In preparing the spring 2025 Mental resources National Dialogue summary, we experimented with using artificial intelligence. What added value does it bring, and what should be considered when using it? In this blog, Elina Henttonen from Valtaamo Oy shares the lessons and insights gained from this trial.
By Elina Henttonen
Artificial intelligence can strengthen democracy in many ways—if it is applied responsibly and transparently. Already today, AI is used to structure material from deliberative citizen discussions, process municipal feedback, and analyse public consultation responses in legislative drafting.
Its power lies in handling, classifying, and summarising large volumes of citizen perspectives and weaving them into comprehensive overviews. At best, AI can improve the knowledge base of decision-making and, in doing so, foster citizen participation. At the same time, however, we must remain mindful of issues such as privacy, the functioning of algorithms, and the responsibility of interpreting results.
The goal of National Dialogues is to strengthen trust, participation, and democracy. A single round of National Dialogues can involve more than a hundred separate dialogues (and we hope many more in the future). Each dialogue generates written records, which can range from a few pages to dozens. Unsurprisingly, such vast datasets are demanding and time-consuming to process.
This is why we wanted to explore how AI could support the analysis and reporting of National Dialogue material, and more broadly, how it could help structure experiential knowledge. Our aim was also practical: to make producing summaries faster and more cost-efficient.
For the spring 2025 Mental resources dialogues, I tested AI as a support tool for data management and analysis. Having prepared several similar dialogue summaries before, I first created my own “manual” thematization of the 117-dialogue dataset. This gave a solid baseline to compare the AI’s results with my own sense of the material—and to judge how accurately AI captured the essence of it.
I then used two different tools: Notebook LM (based on Google’s Gemini language model) and Skimle, a Finnish tool still in testing. I was curious about how each tool summarized, categorized, and themed the data, and what kinds of questions and perspectives they raised. I also wanted to compare their outputs with each other and with my own analysis. Could AI surface something I had missed? Would it enrich the analysis with new angles—or would it flatten out diverse voices?
What I found was encouraging, though not surprising: AI is effective at structuring large datasets. Its categories differed somewhat from mine, but the substance was largely the same. This confirmed that my own analysis had captured the key themes comprehensively. Yet I had secretly hoped AI would challenge my framing and reveal something entirely new. That did not happen.
Despite its usefulness, AI cannot replace thorough human work. A researcher needs to be familiar with the material to evaluate the AI’s categorisations and interpretations. In my experiment, AI did not hallucinate or produce false content, but without a deep understanding of the dataset, it would have been difficult to judge the significance of its suggestions.
Another feature of AI-generated content is its initial polish: at first glance, it can look impressively smooth—even brilliant. But closer examination often reveals flatness, repetition, overlaps, empty phrases, and a weakening sense of narrative.
It is also the researcher’s responsibility to situate observations in context. In my trial, I had fully anonymised the dataset, so AI had no metadata such as organiser or participant details. While it still produced meaningful classifications, linking insights back to context required a clear system on my part.