training data, and concerns about integrity. Of the 9 articles in this category, 6 are categorised as dealing specifically with long-term risks. These mostly concern the idea that current AI development is too fast and that future systems could easily lead to Artificial General Intelligence (AGI).
The fourth and fith categories relate to descriptions of what what AI is and various applications in which it might be used. Here we find examples of how AI can be used by the police to help solve crimes, or how hospitals can use it to digitise many of their systems, or how it can help reduce the administrative burden in public institutions. The sixth and seventh categories are related, but while the “economy” category focuses on the economic sector (most often commerce) in general, the “labour market” category focuses mostly on the impact of AI adoption on people’s jobs. In the latter category, while most authors agree that AI will transform the labour market, the disagreement lies in whether this is seen as a positive development (for example, that AI will create more jobs and opportunities) or whether this development is negative (for example, in relation to the first jobs young people take).
As for the other categories, “Education” focuses on the development of skills related to AI as well as further investment in more flexible education systems to cope with rapidly changing markets, “Research” includes articles calling for more (interdisciplinary) research, “Sustainability” focuses on diferent ways to ensure that AI and its developments will be sustainable in the future, while “Techno-optimism” includes two articles that mainly push the framework that any technological change is beneficial and should be supported and accelerated.
Figure 1 illustrates the evolution of the assigned risk categories over time, as determined by the coders. It is noteworthy that from January 2023, which coincides with the release of ChatGPT, there was a significant increase in the number of articles published. This increase was most pronounced for short-term risk articles, while long-term risk articles showed a more Long-Term Near-Term No Risk Total modest increase. Overall, the majority of articles did not contain a core argument about risk. An example of long-term risk comes form an article accusing “leading AI companies [to play] Russian roulette”11, with the article making the argument that AI development in general is moving too fast and that techniques such as ChatGPT are only a small step away from Artificial General Intelligence (AGI) and therefore a major existential risk to humanity. An example of short-term risk comes from an article a about the current use of AI by law enforcement and how it is unreliable (for example, when used for facial recognition) due to racial bias in the training data12. Finally, an example of no risk comes from an article entitled “AI is redrawing the labour market and our skills needs13, which does not identify any risks at all, but instead argues that Sweden needs to invest in the data-driven economy in order not to fall behind the AI development of other countries.
As Table 6.3 shows, most of the articles (31 out of 74) were written by authors associated with the academic community, such as professors or researchers, followed by those associated with industry - most often working in the field of data or computer science; politicians, either speaking on behalf of their party or on behalf of the government; and representatives of NGOs. In most cases, articles were written by one or more authors from a single group, with the only exception being collaborations between those working in industry and those working in academia. Looking at how the diferent types of authors wrote about risk, we find that those associated with NGOs and politicians never wrote about long-term risk. Instead, these two types either talk equally about short-term risk, or do not mention risk at all, or have a clear preference for the latter, as is the case with politicians. Thus, during the studied period, politicians tended to focus their pieces on neutral aspects of AI rather than highlighting risk.
Given that so many of the long-term risk articles are written by those who belong to academia, it might be interesting to look at the type of academic discipline they belong to, as table 6.3 shows. It is interesting to note that while computer science is the most common category, none of the authors here wrote an article in the long-term risk category, opting for short-term risk or not mentioning risk at all. This suggests that, of all the academic disciplines, computer scientists in the Swedish context do not want to focus on the long-term aspects. Instead, articles on long-term risk were most often written by someone working in mathematics, economics or philosophy. In fact, all four articles written by a mathematician was by the same author. This suggests that a few academics with experience in diferent academic fields are highlighting existential risks, while the majority are not. Another finding is that AI is not (yet) ’owned’ by politicians, but rather by an academic elite.
This study analysed the evolution of the AI debate in a national context, based on a human reading of the argumentative texts to identify the evolution over time. We began our study with the expectation that the character of the AI debate has changed over the past few years, and that this should be a top-down development. As the dataset contains arguments and is relatively small, we expected that human expert coding would be essential, as opposed to categorising pure political news material. We therefore conducted a qualitative content analysis.
Our results indicate that the emphasis on short-term risks has increased in the Swedish
public debate since the release of ChatGPT. This finding complements previous analyses of the
prevalence of risks in the news media, for example Nguyen [
Another reflection is that the Swedish AI debate seems to be somewhat more nuanced than,
for example, the debate in Canada, where Dandurand et al. [
A relevant follow-up to our study would be to analyse the AI debate in other national EU contexts: neighbouring countries such as Norway and Denmark, or more contrasting cases such as Spain, in order to be able to generalise our findings to a greater extent. EU regulation is still in its infancy, and country diferences are likely to continue to play an important role.
