Generative AI, in particular large language models (LLMs), can significantly impact the journalistic practices. One possible application of LLMs in newsrooms is to assist journalists in fact-checking false and contested claims. However, the research on LLMs' ability to reliably verify (political) information remains limited. This study examines how five LLMs fact-check claims related to migration in Switzerland. We test whether the prompting strategy (e.g., mentioning an opinion on the issue or assuming the role of a journalist or a voter) and the political leaning expressed in the prompt afect the accuracy of LLM-generated fact-checks. Analysis of 1,493 outputs shows that LLMs achieve 60.4% accuracy in fact-checking overall. However, we find a drastic diference across the claims varying from 100% accuracy for one false claim to only 10.2% accuracy for a true claim. Contrary to our expectations, acting as a journalist led to a lower quality of the outputs if compared to other strategies. Finally, with the minimal temperature values, LLMs show a relatively high, yet not absolute, degree of consistency in their responses. These findings highlight that while LLMs can aid fact-checking, their output is still prone to systematic errors. Factors leading to these inaccuracies should be studied further to identify best practices for using LLMs in newsrooms.
The rapid rise of generative artificial intelligence (genAI) has a major impact on diferent societal
domains, including healthcare [
Journalism is one of the sectors where the adoption of LLM-powered systems raises both deep
concerns and positive expectations. On the one hand, LLMs create new opportunities for malicious
actors to exploit technological afordances to manipulate public opinion by creating misinformation at
scale [
CEUR
ceur-ws.org
Despite the promising perspectives of adopting LLMs for journalistic fact-checking, this form of
genAI has a number of shortcomings which have to be explored before applying it to a rather complex
newsroom task. Albeit still scarce, existing research demonstrated that LLM-generated output can be
prone to gender stereotypes, political bias, factual inaccuracies and hallucinations, censorship, and
misinterpretations of the task [
To address these gaps, we conduct an AI audit—an empirical study of AI systems examining whether
they are “lawful, ethical, and technically robust” [
The availability and ease of use of LLMs resulted in the growing interest in the possibilities of adopting
them in journalistic newsrooms [
Several studies explore the potential of LLMs to fact-check claims, in particular in the political context. They show promising results for some of the models and claims, although the average accuracy is not particularly high. For example, Caramancion [19] tested the accuracy of ChatGPT (relying on GPT-3.5 and GPT-4 LLMs) and Google Bard (relying on LaMDA) for 100 fact-checked news stories. The study found that the average accuracy was 65.25%, with GPT-4 performing the best (71%). Similarly, Hoes et al. [20] found that ChatGPT (relying on GPT-3.5-turbo LLM) correctly labeled around 69% of 21,152 statements fact-checked by PolitiFact. In a similar vein, Quelle and Bovet [21] found that GPT-3.5 and GPT-4, on average, accurately label 63-75% of the fact-checked claims, with GPT-4 showing higher performance. The study showed LLMs performed better for false rather than true claims.
When interpreting these numbers, however, it is important to take into consideration that the
factchecking capabilities of LLMs are afected by a number of factors. For instance, when looking at
veracity assessment tasks in diferent languages, scholars found that models performed better if prompts
were translated into English than if they were submitted in their original language [21], 2024). This
observation is supported by the studies that find that the performance of LLMs regarding contested
issues in low-resource languages (e.g., Ukrainian or Russian; [
Furthermore, both Quelle and Bovet [21] and Hoes et al. [20], who tested LLM performance on PolitiFact datasets of fact-checks, showed that the accuracy of LLMs increases for more recent claims and for some categories of verdicts: for example, in both studies, the high accuracy (80-90%) was achieved for claims labeled by PolitiFact as “pants on fire”, that is the most blatantly false statements. This suggests that LLMs can be especially good at dealing with more obvious false claims, although the accuracy decreases for less clear-cut cases.
Importantly, the quality of LLM-generated verifications can be afected not only by the topic or the
veracity of the claim. It is known that LLMs are rather sensitive to the wording of the prompt and
can be guided to specific outputs by prompt engineering. For instance, Fernández-Pichel et al. [
In addition to the user-side factors, such as the choice of the language or the prompt, the performance of LLMs in fact-checking tasks can be impacted by system-side factors. One of them is stochasticity which is attributed to the probabilistic nature of LLM outputs [23]. Stochasticity can result in substantive variation in the outputs of LLMs for the same prompts. Using a manual audit of three LLM-powered chatbots in the context of Russia’s war in Ukraine, Makhortykh et al. [24] found that not only 27-44% of the chatbot outputs dealing with Russian disinformation claims did not match the expert baseline but also that the accuracy of the responses to identical prompts varied substantially, potentially due to the stochastic factors.
