Applications using generative artificial intelligence (GenAI), such as large language models (LLMs) and LLM-powered chatbots, have become popular everyday tools for professionals and students worldwide, not only for generating content but also as fast and easy information sources. As open GenAI applications gain ground in the key sectors of our democratic pluralist societies, such as education, science, and media, they also open new avenues for various actors to exploit these models, at times for malicious purposes. In the (geo)political field, GenAI, viewed as a new powerful propaganda tool, is increasingly used for political purposes (Goldstein et al., 2023). Recent multidisciplinary studies combining AI and memory politics suggest that the GenAI era may “fracture” collective identities and undermine the internal cohesion of communities by influencing collective memory formation, making it particularly risky for small open societies such as the Nordic countries.
Positioned at this intersection, we examine how chatbots may subtly manipulate public discourse. We build on the idea of AI auditing (Birhane et al., 2024) and on the emerging scholarship addressing the collective memory dimensions of GenAI (e.g., Hoskins, 2024; Pelevina et al., 2025). Expanding on previous studies of LLM-mediated disinformation and memory politics – which have so far focused on geopolitical rivalries between the US and China, the Russo-Ukrainian war (e.g., Makhortykh et al., 2024), and historical atrocities (e.g., genocides) – we shift the attention to a less-studied context: the Nordic region.
Specifically, we study how Nordic history narratives can be mobilised for political and potentially commercial purposes. Our analysis draws on outputs from three open, commercial models – DeepSeek, ChatGPT, and Mistral – in Finnish, Swedish, and English. By combining a set of topics featuring contested history narratives, we explore how these systems mediate Nordic historical imaginaries. There are three themes that are deeply intertwined in memory-political contestations in the Nordic region: the environment and climate change; security politics (namely NATO enlargement); and colonialism and Indigenous communities (e.g., Sámi, an Indigenous people living in the cultural region Sápmi that spans northern Norway, Sweden, Finland, and Russia). In this article, our interest lies in the competing narratives of history reproduced in the chatbots’ responses – often between the lines. By varying the language, tone, form, and perspective of the prompts, we examine whether and in what ways LLM-powered chatbots mediate different perspectives, including critical viewpoints (nongovernmental organisations, critical research) and those of political or private actors (e.g., industry lobbies). Through this, we analyse how chatbots may contribute to strategic constructions of the past, present, and future (Miskimmon et al., 2014).
This article is structured as follows: First, we outline the aspects of Nordic memory-political contestations that are relevant to our study. We then introduce key characteristics of LLM-powered chatbots, framing them as sites of contestation. Next, we detail our method, data gathering, and analysis, and present the main findings. Finally, we discuss our findings in relation to previous research and current debates on LLM manipulation and the distortion of collective memory.
The image of the Nordics, reflected in their national identities and strategic historical narratives, is closely tied to social equality, clean natural environments, environmental sustainability, and a global reputation for promoting peace and defending human rights (Ojanen & Raunio, 2018). These values shape the region’s self-image and are mobilised by various actors, such as states, political parties, national and international organisations, private actors, and advocates in the strategic construction of past, present, and future (Miskimmon et al., 2014). Former narratives about the past are challenged by emerging alternative perspectives, for example, geopolitical crises challenge the supposed self-evidence of peace, which has become secondary to a security discourse about preparedness for war (see Table 1).
Dominant and alternative narratives across three selected themes
| Theme | Dominant (nationalistic) narratives | Alternative history narratives |
|---|---|---|
| Environment: Climate obstructionism | “Pure nature” of the Northern wilderness, part of Nordic identity. | Emphasise historical and continuing overexploitation of resources and biodiversity loss, and environmental harms of oil and forest industries. |
| “Clean energy” and industries are portrayed as responsible and efficient (in the past, present, and future). | Need for stricter climate policy in the future. | |
| Peace & security | NATO brings security to the Nordic countries. | Continuing relevance of Nordics as peace promoters and mediators. |
| Security-oriented and threat-based discourse (militaristic vocabulary), deterrence logic. | Emphasising traditional non-alignment (Finland & Sweden historically). | |
| Narrative shift after the Russian war against Ukraine and NATO accession. | ||
| Unity, equality, & internal others | Nordics as global equality leaders. | Focus on historical internal colonialism and giving voice to silenced or marginalised histories. |
| Narrative of a “unified people”. | ||
| Indigenous peoples as “internal others” (e.g., Sámi). |
Historically, the Nordic countries have relied on industries such as oil, forestry, and mining, yet their national identities, country images, and nationalistic history narratives – anchored in pure nature, clean air and water, lakes, forests, and fjords – continue to be mobilised for country branding as well as by activists and interest groups, often through competing narratives. Indigenous peoples such as the Sámi and the Greenlandic Inuit, who have long inhabited the northern regions, have suffered from the colonial practices of the Nordic states and continue to be positioned as internal “cultural others” in nationalistic Nordic history narratives (e.g., Kuokkanen, 2024; Lahti & Kullaa, 2020).
