Master’s Thesis: Towards Reliable Automatic Fact-Checking of Swedish Political Debates

Background

The non-profit automatic fact-checking application Sanningskollen was recently released with the aim to fact-check political debates and interviews ahead of the Swedish election in 2026. It automatically transcribes political discussions and verifies each speaker’s claims using retrieval-augmented techniques. Sanningskollen thus represents a significant first step toward lowering verification thresholds and making Swedish political discussions more fact-aligned.

With Sanningskollen as a starting point, this project will look at further possible improvements to automated fact-checking for Swedish political discourse. The guiding questions are:

• Assuming we have a fact-checking system, what are the metrics by which we assess its reliability?
• How do we test these?
• What does it mean to be reliable in this context?

The project

The project consists of the following steps:

• Perform a literature survey to identify established automated fact-checking systems, methods for benchmarking automatic fact-checkers and methods for measuring the severity of different fact-check errors, tied to their potential consequences.
• Analyse the performance of Sanningskollen for fact-checking Swedish political discourse, grounded in the findings from the literature survey. This could also involve a deeper analysis of potential bias in the system, tied to questions such as “What statements are judged to be ‘false’ by the system and why?” and “Is the system’s justification really a good, impartial justification?”.
• Compare fact-checking Swedish politics with, for example, US politics – what are the commonalities and significant differences? What are their implications for transferring automatic fact-checkers between the political discourse of different countries?
• Based on the analysis above, investigate and implement suitable improvements to Sanningskollen or similar automatic fact-checkers.

It is important that the project aligns with the interests of the thesis worker(s). Therefore, we are willing to adapt the project according to your preferences.

Who should apply?

This project is suitable for students who want to learn more about reliable and applied LLM systems. You should ideally have experience with Python and NLP, and familiarity with LLM APIs, prompting and evaluation, or information retrieval/RAG. Experience with experimental design, annotation or statistics is a plus, as is an interest in politics, journalism or media. Speech/audio processing experience is a bonus for work on speaker identification.

Students interested in publishing papers on their work are also encouraged to apply.

The project can be carried out by one or two students, but we encourage students to work in pairs. Since the work involves the Swedish language, we recommend that at least one of the applicants speaks Swedish.

Why apply?

As a thesis worker, you will get the opportunity to:

• Learn to scope, discuss and present research ideas
• Work hands-on with a live fact-checking system used by the public during the 2026 Swedish election
• Learn more about applied NLP, LLMs and retrieval-augmented verification
• Work with the full chain from speech to verdict: Swedish speech recognition, speaker identification, claim extraction and evidence retrieval
• Build and use a benchmark of real Swedish political debates with human-reviewed fact checks
• Learn about tools for efficient ML development
• Directly contribute to important NLP research for social science
• Be part of a non-profit effort to increase truthfulness in Swedish political debates

Who will you be working with?

You will mainly be working with:

• Lovisa Hagström, industrial supervisor at eghed
• Bastiaan Bruinsma, academic supervisor at DSAI, Chalmers University of Technology
• Moa Johansson, examiner at DSAI, Chalmers University of Technology

You will also be offered the opportunity to work from eghed’s office in Gothenburg and be part of the eghed community.

About eghed

eghed is a specialist consultancy within Data, Machine Learning and AI. We work with technically demanding problems where advanced analytics, mathematical modelling and software engineering meet real-world applications.

As a thesis student, you will work closely with experienced consultants and get the opportunity to turn theoretical methods into solutions that operate on real data.

How to apply

Send your application to thesis@eghed.se.

Please include:

• Your CV
• A personal letter introducing yourself (or yourselves), your relevant background and why you would like to work on the project
• Your Ladok transcripts

Applications are reviewed continuously and candidates may be chosen before the application deadline – so if you are interested in this project, do not hesitate to apply!

For questions about the project, contact lovisa.hagstrom@eghed.se.

References

• Guo, Zhijiang, Michael Schlichtkrull, and Andreas Vlachos. “A survey on automated fact-checking.” Transactions of the Association for Computational Linguistics 10 (2022): 178–206.
• Quelle, Dorian, and Alexandre Bovet. “The perils and promises of fact-checking with large language models.” Frontiers in Artificial Intelligence 7 (2024): 1341697.
• Wang, Yuxia, et al. “Factcheck-Bench: Fine-grained evaluation benchmark for automatic fact-checkers.” Findings of the Association for Computational Linguistics: EMNLP 2024 (2024).
• Si, Chenglei, et al. “Large language models help humans verify truthfulness – except when they are convincingly wrong.” Proceedings of NAACL-HLT 2024 (Volume 1: Long Papers) (2024).
• Warren, Greta, Irina Shklovski, and Isabelle Augenstein. “Show me the work: Fact-checkers’ requirements for explainable automated fact-checking.” Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (2025).
• Wintersieck, Amanda L. “Debating the truth: The impact of fact-checking during electoral debates.” American Politics Research 45.2 (2017): 304–331.
• PolitiFact
• Google Fact Check Tools API