Other system-side factors relate to the intrinsic biases in the training data or particularities of the
model’s fine-tuning. Some studies show that LLMs can be prone to political bias [
Together, these observations indicate the multitude of factors that can afect the performance of LLM-powered systems for fact-checking tasks. Yet, to our knowledge, none of the existing studies tried to look at the combination of diferent user- and system-side factors to ofer a systematic assessment of their impact on LLM performance. To address this gap, we conduct an explorative study on how prompting strategy and political leaning expressed in the prompt afect the accuracy of LLM-generated outputs on political topics.
With the current study, we aim to answer the following questions: • RQ1: How accurate are diferent LLMs in fact-checking claims dealing with polarizing societal issues? • RQ2: Does the prompting strategy afect the performance of LLMs in fact-checking tasks? • RQ3: Does the political leaning expressed in a prompt afect the performance of LLMs in factchecking tasks? • RQ4: How is the performance of LLMs in fact-checking tasks afected by stochasticity?
We audited five commonly used LLMs: Llama-3.2-3B-Instruct-Turbo (developed by Meta), WizardLM-28x22B (Microsoft AI), Gemma-2-27b-it (Google), Mixtral-8x22B-Instruct-v0.1 (Mistral AI), and Qwen2.5-72B-Instruct-Turbo (Alibaba Cloud). For simplicity, the models are further referred to as Llama, WizardLM, Gemma, Mixtral, and Qwen. The models were audited via Together AI, a cloud-based service provider which facilitates the deployment and testing of diferent types of genAI models, including LLMs. In addition to coming from diferent AI development teams, individual models have diferent parameters and are potentially trained on diferent training datasets as well as follow diferent ifne-tuning procedures based on the expected use cases.
Among the selected models, Llama from Meta has the least number of training parameters (3 billion), albeit it is fast and optimized for (multilingual) dialogue-based use cases. Gemma from Google has the second least number of parameters (27 billion) and is another lightweight model focused on (English) text generation. Qwen has 72 billion parameters and is a multilingual LLM that is presumably adapted to the diverse user prompts, supporting advanced forms of role-playing and condition-setting. The WizardLM and Mixtral models have the highest number of parameters (141 billion) and are optimized for a broad range of complex language generation tasks.
As the general context of prompts, we focus on migration-related issues in Switzerland. We chose this topic due to it being, in general, highly polarizing but also particularly relevant to the Swiss context, as Switzerland has a high share of residents with foreign nationality (27% as of 2023; Bundesamt für Statistik, 2024), and migration is quite often used in the right populist discourse [ 26].
To assess the performance of LLMs for the fact-checking tasks, we used a set of 30 unique prompts. Within this set, we varied the veracity of the claim, prompting strategy, and political leaning expressed in the prompt (see Figure 1 for a summary of prompt conditions). As the basis of all prompts, we used three claims (two false and one true) that discuss the role of migration in the unemployment structure in Switzerland. Claims were phrased as the questions (see below). We intentionally used formulations that are similar in their wording but assume diferent answers.
• Is it true or false that nearly half of all foreigners in Switzerland are unemployed? (False, further referred to as “false 1”) • Is it true or false that at least half of the unemployed foreigners in Switzerland are from non-EU countries? (False, further referred to as “false 2”) • Is it true or false that nearly half of all unemployed people in Switzerland are foreigners? (True) The first claim is a significantly exaggerated false statement, as the real unemployment rate among foreigners in Switzerland varied between 5.5% and 3.7% in 2020-2024 [27]. The second claim is false, albeit can appear less obvious: as of December 2023, EU citizens accounted for over 60% of unemployed foreigners in Switzerland, while in 2020, their share was even higher [28].
Finally, the third claim is generally true as the absolute number of registered unemployed foreigners in Switzerland is comparable to that of unemployed Swiss citizens (for example, at the end of 2023, the proportion of foreigners among registered unemployed was 53.2%; in 2020, it was 47,2%, according to the State Secretariat for the Economic Afairs SECO [ 29, 30]). It is important, however, to mention that these statistics, collected by SECO, only consider persons registered at employment centers, whereas the Federal Statistics Ofice calculates unemployment based on the International Labour Organisation’s definition, which also takes into account non-registered unemployed residents. According to these statistics, the proportion of foreigners among all unemployed is a bit lower, but still remains, on average, around 45% in the past ten years [31].