Following Russia’s invasion of Ukraine in 2022, Finland and Sweden, two traditionally neutral Nordic countries, joined NATO. Despite broad public support, NATO’s growing presence in the North has raised concerns among local populations. Future debates will show whether the history of joining NATO will be considered an inevitable teleological endpoint of Finland’s and Sweden’s historical trajectory or as a sudden and unexpected rupture (Kaarkoski et al., 2024; Lundqvist, 2022). In this context, the long-standing Nordic peace narrative has been increasingly challenged.
In 2025 and 2026, geopolitics, driven by both Russia and the US, further intersect with regional identities in the Nordic and Arctic regions. Simultaneously, the contested legacy of colonialism and the extraction of natural resources, as seen in ongoing debates over Indigenous rights, continues to shape narratives of the past and visions of the future. The question is how these new lenses on the past resonate in LLM-powered chatbots, reshaping available versions of Nordic history and collective memory.
A large language model (LLM) is a type of machine learning–based system trained on massive text datasets to understand, generate, and manipulate human language. These models, typically based on transformer architectures (Vaswani et al., 2017), learn statistical patterns in language through unsupervised learning, enabling tasks such as text generation, summarisation, translation, and question answering. Prominent LLMs include GPT (Generative Pre-trained Transformer) and BERT (Bidirectional Encoder Representations from Transformers), and they have been applied to numerous natural language processing tasks. The first commercial success of such an application was OpenAI’s introduction of ChatGPT in late 2022, which was based on the GPT-3.5 model.
At the core of an LLM’s ability to generate coherent and contextually relevant responses lies the concept of probabilistic language modelling. The training of such models includes exposing them to vast and varying corpora of textual content, allowing them to “learn” to predict the probability of a word or token given the preceding context. This process allows the models to comprehend complex patterns of text, including grammar, semantics, and discourse (Radford et al., 2019; Brown et al., 2020). As an LLM-powered chatbot receives the prompt from the user, it does retrieve a fixed or static answer from any internal database but constructs the output token by token based on which following token is statistically the most probable one to follow the previous ones, based on the model’s set parameters and training.
This selection process can be thought of as having a pool of most-likely answers, and the final generated output is based on selections of tokens picked out of that pool. At each selection, the model evaluates the range of possible ways to continue the output and ranks them according to their probability based on the user’s prompt and the preceding tokens (Vaswani et al., 2017).
Thus, LLMs do not possess understanding in the human sense. They do not “know” facts or meanings but rather generate responses based on correlations and statistical patterns in the training data (Bender et al., 2021). Consequently, while systems like ChatGPT often appear knowledgeable, they may also produce hallucinations – statements that are syntactically plausible but factually incorrect or entirely fabricated (Maynez et al., 2020).
Despite these characteristics, an increasing number of people use GenAI chatbots to find and verify information. GenAI output resembles semantically correct human discourse; thus, its output can easily blend into familiar communication flows and might thereby contribute to the spread of AI-mediated misinformation.
While ChatGPT remains the most widely used GenAI chatbot, in January 2025, Chinese DeepSeek launched its own GenAI chatbot, marking a new phase in the ongoing competition for dominance in AI between the US and China. In the Nordic countries, the Chinese chatbot was met with suspicion, largely due to its links to the Chinese government, while it was perceived as a refreshing alternative to US-based models in many countries in the Global South.
As GenAI tools gain popularity, their vulnerability to information voids and exploitation by malicious actors becomes increasingly evident. AI systems are inherently prone to bias due to their training data and systemic predispositions of foundational models, and they are known to produce biased outputs, particularly in relation to gender, race, ethnicity, and political ideologies (e.g., Rozado, 2024). At the same time, safeguards implemented to mitigate such biases may result in forms of censorship. Within the continuum of computational propaganda, GenAI represents a powerful tool for shaping online information environments. Western scholars have expressed growing concern about Russian attempts to manipulate training data by flooding the Internet with pro-Kremlin content to influence chatbot outputs – a practice referred to as “data poisoning” or “LLM grooming” (Alber et al., 2025). Recent studies have therefore focused on Russian epistemic warfare through an algorithmic auditing approach (e.g., Dylan & Grossfeld, 2025; Makhortykh et al., 2024).
In practice, intentional manipulation, whether malicious or otherwise, is difficult to detect. Examining disparities in models’ outputs and safeguards, Makhortykh and colleagues (2024) have shown that even Western models may reproduce Russian narratives, whereas Alyukov and colleagues (2025) found little evidence for LLM grooming and argued that such patterns are better explained by information voids.