This work was supported by the Marianne and Marcus Wallenberg Foundation and the Wallenberg AI, Autonomous Systems and Software Program – Humanity and Society. Bastiaan Bruinsma is grateful for support from the GATE Project, funded by the European Union’s Horizon 2020 WIDESPREAD-2018-2020 TEAMING Phase 2 programme under Grant Agreement No. 857155.
One of the authors, Moa Johansson, declares to have co-authored two opinion articles analysed here - “Skrämsel om AI döljer de verkliga problemen”, published in Göteborgs Posten on 24-022023, and “Alla måste lära sig mer om artificiell intelligens”, published in Göteborgs Tidningen on 02-04-2019. To avoid potential conflicts of interest, both articles were coded by one of the other two coders. [22] W. L. Bennett, S. Iyengar, A New Era of Minimal Efects? The Changing Foundations of
Political Communication, Journal of Communication 58 (2008) 707–731. doi:10/dxdvrv. [23] A. Shehata, J. Strömbäck, Not (Yet) a New Era of Minimal Efects: A Study of Agenda Setting at the Aggregate and Individual Levels, The International Journal of Press/Politics 18 (2013) 234–255. doi:10/f2z2jt. [24] M. Djerf-Pierre, A. Shehata, B. Johansson, Media Salience Shifts and the Public’s Perceptions About Reality: How Fluctuations in News Media Attention Influence the Strength of Citizens’ Sociotropic Beliefs, Mass Communication and Society (2024) 1–26. doi:10/gtgbvm. [25] P. Barberá, A. Casas, J. Nagler, P. J. Egan, R. Bonneau, J. T. Jost, J. A. Tucker, Who Leads? Who Follows? Measuring Issue Attention and Agenda Setting by Legislators and the Mass Public Using Social Media Data, American Political Science Review 113 (2019) 883–901. doi:10/gf9r33. [26] M. Zhang, Z. Chen, X. Liu, J. Liu, Theory and Practice of Agenda Setting: Understanding Media, bot, and public agendas in the South Korean presidential election, Asian Journal of Communication 34 (2024) 24–56. doi:10/gs9n4b. [27] L. Xian, L. Li, Y. Xu, B. Z. Zhang, L. Hemphill, Landscape of Generative AI in Global News:
Topics, Sentiments, and Spatiotemporal Analysis, arXiv preprint arXiv:2401.08899 (2024). [28] O. Häggström, Tänkande maskiner: Den artificiella intelligensens genombrott, Fri Tanke,
Stockholm, 2023. [29] Y. Elazar, N. Kassner, S. Ravfogel, A. Ravichander, E. Hovy, H. Schütze, Y. Goldberg, Measuring and Improving Consistency in Pretrained Language Models, Transactions of the Association for Computational Linguistics 9 (2021) 1012–1031. doi:10/gpbd5z. [30] L. Hagström, D. Saynova, T. Norlund, M. Johansson, R. Johansson, The Efect of Scaling, Retrieval Augmentation and Form on the Factual Consistency of Language Models, in: H. Bouamor, J. Pino, K. Bali (Eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Singapore, 2023, pp. 5457–5476. doi:10/gtn9qc. [31] B. Wang, X. Yue, H. Sun, Can ChatGPT Defend its Belief in Truth? Evaluating Llm Reasoning Via Debate, in: H. Bouamor, J. Pino, K. Bali (Eds.), Findings of the Association for Computational Linguistics: EMNLP 2023, Association for Computational Linguistics, Singapore, 2023, pp. 11865–11881. doi:10/gtn9qg. [32] P. West, X. Lu, N. Dziri, F. Brahman, L. Li, J. D. Hwang, L. Jiang, J. Fisher, A. Ravichander, K. Chandu, B. Newman, P. W. Koh, A. Ettinger, Y. Choi, The Generative AI Paradox: “What It Can Create, It May Not Understand”, in: The Twelfth International Conference on Learning Representations, 2024, pp. 1–29. [33] E. Rønningstad, E. Velldal, L. Øvrelid, A GPT among Annotators: LLM-based Entity-Level Sentiment Annotation, in: S. Henning, M. Stede (Eds.), Proceedings of The 18th Linguistic Annotation Workshop (LAW-XVIII), Association for Computational Linguistics, St. Julians, Malta, 2024, pp. 133–139. [34] P. Törnberg, Best Practices for Text Annotation with Large Language Models, 2024. doi:10 /gtn9qf. [35] H. Kim, K. Mitra, R. L. Chen, S. Rahman, D. Zhang, MEGAnno+: A Human-LLM Collaborative Annotation System, 2024. doi:10/gtn9qd.