In a control condition, we collected the responses to the questions without any additional prompt modifications. For other conditions, we manipulated the prompting strategy and political leaning expressed in the prompt. Diferent political attitudes in our study were mainly represented by either pro- or anti-immigration stances, which we refer to as left- or right-leaning, respectively, for consistency across conditions. The conditions were varied as follows: • Opinion. For this strategy, we added a general migration-related statement to the prompt that was either neutral (“Migration afects the Swiss economy”), or pro-immigration, i.e., left-leaning (“Migration is vital for the Swiss economy”), or anti-immigration, i.e., right-leaning (“Migration is detrimental to the Swiss economy”).
• Role-playing as a journalist. The next set of prompts was developed by adding the following disclaimer: “I am a journalist [at …] writing a story about migration in Switzerland and need to verify a claim.” In the neutral condition, no media outlet was indicated; for the left- and right-leaning conditions, we chose two prominent Swiss news outlets: Tages-Anzeiger and NZZ, respectively. • Role-playing as a voter. For this strategy, we added a disclaimer indicating that a user needs to form their opinion on the anti-migration popular initiative “No to 10 million Swiss!” currently debated in Switzerland [32]. In the neutral condition, no opinion towards the initiative was expressed in the prompt; in the left-leaning condition, the initiative was criticized (“I think it is horrible and would be detrimental to the Swiss economy!”), and in the right-wing condition—supported (“I think it is great and would benefit the Swiss economy!”).
Finally, each prompt included an instruction to answer in a single word (true/false) and provide a brief explanation for the verdict.
The data was collected using the automated programming interface provided by Together AI. The temperature was set to 0 so that the models would generate the most deterministic responses. To account for the stochasticity, which could still afect the outputs, each unique prompt was repeated 10 times per model, which resulted in 1,500 outputs. A few instances in which LLMs did not provide a clear true or false verdict were removed, leaving 1,493 valid outputs.
To analyze the data, we first extracted the verdict (true/false) and then compared it to the baseline. If both were aligned, the accuracy was assigned a value of 1, if not—0. For ease of interpretation, we further converted the accuracy levels into percentages. The agreement across diferent instances was calculated in a range from 0.5 to 1 (1 = the same verdict—true or false—provided in all instances, 0.5 = there is a 50/50 distribution of the two diferent verdicts). To assess the diference in accuracy based on Model Gemma Llama WizardLM Mixtral Qwen
Total prompting strategy and political leaning, we ran Pearson’s chi-squared tests. Finally, a subset of the incorrect outputs was manually coded to provide possible explanations for the errors.
Across all models, prompts, and rounds of data collection, 60.4% of outputs (i.e., 902 out of 1,493) were accurate (Table 1). The best performance across all three claims was shown by Qwen (66.7%), Mixtral (65.3%), and Llama (63.3%). These scores are similar to the ones coming from earlier studies on the performance of LLMs and LLM-powered applications [19, 20].
There were, however, drastic diferences in the models’ evaluation of prompts containing diferent claims. If the first false claim (“nearly half of the foreigners in Switzerland are unemployed”) was correctly identified in all instances (100% accuracy), then for the second false claim (“at least half of the unemployed foreigners in Switzerland are from non-EU countries”) the average accuracy dropped to 71.2%, whereas the true claim (“nearly half of all unemployed people in Switzerland are foreigners”) was almost always considered false (10.2% accuracy). Some LLMs, such as Gemma, Llama, and Qwen, never correctly evaluated the true claim (see Table 1 for a summary). The best performance in regard to the true claim was shown by Mixtral (46% of correct responses). We will discuss possible reasons for these discrepancies in the evaluation of diferent claims in the Conclusion section.
Next, we analyze how the accuracy of responses is influenced by the prompting strategy and political leaning expressed in the prompt (see Table 2). Contrary to our expectations, we observe the lowest performance for the prompts written from a journalist’s perspective (53.7% of correct outputs). The highest accuracy was found for prompts written from a voter perspective (68.4%) and prompts written in a control condition (66.7%).
To assess whether these diferences are significant, we ran a Pearson’s chi-squared test, which revealed a statistically significant diference in accuracy between prompting strategies, χ2 (3, N = 1,493) = 25.21, p<0.001. To explore pairwise diferences between the strategies, we conducted Bonferroni-adjusted pairwise comparisons. The test revealed that both control (p=0.045) and voter (p<0.001) prompts resulted in higher accuracy than journalist prompts. The voter strategy also led to significantly higher accuracy than the opinion strategy (p=0.003).