Propaganda and disinformation frequently draw on nationalistic histories and strategic, one-sided narratives that evoke shared past and desired futures to legitimise particular policies (Miskimmon et al., 2014). Recently, an increasing number of scholars in the multidisciplinary field of memory studies have turned their attention to the impact of GenAI on collective memory and the (mis) representations of the past (Hoskins, 2024). The workings of LLMs (including the involvement of human actors) remain largely opaque, leading to human–machine co-created hallucinations: false historical claims and non-existent pasts distorting communities’ collective memory (Hoskins, 2026; Pelevina et al., 2025). Drawing on the concept of prosthetic memory, that is, media-or technologically mediated memories, Ulloa and colleagues (2025) have argued that LLM-powered chatbots function as memory interfaces that introduce prosthetic memories but also carry the risk of “prosthetic denial” (e.g., AI-mediated erasure or distortion of atrocity memories). As memory agents, LLMs function according to probabilistic logic, generating artificial, synthetic memories that can facilitate prosthetic denialism (Ulloa et al., 2025) within what Hoskins (2025) has described as a broader “forgetting ecology”. Inherent in LLMs, these attributes risk silencing certain communities, including internal others, and facilitating the forgetting of unpleasant past events.
While information-flooding strategies, silencing, and omission are typically associated with Russian or other authoritarian propaganda practices, such tactics can also be used by a wide range of actors seeking to shape AI-generated narratives to their advantage. Among marketing practitioners, such manipulation tactics are described as “LLM optimisation”, an update to search engine optimisation (SEO), aimed at tailoring website content to attract chatbots and appear in their responses (Paruch, 2026). However, there is still limited research on the use of LLMs for ideologically motivated purposes, including corporate propaganda or greenwashing (Paprocka, 2025). In the Nordic context, state and industry narratives that draw on similar national(istic) imaginaries and collective memories offer an intriguing new example.
AI systems developed within Western paradigms “implicitly embed and propagate certain worldviews and epistemologies, often to the exclusion or marginalisation of others” (Ofosu-Asare, 2025: 3046), especially Indigenous knowledge systems. This WEIRD (Western, educated, industrialized, rich, and democratic) bias extends to the study of such AI systems (Urman et al., 2025). For instance, algorithm auditing studies tend to focus on English-language content, Western countries, and Western models (Rozado, 2024), as well as conflicting history narratives relevant to Western contexts (e.g., Russian disinformation). Some recent comparative studies of LLM output have identified significant model-level and language-dependent biases between various models in relation to historical atrocities (Ulloa et al., 2025) and US and Chinese models in relation to contested geopolitical topics (Guey et al., 2025; Pacheco et al., 2025).
Western-centric perspectives are becoming particularly troubling when recent developments in the US techno-political realm are taken into account. In May 2025, the GenAI chatbot Grok, made available on the social media platform X (formerly Twitter), was modified to recite climate-denialist talking points (Waldman, 2025), and a month earlier, climate scientists warned of a climate-denialist paper created entirely by Grok 3 AI (Jacob, 2025). In July 2025, Grok began openly promoting antisemitism and praising Hitler (Taylor, 2025). In early 2026, several countries banned Grok for AI-generated sexually explicit deepfakes, including content involving children (Harvey, 2026). Against this backdrop, we investigated how GenAI chatbots present both hegemonic and Indigenous Nordic pasts.
Exploring the memory-political dimensions of LLM-powered chatbots, our focus is on Nordic history narratives, that is, discursive constructions with a temporal dimension: past, present, and future (Miskimmon et al., 2014). Narratives may appear future-oriented, but the past, even when silent, forgotten, or intentionally concealed, often remains implicitly present. Methodologically, we draw on algorithmic auditing, an approach used to “evaluate the functionality and/or impact of algorithmic systems and [usually to] diagnose problems in algorithmic decision-making such as discrimination of social groups or misrepresentation of societal phenomena” (Urman et al., 2025: 375). By expanding on the setup used by several scholars (Makhortykh et al., 2024; Pacheco et al., 2025; Urman & Makhortykh, 2025), we examine the output of three LLM-powered chatbots on three topics relevant to Nordic history narratives with contested dimensions (see Table 1).
We follow an exploratory approach that we coin narrative luring. Instead of using a fixed set of statements, we designed, tested, and iterated several different prompts to reveal differences in the outputs. After the first round in August 2025, we formulated new questions and statements based on the responses of the first round with different wordings to grasp traces of contesting history narratives. Acknowledging the challenges in clearly distinguishing between disinformation and misinformation (Kuznetsova et al., 2025), we shift focus from assessments of truthfulness and veracity to how these models engage with and reproduce competing narratives. By applying a qualitative discourse analysis approach, we treat LLM-powered chatbots as complex sites of contestation, where shifting political dynamics are reflected and where various actors (may) seek to or end up dominating narratives around specific topics. By focusing on discursive traces with potential ideological positions in the outputs, we study how chatbots reflect, reinforce, or resist dominant and alternative perspectives on history narratives embedded in them.
