Introduction
Using generative AI (GenAI) in the public sector comes with specific concerns related to accuracy, effectiveness, ethics and trust. In an attempt to address these and other challenges, the Dutch Government published a Government-wide vision on generative AI (GenAINe 2024). In this vision, they encourage “responsible experimentation and a learning approach” in order to “use generative AI innovatively […] while exploring its potential”. In an article for Tiro, Gijs Freriks gave examples of uses of GenAI in parliamentary reporting (Freriks, 2023). Which factors should we consider during “responsible experimentation” with GenAI in a parliamentary context?
Misinformation
A parliament should be very careful not to become a super-spreader of misinformation. One of the main challenges of building a use case for GenAI in parliamentary settings is the high standard of accuracy required. Parliamentary discussions are often nuanced. Factoring out the human element in judging these nuances comes with risks. Even minor inaccuracies in AI-generated summaries can misrepresent the key points of a parliamentary debate, leading to misinformation and potentially resulting in confusion or, worse, erosion of trust in institutions.
Black box
GenAI can process large amounts of text quickly, but its outputs can be inconsistent and can vary based on factors such as the phrasing of the prompt or the data it has been trained on. This is complicated further by the lack of transparency surrounding the algorithmic systems that make GenAI possible, commonly referred to as the “black box”: information goes in and decisions come out, but we have little idea how they arrived at those decisions (Jiménez, 2024). This variability can lead to unpredictable results, making GenAI risky for tasks that demand high reliability. In parliamentary contexts, stakes are high and public accountability is key. These risks make it necessary to double-check AI-generated content.
Context
GenAI’s struggle with context is another major obstacle to its effective use in parliamentary settings. Parliamentary language is often nuanced, and speakers may use rhetoric or implied meanings that require contextual knowledge. For example, parliamentary debates frequently involve layered discussions on policy in which MPs refer to previous laws, policies or political events. Without full comprehension of this context, GenAI might generate responses that miss these connections, thus failing to capture the debate’s true essence. This lack of contextual understanding is a major limitation, especially when nuance and accuracy are crucial.
Effective prompts
The performance and effectiveness of GenAI is heavily dependent on the clarity and specificity of the prompts given to it (STS 2024). In other words, the AI’s output quality is closely linked to the human operator’s skill in creating precise and effective prompts. Crafting a well-structured prompt involves understanding the AI’s strengths and weaknesses and the standards of parliamentary discourse. Inconsistent prompt quality can lead to variability in GenAI’s output, which limits the reliability of the AI as a tool for parliamentary use. If different users achieve varying results, GenAI’s usefulness is compromised, as parliamentary work demands consistency, clarity and precision.
Ethical concerns
Introducing GenAI in parliamentary work also brings ethical concerns, particularly around transparency, accountability and public trust. As a public institution, a parliament must uphold the highest standards of integrity. The use of AI-generated content may lead the public to question the objectivity and accuracy of a parliament’s communications.
Biases
To preserve trust, any AI-assisted work would need to be transparent about the AI’s role and ensure that generated content remains free from misrepresentation or bias. This is complicated due to the “black box” nature of AI mentioned earlier. There is a risk that AI could unintentionally inject subtle biases into its outputs, since the models are trained on vast datasets that include historical and cultural biases (Jnini, 2024). If not carefully monitored, such biases could shape the tone or interpretation of parliamentary debates in ways that misrepresent political positions or policy issues. Public trust in parliamentary information could be undermined if the public perceives AI-generated content as unreliable.
Closing remarks
In conclusion, what does “responsible experimentation” with generative AI for parliamentary reporting imply? Experimentation is key for understanding GenAI’s potential. It can be a surprising and sometimes even useful tool. It needs to be welcomed that Governments create room for experiments but, in order for them to be successful, the specific shortfalls and risks of GenAI need to be acknowledged, described and effectively addressed.
Deru Schelhaas is a parliamentary reporter in the Dutch House of Representatives. He is a member of a working group that investigates and tests relevant technical developments for parliamentary reporting and broadcasting.
References
Freriks, G. (2023) New uses for AI in parliamentary reporting. Tiro 2/2023. Available at https://tiro.intersteno.org/2023/12/new-uses-for-ai-in-parliamentary-reporting/
GenAINe 2024 = A Government-wide vision on generative AI of the Netherlands. Government of the Netherlands. Available at https://www.government.nl/documents/parliamentary-documents/2024/01/17/government-wide-vision-on-generative-ai-of-the-netherlands
Jiménez, M. (2024). Unboxing the Black Box of AI. Utrecht University. Available at https://www.uu.nl/en/organisation/in-depth/unboxing-the-black-box-of-ai
Jnini, S. (2024). Is GenAI biased? How to stop algorithms from reproducing human errors and prejudices. Atos. Available at https://atos.net/en/blog/is-genai-biased
STS 2024 = Effective Prompts for AI: The Essentials. MIT Management: STS Teaching & Learning Technologies. Sloan Technology Services. Available at https://mitsloanedtech.mit.edu/ai/basics/effective-prompts/