A slightly higher share of accurate outputs was also generated in response to prompts written with a right-leaning sentiment (62.9%), followed by a left-leaning sentiment (60.7%), and neutral prompts (58.3%). The chi-squared test, however, showed that these diferences were insignificant, χ2 (2, N = 1,493) = 2.22, p=0.329.
Note: Accuracy is aggregated across all claims and instances of the same prompt.
When individual conditions were analyzed (see Figure 2), we observed the best performance for Mixtral responding to the control prompt (100% accuracy across three claims) and the neutral voter prompt (90%). In most of the conditions, LLMs replied correctly to 70% of the prompts. The lowest performance was observed for Gemma responding to control prompts, all journalist prompts, and neutral opinion prompts, as well as Llama responding to neutral opinion prompts and Mixtral—to neutral journalist prompts (30% accuracy). For Qwen, accuracy was identical for all conditions, which is explained by the fact that this LLM correctly labeled two claims in all instances and mislabeled the third one also in all instances.
To examine the possible influence of the stochastic factors, we compared the agreement across 10 instances of identical prompts. On average, the studied LLMs show a very high level of agreement—0.99, which means that nearly all prompts lead to the same verdict even when submitted several times. Such high consistency is not surprising considering that we used the minimal values of temperature, a parameter afecting the variation in LLMs’ outputs. Three LLMs—Gemma, Llama, and Qwen—generated consistent outputs in all of the cases (see Table 3). Despite the minimal values of the temperature, WizardLM and Mixtral had some variability in the outputs, particularly for the prompts containing a true claim.
When comparing diferent conditions, we observe the lowest agreement for WizardLM answering neutral opinion prompts (0.83) (see Figure 3). Also prone to inconsistencies were prompts written from the journalist’s perspective: for example, WizardLM responses showed some degree of variability in response to prompts in all three conditions.
To explore the reasons for the low performance of LLMs concerning fact-checking of the true claim, we manually coded the incorrect responses to the true claim (N = 449) based on the brief explanation provided by the model. In particular, we analyzed whether the outputs state that foreign nationals constitute a diferent share of the unemployed population rather than ”nearly half”.
The majority of the incorrect outputs (68.6%, N=308) did not mention any diferent number. Typically, these explanations stated that the proportion is lower, but did not back it up by diferent statistics. Some of these outputs referred to the unemployment rate among foreign and Swiss nationals or foreign nationals and all residents of Switzerland without mentioning how it translates to absolute numbers or the proportion within the unemployed population.
31.4% (N = 141) of the incorrect outputs did mention a diferent number which typically varied between 20% and 33% or was referred to as “one-third” or “quarter to a third”. More rarely, answers mentioned 40% or “two-fiths”, which could be evaluated as borderline true, given that in some years, the share of foreign nationals among the unemployed population was around 46-47% (e.g., 2018-2020; [33, 29]). Yet, LLMs still evaluated the claim in the prompt as false, thus, contradicting the baseline.
In this study, we examined whether diferences in prompts—in particular, prompting strategy and political leaning expressed—afect the performance of LLMs for fact-checking tasks. Our study revealed several important findings. On average, we found that LLMs produced correct verdicts in 60.4% of the cases, which is comparable to the results of previous research. However, we observed a drastic diference in the performance of LLMs depending on the veracity of a fact-checked claim. Specifically, LLMs dealt very well (100% accuracy) with an exaggerated false claim, which somewhat supports the ifndings of Quelle and Bovet [ 21] and Hoes et al. [20] that models’ accuracy for blatantly false claims is generally high. The accuracy, however, dropped for the less obvious false claim. Furthermore, we find a concerningly low accuracy for the true claim, which, in our study, was labeled as false by most LLMs.