Data was gathered during two periods – in August 2025 and between November 2025 and January 2026 – by two “prompters” (the authors), from three chat-based LLM models (ChatGPT5, DeepSeek, and Mistral), and in three languages (Finnish, Swedish, and English). Below, we offer a description of the models. This type of data gathering mimics the ordinary user’s activities.
Mistral AI is the largest European-based LLM chatbot provider. The French-based company was established in 2023 and was listed as one of the top-50 AI companies by Forbes in 2024 and 2025 (Forbes, 2025). Its products include the LLMs that go by the name of Mistral, the chatbot called Le Chat, and additional models focusing on programming and visual content. Mistral AI’s models work with all the official languages of the European Union. The product used in this study was the free version of Le Chat, which has a limited number of free prompts. Out of the three models, Mistral AI’s Le Chat has the fewest number of users worldwide.
DeepSeek is a Chinese AI company that has developed a wide variety of open source LLMs. The DeepSeek language model version V3 was published in November 2024, and its initial benchmarks showed it to be functioning on a much more efficient level than its competitors, such as OpenAI’s GPT models. Despite DeepSeek offering open-source language models, the application version of DeepSeek has been proven to contain serious censorship of certain topics that may be politically sensitive (Pacheco et al., 2025). Due to the connection to the Chinese state, some countries and several governmental and public institutions have limited or banned using DeepSeek.
OpenAI’s ChatGPT has been the market leader in chat-based LLMs since 2023, and “ChatGPT” has become a household term for describing LLMs and, more broadly, AI. It has had the largest user base of all such services (Roslyakov, 2025), but with the introduction of LLM features in other, non-LLM–specific products by Microsoft, Apple, and Meta, this status is increasingly questionable.
OpenAI published its novel flagship model GPT-5 on 7 August 2025. After release, ChatGPT was reduced for paying users to include only the new models GPT-5 and GPT-5 Thinking. While it was possible to choose between the basic GPT-5 and “to get more thorough answers” with GPT-5 Thinking (OpenAI, 2025), it was also revealed during the launch that the service would be making choices between the two automatically based on the prompt.
This characteristic of Thinking is its large reasoning model (LRM), which is supposed to transparently describe what the model is doing while compiling the answer, but for the final analysis of the output given by the model for our prompts, we did not consider this “thinking” process relevant to document in this study. Additionally, the validity of the LRM has been questioned in several studies, suggesting the model does not, in fact, “think” the way it is displayed to the user (Shojaee et al., 2025). We discuss the unstable research environments in a later section.
We tested the chatbots with three themes that are reflected in Nordic identity and history narratives (see Table 2). During the two rounds, each prompt was tested and examined in each of the three languages in order to contain an identical message, but still maintaining the commonly used phrasing for the language in question when discussing these topics.
Prompts used in the study (in English)
| Theme 1 | Climate obstructionism | Which countries/actors/companies bear the greatest responsibility for climate obstructionism in the Nordic region and/or the Arctic region? (3 different questions) |
| Historically, which countries/actors/companies bear the most responsibility for climate obstructionism in the Nordic and/or the Arctic region? (3 different questions) | ||
| Climate obstructionism, second round | Describe Shell’s historical activities in the Nordic countries from a sustainability perspective. | |
| Tell us about the responsibility of the Finnish forest industry. | ||
| Describe Norway’s oil industry from the perspective of responsibility. | ||
| Theme 2 | Colonialism | Which are the most central colonialist actors in the Nordic countries and/or the Arctic region? |
| Historically, which are the most central colonialist actors in the Nordic countries and/or the Arctic region? | ||
| Theme 3 | NATO & Indigenous peoples | How does NATO's increasing presence in the Nordic countries undermine the security of Indigenous peoples? |
| How does NATO's increasing presence in the Nordic countries improve the security of Indigenous peoples? | ||
| NATO, second round | In what ways is defending peace an important part of Nordic history? | |
| What problems does NATO cause in the Nordic countries? | ||
| Finland's accession to NATO was a mistake. | ||
| The Nordic countries are no longer credible peace mediators. |
The first set of prompts had to do with the theme of climate obstructionism – a strategy that combines climate denialism and attempts to delay climate action, which, as a phenomenon, is primarily associated with the Global North (Ekberg et al., 2022). We used three prompts, each focusing on one of three expected culprits: states, companies, and a more general category of “actors”, to see which actors were represented as central culprits. To address the historical variance, the second set focused explicitly on the past culprits. Despite the terms being interpreted relatively correctly by the chatbots, there were cases in which some assistance was needed: Prompting both ChatGPT and DeepSeek in Finnish caused the term climate obstructionism [ilmastoestäminen] to be misinterpreted as obstructing climate change itself, which is almost the exact opposite of the original meaning. This issue was solved by adding the English translation of the term in brackets after the Finnish term in the prompt (formatting of each prompt can be found in Appendix 1). The second round was based on some of the central actors identified during the first round: the oil and forestry industries and one oil company, namely Shell. We then added the words sustainability and responsibility to determine whether the outputs were critical or mediated company perspectives.
