Large Language Models for media and democracy: wrecking or saving society?

With this workshop we intend to map the salient technical and societal issues that emerged around foundational models and to discuss recent developments to address them.

Over the past years, foundational models, including large-language models (LLMs) and multi-modal systems, have significantly advanced the possibilities regarding the understanding, analysis, and generation of human language [1]. These models, based on artificial neural network computations trained on a large scale of documents, open the possibility of performing advanced tasks related to language, audio, video, and image processing.

However, these models are relatively novel and there are several crucial open issues that need to be addressed. These issues are related both to the development and the impact of these models on society, especially on media and democracy.

Reliability and transparency

Technical issues regard the reliability and the transparency of LLMs. While LLMs are very good at providing believable and convincing responses to users, assurances on the quality of the information provided should be given. For example, LLMs often suffer from the so-called hallucination issue, that is, they generate believable but false statements. Also, they are often trained on large but biased datasets, and this might exacerbate the presence of stereotypes and undesired associations in the responses given. While preliminary solutions to these problems have been proposed (e.g. [4,5]), mapping these open issues and discussing potential solutions is important to enhance a conscious use of these models and an understanding of the pitfalls and limitations of the solutions proposed so far to known issues. For example, debiasing may result in bleached and uninteresting responses.

Use and misuse

Societal issues concern the possible use and misuse of LLMs. Media, as communication tools for storing and delivering content, can find in LLMs a powerful tool to automate and accelerate the information generation process. However, such automation should be performed by understanding the limits, risks, and capabilities of LLMs first, in order to guarantee the quality of the output generated. For example, while it could be hazardous to create content exclusively based on LLMs, other tasks like content curation or engagement enhancement could significantly benefit from LLMs.


Besides their effect through the media, democracy can be affected directly by LLMs, because of the influence that these models can have on the generation and propagation of disinformation [2], e.g. by deliberately prompting to generate false or unethical information. However, LLMs can also be part of the strategy to contrast misinformation spreading, by allowing a deeper and more advanced assessment of the information spread online.

In light of these considerations, the legal domain should provide tools to govern the adoption of these models and their effect on society. The AI ACT represents a beginning in this direction, although further developments are needed [3].

In this workshop, we intend to showcase, discuss, and advance recent developments regarding LLMs, both from the technical and the societal point of view.

[1] Manning, Christopher D. (2022). "Human Language Understanding & Reasoning". Daedalus. 151 (2): 127–138. doi:10.1162/daed_a_01905. S2CID 248377870.

[2] European Digital Media Observatory (2023) on/

[3] Helberger, N. & Diakopoulos, N. (2023). ChatGPT and the AI Act. Internet Policy Review,


[4] Y Liu, Y Yao, JF Ton, X Zhang, RGH Cheng, Y Klochkov, MF Taufiq, H Li. Trustworthy LLMs: a Survey and Guideline for Evaluating Large Language Models' Alignment. arXiv preprint arXiv:2308.05374, 2023.

[5] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Proceedings of the 34th International Conference on Neural Information Processing Systems (NIPS'20). Curran Associates Inc., Red Hook, NY, USA, Article 793, 9459–9474.

Large Language Models

Practical information

Save the date
The workshop will be held on 23 and 24 April 2024 at CWI. Go to the programme.

Registration is closed

See here the Call for Posters

This workshop is organized with the Hybrid Intelligence Center (VU Amsterdam and TU Delft) and Hands4Grants.


A broad range of topics will be discussed:

  • Technological issues of LLMs: addressing hallucination, bias, creating transparency, respecting IPR, authority & quality, augmenting models with value-awareness & ethics, specific (Dutch) contexts and critical decisions within long-tail distributions.
  • Solutions to issues, among others Retrieval Augmented Generative AI (RAG), debiasing, augmentation, filtering (RLHF), emerging properties.
  • Users of AI in media and democracy: election tools, cyber crime, startup companies.
  • Societal issues of LLMs: risks of using AI solutions in media and democracy, authentication, plagiarism, diversity, adaptability, privacy and leakage, misinformation spreading.
  • Solutions to risks: prescription, regulation, liability, AI-literacy
  • Governance