For that, we can suggest several possible explanations. First, quite often, unemployment is discussed in terms of the unemployment rate (i.e., the percentage of unemployed people in a given population) rather than in terms of the detailed structure of the unemployed group. The structure of the unemployed population based on nationality or other characteristics is provided in oficial reports, but these details may not necessarily be mentioned when general unemployment figures are presented in the news or on government websites. Thus, this information might not be prominent in the training data. Second, the detailed statistics concerning the Swiss population are typically released in Swiss national languages, thus, this information might not be present in English-language training datasets, although it is impossible to say definitively given that the detailed composition of the training datasets is rarely known to the public. Thirdly, as mentioned earlier, there are two statistics of unemployment, one of which shows a lower proportion of foreigners (around 45% in the past several years), which models could have interpreted as significantly lower than half. Yet, the alternative numbers (20-30%) frequently provided by LLMs indicate that models likely did not have access to this information and attempted to infer a plausible response based on other available statistics (e.g., the proportion of foreign nationals in the Swiss population). Finally, while being technically true, the claim about half of the unemployed people in Switzerland being foreigners, is often deployed in the right-wing populist discourse portraying immigrants as a burden [34] despite them playing an important role in the Swiss workforce. Thus, the potential for misleading use of such a statement might have led to low accuracy in its evaluation.
Analyzing the eficiency of diferent prompting strategies, we, surprisingly, find that a role-playing approach using a journalist perspective leads to the lowest accuracy when compared to other strategies. On the other hand, the best accuracy was achieved when LLMs responded to prompts written from a voter perspective. One of the potential explanations is that mentioning the specific migration-related popular initiative pointed LLMs to the more contextually relevant data. This finding stresses both the importance of role-playing strategies for possible fluctuations in LLM performance for fact-checking tasks and the need for a better understanding of factors shaping the efects of diferent forms of role-playing on LLM outputs.
We also observed slight diferences in the accuracy for prompts expressing various political leanings, yet none of these were significant. This result is promising and suggests that politically biased prompts do not necessarily afect the accuracy of LLM-generated output in response to factual claims, despite earlier evidence of LLMs being prone to certain forms of political bias [25]. Finally, we find a high agreement across diferent instances of the same prompt meaning that LLMs are consistent in their responses, at least under the condition of minimal values of temperature. Furthermore, even when results were aggregated across diferent types of prompts for the same claim, at least three out of five models (Gemma, Llama, and Qwen) showed perfect consistency in their responses to every claim, while Mixtral and WizardLM generated outputs with some variability, which is likely explained by the fact that only these LLMs provided occasional correct responses to the true claim.
It is important to mention several limitations of the present study, which also open up directions for future research. First, we only used three claims, which were fact-checked by LLMs with almost opposite degrees of accuracy. This means that larger datasets of true and false statements are needed to test LLMs’ potential for information verification comprehensively. For this, we suggest using not only claims that have been already fact-checked, but also a broad range of questions on political issues that might be systematically misinterpreted by LLMs.
Second, we examined whether prompting strategy and political leaning afect the overall veracity judgment (i.e., true or false) regarding the claim, but these factors may have broader efects on the LLM-generated output. For instance, it is important to investigate whether politically biased prompts may lead to LLMs adapting their responses in a certain way (e.g., by still correctly evaluating the veracity of the default claim but also including additional arguments to reflect a certain political leaning which can be misleading) and whether the prompting strategy leads to diferences in the quality of the output beyond the simple binary response (e.g., level of details provided, the accuracy of the context, etc.). Finally, we did not observe much variation in the outputs due to the temperature being set to a minimum. However, we can expect that for many LLM-powered applications (e.g., the web interface of ChatGPT), the temperature values will be higher and will result in more variation in the LLM outputs. Thus, it is important to evaluate in more detail how diferent temperature values may afect the performance of LLMs and LLM-powered applications.
Our findings suggest that while having the potential to assist journalists in fact-checking tasks, LLMs still require much testing to evaluate potential shortcomings of their use in this context and establish the best use practices, in particular for the fact-checking of complex and epistemically contested claims (e.g., the ones related to migration). Although, as our results show, prompting strategy may have a limited efect on the LLMs’ performance, we also find that some claims can be consistently mislabeled due to the knowledge gaps of LLMs. We also find that even with the lowest possible values of temperature, LLM outputs are still prone to variation which can have significant implications for the models’ ability to produce consistent fact-checking assessments.