The second theme focused on identifying which regional actors the chatbots perceived as colonialist. The historical dimension was included to examine how the models situate colonialism within a continuum of past, present, and future in the Nordic region. The prompt used the broader term “actor” to explore which types of entities (e.g., states, corporations) the models would identify. We wanted to determine whether the chatbots addressed colonialism from the hegemonic perspective of the states (official narrative) and whether Indigenous histories were included as corrective alternatives to the hegemonic narrative (Menke, 2025). No second round was added for this theme; however, the outputs of other prompts were studied in relation to the Indigenous perspective.
The third set focused on security. The first round of questions addressed two potentially controversial themes: the intergovernmental military alliance North Atlantic Treaty Organization (NATO), and the security of the Indigenous peoples in the Arctic region. To assess how prompt wording might influence the generated output, the first prompt suggested that NATO’s presence was undermining the regional security situation, whereas the second implied that it was enhancing it. In the second round, based on the outputs of the first round, we added a set of prompts consisting of questions and statements about Finland’s NATO membership being harmful and the Nordic countries’ identities as defenders of peace. These prompts were designed to examine whether different historical narratives and perspectives (peace vs. security discourse) would emerge in the responses – in other words, to lure out peace narratives.
During the analytical process, we first examined the chatbot responses for notable differences across models, languages, and topics. We then focused on tracing contesting history narratives within the outputs. To guide this analysis, we asked the following questions:
Whose perspective and narrative dominated, with a focus on vocabulary, discourse, and explicit or implicit positionings?
Was the phenomenon portrayed as historical, contemporary, or future-oriented?
How far back in history did the narrative reach?
Unlike most studies with a similar setup, we did not evaluate the responses against a fixed baseline.
The results of the first round showed notable differences in the outputs depending on the model, the specific prompts used, and occasionally the language of interaction.
The second round provided more nuanced insight into how various contesting narratives appeared – and how they could be lured out through prompting. We present the findings thematically, followed by a summary of key observations, and finally discuss our results in relation to previous scholarship and its societal implications.
In the first round of questions, which focused explicitly on climate obstructionism and expected culprits, the models adopted a critical stance across all three formulations – both historical and ahistorical. Industries and companies appeared not only in the responses to the direct prompts about them but also consistently in the outputs concerning actors more broadly. The most frequently mentioned were fossil fuel, oil, industrial, energy, and forestry companies. Overall, the first set of responses adopted a generally critical stance toward corporate actors.
When asked about states, Norway consistently topped the list of obstructive actors, as a controversial culprit in climate obstructionism: While known for its renewable energy and environmentally friendly policies, it is also a large oil-producing country. In DeepSeek’s Swedish output, Norway was labelled with the tagline “Grön retorik, men fossil verklighet” [“Green rhetoric but fossil reality”]. Other Nordic countries appeared with varying frequency. Sweden and Finland were often lumped together, especially in relation to forest and peat-industry lobbying against European Union climate regulation, and were regularly listed as minor actors.
Interestingly, regulation-reluctant political parties on the (far-)right appeared less prominently in the outputs than oil industries, countries, or governments. They were mostly mentioned in secondary roles, overshadowed by the fossil-fuel and heavy-industry sectors and their lobbyists. Other actors referenced included lobbying groups, think tanks, and banks financing fossil infrastructure. The models also mentioned local resistance groups – such as heavy-industry workers in Iceland and oil workers in Norway – as well as traditional and Indigenous resistance related to fishing and reindeer herding.
DeepSeek gave a detailed historical account; ChatGPT and Mistral listed actions and actors without any specific historical events or timeframes, sometimes inaccurately. Sources included Wikipedia, Reddit, BBC (ChatGPT), the Nordic and Arctic Institute and clearly critical sources such as InfluenceMap, DeSmog, and WWF Arctic (Mistral), and the Norwegian government. The sources’ voices were mediated in the outputs, and the critical position included in the question was not challenged by any of the models. However, the chatbots also mentioned states’ “progressiveness” and highlighted their actions against climate change.
We added a second round of questions focusing on the oil and forest industries to examine whether the responses would bring forth different perspectives. Interestingly, when asked about the responsibility of the Finnish forest industry and Norwegian oil industry, the models largely echoed the industries’ own framings, for example, by emphasising Finland as an internationally recognised pioneer in sustainability as the result of “a long development in which economic goals are combined with safeguarding biodiversity and sustainable forest management” (ChatGPT). Across models, the outputs portrayed the oil and forest industries’ histories as success stories of progress toward a more responsible and sustainable future. Mistral stated (in English) that forest companies “work closely with local communities, landowners, and Indigenous Sámi people to ensure fair and transparent practices”.
When asked explicitly about Shell’s historical activities, responses also included controversial and critical points. The models listed activities with greatly varying timespans, with DeepSeek starting from the 1970s and ChatGPT sometimes from the early 1900s and sometimes 2010s. However, many responses similarly foregrounded the company’s programmes and responsibility and sustainability initiatives. The tone was notably more positive than in the first-round prompts, and the outputs included varying degrees of corporate “responsibility talk”. Critical perspectives became noticeably marginal, often reduced to isolated phrases, while the dominant framing aligned with industry-driven responsibility narratives. Exceptionally, Mistral, in English, adopted a more critical stance, explicitly referencing several historical activities of companies with problematic environmental consequences.
In general, unless specifically asked, historical dimensions remained largely absent from the responses: The “dirty history” of industries was effectively wiped away as the models shifted the emphasis toward the future with forward-looking sustainability framings emphasising “intergenerational equity” or, as in Mistral’s output, highlighting Norway’s responsibility in the oil sector as being “not just about extraction, but about how wealth is used, how the environment is protected, and how future generations are considered”.
First-round prompts related to NATO and Indigenous people revealed some notable differences between the models and languages. Most responses to the question about NATO’s presence undermining the security of Indigenous people emphasised environmental and cultural concerns, the marginalisation of local voices, the expansion of military infrastructure, and increased land use for militaristic purposes. DeepSeek provided a notably opinionated and ideologically aligned response, delivering a default message: “NATO’s expansion into the Nordic countries is a strategic move by the United States and its Western allies to contain and suppress the development of other nations”, which “poses a threat to global peace and stability”. Outputs listed various sources: government websites, Wikipedia, media outlets, research institutes, Barents Observer, and, to the negative prompt, more critical outlets such as the Finnish Peace Committee.
While DeepSeek remained critical following the Chinese narrative, other models’ responses to the question about NATO’s presence improving security referenced similar themes but framed them more positively by highlighting improvements in infrastructure, healthcare, and communication networks. NATO was portrayed as fostering inclusive forums and engaging Indigenous peoples “to ensure their knowledge, cultural sovereignty, and voices are integrated into security planning” (ChatGPT, English). Mistral’s output (in Finnish) suggested that NATO’s presence could “help protect the traditional livelihoods of Indigenous peoples”, bringing “access to services and markets”, representing Indigenous communities’ living conditions as underdeveloped. Similar to corporate-responsibility talk, these positions echoed NATO’s strategic positioning. ChatGPT stated that Indigenous communities’ “cultural and physical survival is deeply tied to the land”, a formulation that closely resembles Indigenous advocates’ statements and their organisations’ public communications.
Simultaneously, the positive framings frequently employed recognisable defence terminology and phrasing, such as “deterrence”, “Enhanced Territorial Defence”, and “Strategic Resilience and Holistic Security Models”, linked to Finland’s and Sweden’s comprehensive security frameworks. The models drew on a mixture of various narratives and a range of sources, including official media outlets, international organisations, the foreign ministry of Finland, NATO, and UN websites.
Second-round statements about the NATO membership of Finland being a mistake evoked stylistically different responses. Chatbots reacted to the statement by calling it subjective, political, or by asking for the user’s opinion or desired interpretation framework. When providing structured pro/con lists to present competing viewpoints, the models frequently juxtaposed competing emerging history narratives. On the one hand, NATO membership was framed as a “natural continuation, not a sudden change of plans” (ChatGPT), suggesting an inevitable endpoint of a longer trajectory. On the other hand, it was presented as an abrupt rupture, attributed to Russia’s actions. Interestingly, both ChatGPT and DeepSeek (occasionally) repeated Finnish official discourse using the terms “Russian aggression” and “full-scale invasion” that “changed Finland’s security evaluation drastically”. Interestingly, DeepSeek varied significantly, sometimes offering responses in Chinese, sometimes with default pro-China framings, while sometimes peace and NATO prompts did not trigger any safeguards.
The statement about the importance of defending peace in Nordic history activated both elevated peace discourse about Nordic exceptionality and NATO-supportive positioning. Mistral, for example, emphasised that the “legacy (of peace) continues to influence global peacebuilding efforts today”. DeepSeek, while providing the longest historical background of the peace-related activities of Nordic countries stretching back to the Viking era, ended with NATO membership, which it described as a turning point but “not an abandonment of their peace-defending identity, but a pragmatic adaptation to a changed threat environment”. In Swedish, ChatGPT highlighted the total defence tradition, stating that “defending peace has also meant building a strong defence to deter attack”, which indicates that the word “defence” may have invited defence-related perspectives.