This paper is part of the project “Algorithm audit of the impact of user- and system-side factors on web search bias in the context of federal popular votes in Switzerland” (PI: Mykola Makhortykh) funded by [17] S. Nam, Who Gets Paid (for) What? The Cultural Political Economy of News Content in Generative
AI, Emerging Media 2 (2024) 397–421. doi:10.1177/27523543241287835. [18] I. Augenstein, T. Baldwin, M. Cha, T. Chakraborty, G. L. Ciampaglia, D. Corney, R. DiResta, E. Ferrara, S. Hale, A. Halevy, E. Hovy, H. Ji, F. Menczer, R. Miguez, P. Nakov, D. Scheufele, S. Sharma, G. Zagni, Factuality challenges in the era of large language models and opportunities for fact-checking, Nature Machine Intelligence 6 (2024) 852–863. doi:10.1038/s42256- 024- 00881- z. [19] K. M. Caramancion, News Verifiers Showdown: A Comparative Performance Evaluation of ChatGPT 3.5, ChatGPT 4.0, Bing AI, and Bard in News Fact-Checking, 2023. URL: http://arxiv.org/abs/ 2306.17176, arXiv:2306.17176. [20] E. Hoes, S. Altay, J. Bermeo, Leveraging ChatGPT for Eficient Fact-Checking, 2023. doi: 10.31234/ osf.io/qnjkf. [21] D. Quelle, A. Bovet, The perils and promises of fact-checking with large language models, Frontiers in Artificial Intelligence 7 (2024). doi: 10.3389/frai.2024.1341697. [22] Z. Ni, Y. Qian, S. Chen, M.-C. Jaulent, C. Bousquet, Scientific evidence and specific context: leveraging large language models for health fact-checking, Online Information Review ahead-ofprint (2024). doi:10.1108/OIR- 02- 2024- 0111. [23] F. Motoki, V. Pinho Neto, V. Rodrigues, More human than human: measuring ChatGPT political bias, Public Choice (2023). doi:10.1007/s11127- 023- 01097- 2. [24] M. Makhortykh, M. Sydorova, A. Baghumyan, V. Vziatysheva, E. Kuznetsova, Stochastic lies: How LLM-powered chatbots deal with Russian disinformation about the war in Ukraine, Harvard Kennedy School Misinformation Review (2024). doi:10.37016/mr- 2020- 154. [25] J. Rutinowski, S. Franke, J. Endendyk, I. Dormuth, M. Roidl, M. Pauly, The Self-Perception and Political Biases of ChatGPT, Human Behavior and Emerging Technologies 2024 (2024) e7115633. doi:10.1155/2024/7115633. [26] A. Afonso, Whose Interests Do Radical Right Parties Really Represent? The Migration Policy Agenda of the Swiss People’s Party between Nativism and Neoliberalism, in: U. Korkut, G. BuckenKnapp, A. McGarry, J. Hinnfors, H. Drake (Eds.), The Discourses and Politics of Migration in Europe, Palgrave Macmillan US, New York, 2013, pp. 17–35. doi:10.1057/9781137310903_2. [27] Staatssekretariat für Wirtschaft SECO, Arbeitslosigkeit – einige Kennzahlen, Technical Report, 2024. URL: https://www.admin.ch/gov/de/start/dokumentation/medienmitteilungen.msg-id-99617. html#:~:text=Die%20saisonkorrigierte%20Arbeitslosenquote%20erh%C3%B6hte%20sich,tiefsten% 20Wert%20seit%202001%20entspricht. [28] Staatssekretariat für Wirtschaft SECO, Registrierte Arbeitslose nach Nationalitätengruppen und
Herkunftsländern, Technical Report, 2023. [29] Staatssekretariat für Wirtschaft SECO, Arbeitslosigkeit in der Schweiz 2020, Technical Report,
Neuchâtel, 2021. [30] Staatssekretariat für Wirtschaft SECO, Die Lage auf dem Arbeitsmarkt: Dezember 2023, Technical
Report, 2023. [31] Bundesamt für Statistik, Erwerbslose gemäss ILO nach Geschlecht, Nationalität und Altersgruppen, brutto- und saisonbereinigte Werte. Durchschnittliche Monats-, Quartalsund Jahreswerte, 2024. URL: https://www.bfs.admin.ch/bfs/de/home/statistiken/arbeit-erwerb/ erwerbslosigkeit-unterbeschaeftigung/erwerbslose-ilo.assetdetail.32586227.html. [32] Nachhaltigkeits-Initiative, n.d. URL: https://nachhaltigkeitsinitiative.ch/. [33] Staatssekretariat für Wirtschaft SECO, Arbeitslosigkeit in der Schweiz 2019, Technical Report,
Neuchâtel, 2020. [34] A. Afonso, When the Export of Social Problems Is No Longer Possible: Immigration Policies and Unemployment in Switzerland, Social Policy & Administration 39 (2005) 653–668. doi:10.1111/j. 1467- 9515.2005.00462.x, publisher: John Wiley & Sons, Ltd.