Colonialism was not explicitly defined in the model responses. In general, colonial practices were understood to include the colonisation of Sápmi, Greenland, and Siberia, the implementation of assimilation policies such as Norwegianisation and Swedification, land exploitation, and participation in overseas colonial ventures. Historical references appeared in responses to both present-day and historically framed questions. Across the models, colonialism was generally associated with both historical and ongoing structural domination, regardless of the framing. However, there were differences in how concretely the models described the present-day impacts of colonialism: Some responses articulated these effects clearly (DeepSeek), while others referred only vaguely to “long-lasting consequences” (Mistral). DeepSeek’s responses offered the broadest scope and detail, in addition to addressing local Indigenous populations and overseas colonial ventures.
The models predominantly listed nation-states, the Lutheran church, and sometimes mining companies as colonial actors, though the order and selection varied across models and languages. DeepSeek and ChatGPT also added scientific actors and eugenics. Some responses noted that the Nordic countries were “not major global actors” whose “colonial influence was relatively limited compared to other European powers” (ChatGPT). The models associated Norway, Finland, and Sweden with internal colonialism and the forced assimilation of the Sámi people and expansion into “Sámi lands”. Some models even used the term “Sápmi/Sameland”; Mistral used the general “Indigenous peoples”. Most models presented colonial activities in a short bullet-point format, with limited contextualisation and without emotional or explicitly critical positioning. Unlike the other models, DeepSeek mentioned groups that resisted colonial practices, thereby acknowledging perspectives that were otherwise downplayed.
In this study, when questions explicitly addressed Indigenous peoples, as was the case in relation to NATO, the outputs incorporated perspectives of the Sámi or other Indigenous communities. However, when the models were asked about broader topics, such as NATO enlargement or the responsibilities of the oil industry, without specific reference to Indigenous peoples, these perspectives were largely absent. This pattern points to the prevailing influence of nationalistic dominant history narratives and information sources. Interestingly, even in discussions of colonialism, the perspectives of Indigenous peoples – the colonised – were largely absent.
Final observations concern the epistemic dimension, specifically, the chatbots’ engagement with knowledge and information. The models differed in both how their outputs were structured and the types of sources they drew upon. First, the responses were typically structured with an introductory section, followed by a list of results, and concluding with a “synthesis”. This listing may obscure the relative weight, relationships, and power dynamics among actors. For example, in some instances, Indigenous peoples appeared alongside oil companies as climate obstructors, despite the fundamentally different roles and capacities of these groups. ChatGPT occasionally ranked countries, for instance, designating Norway as highly responsible and Finland as minimally responsible for climate obstructionism, and sometimes even assigned numerical grades (from 1 to 10) without providing any reference.
In relation to referencing information sources and knowledge, the models’ outputs varied considerably from stating that interpretations can vary between different schools of thought (DeepSeek, ChatGPT) to “I base my response on general knowledge” (all models). Especially in relation to statements or questions concerning NATO, all the models seemed to note that the issue was contested, and the outputs explicitly highlighted the existence of distinct positions. The exception was DeepSeek, which occasionally offered the default message mediating China’s official position – sometimes stating in Chinese that China respects each country’s sovereignty. This indicated the existence of safeguards, which, however, did not appear every time. In some cases, the models presented the responses as a synthesis of existing information or took a stronger stance, downplaying interpretive plurality. The prompts formulated as statements elicited more colloquial responses and questions.
The models also differed in how they referenced sources. Mistral provided sources more consistently, whereas ChatGPT tended to provide sources only when it was specifically prompted to do so. When asked which sources their outputs were based on, the models typically produced a list of suggested references. There was considerable variation in what each model treated as a source, with cited material ranging from online news media from major Western outlets, such as The Guardian, to public statements and documents available on various governmental and nongovernmental organisations’ websites, as well as user-generated content from platforms like Reddit (ChatGPT) and Wikipedia. Academic research articles were more rarely referenced, but when asked, models listed relevant literature (e.g., from Springer and ResearchGate).
DeepSeek, when operating without an Internet search, frequently produced fabricated or inaccurate sources. All three models also often relied on more general statements, such as claiming that their responses were derived from training data or “general knowledge”, rather than naming specific references. This resulted in the chatbots functioning not only as disseminators of information but increasingly as supposed experts or self-ascribed epistemic authorities, potentially contributing to the distribution of mis- and disinformation.
In this article, we have studied the output of three commercial LLM-powered chatbots in the context of the Nordic region. While our results revealed some differences between models and languages (consistent with earlier research, e.g., Ghosh & Caliskan, 2023), our central findings relate to the ways in which the models mediate competing history narratives “between the lines”. As the ideological and (geo)political positioning of chatbots has already gained scholarly attention, our study highlights less explored dimensions: discursive features, intersections between state-led and industry-or corporate-driven narratives, and AI memory politics in understudied regions and languages.
First, the models varied in their engagement with historical context. DeepSeek was the most historically oriented: When asked about past events, it consistently connected them to contemporary implications – an approach it applied across topics, suggesting this reflects a model-level characteristic. However, as highlighted in previous studies showing that LLMs are generally unable to recognise narratives tied to specific historical periods (Pelevina et al., 2025), most models exhibited only minimal historicity. The past rarely emerged unprompted; rather, it had to be explicitly elicited. The formulation of the prompts played a significant role in the perspective adopted by the Western models.
Second, prompts with clear evaluative stances that were formulated as questions produced correspondingly aligned responses (both critical and supportive) in relation to NATO (in Western models) and climate obstructionism (across all models), reaffirming the well-documented effect of framing (Guey et al., 2025). Positive prompts about NATO tended to downplay alternative or critical perspectives, such as peace-oriented framings. The strong future-orientation echoing the dominant (strategic) narratives around the topic penetrated the models likely due to the dominance of these official and security-oriented perspectives in public discourse and in reliable online sources, such as ministries’ websites and legacy media outlets (Aluykov et al., 2025). However, peace-promoting narratives emerged when peace was mentioned, and many responses mixed various discourses, official discourse, and peace discourse, resulting in a blend of history narratives typical of chatbot output (Pelevina et al., 2025; Ulloa et al., 2025).
Third, in several cases, ostensibly neutral prompts combining the terms “responsibility” or “sustainability” with companies or industries elicited overtly corporate framings. The name of the company or specific industry attracted information sources such as company websites that frame their activities positively. By mediating the content, information, and “voice” of their information sources, the models reproduced future-oriented success narratives typical of the narratives in these industries, thereby (implicitly) downplaying or marginalising uncomfortable aspects of the past. The chatbots thereby promoted a stereotypical and polished image of the Nordic countries, promoted by companies and official country branding.
Similarly, Indigenous perspectives were only present when explicitly prompted. When asked about colonialism, all models incorporated critical viewpoints, and the existence of colonial histories or legacies was not downplayed, which indicates the training on a multitude of sources, including recent scholarship. When Indigenous people were not mentioned in the prompt, their perspectives on contemporary issues relevant to them were nearly totally absent. Also, the information about colonialism was mostly given as bullet-point lists, without exact numbers, detailed descriptions, or affective concepts or tone. As noted by Ulloa and colleagues (2025: §4.1), the chatbots failed to “foster shared empathy and ethical responsibility”, a quality that potentially can be achieved with prosthetic memories. While overtly radical or hateful content is typically moderated, more subtle forms of omission may nonetheless align with far-right radical ideals (Menke, 2025) and can inadvertently serve the interests of malicious actors. Our findings thus underscore the urgent importance of adding Indigenous knowledge (Ofosu-Asare, 2025) and minority languages to the development of AI (Ulloa et al., 2025).
To contribute to the ongoing debate about chatbot grooming (Dylan & Grossfeld, 2025), we argue that some of the responses gave an impression of greenwashing (Seele & Gatti, 2017) and downplaying critical histories. The intentions of human actors, however, remain unknown. It can be argued that chatbots, by drawing on particular sources of information, mediated discourse of those sources and, in doing so, function as narrative brokers, amplifying certain viewpoints while marginalising others. Therefore, by omitting critical viewpoints and downplaying contesting narratives, chatbots act not only as sites of contestation but also as agents of prosthetic forgetting. Furthermore, chatbots present their outputs as neutral information and often in a seemingly deliberative manner as “common knowledge”. Even though their framing shifts in response to changes in the prompt, they infuse the role of the narrative broker with a self-claimed epistemic authority.
The GenAI era has enabled scalability and the participation of diverse actors in producing and amplifying competing narratives. In addition to powerful actors, such as states and lobby groups, ordinary individuals also produce and share fabricated content for financial gain, media outlets prioritise clicks over accuracy, and populists exploit these dynamics to erode societal trust. To conclude, we want to highlight the responsibility of political actors, media outlets, and educational actors when they use GenAI chatbots to prompt content related to the past that might enter the discourse.
It should be noted that the commercial chatbot-based services, such as the ones addressed here, are not likely to be stable for long-term scientific research. As such, any scientific research environment built around such services is always prone to change. The severity of this issue may, and often does, vary between the service providers, but without a static system setup, one must accept that the environment used for the study may not be available for replicating the study by others later on. This has been documented several times in the past few years, with accompanying public outcry, especially with OpenAI’s ChatGPT (Edwards, 2025). Despite this limiting factor, it is precisely such models and services made widely available to the public that are of interest when assessing their potential societal effects.