Electorally: How artificial intelligence should help with voting decisions

Ask questions about the election program using the AI tool Wahlweise. The model lets you explore and clarify your specific points of interest in the election.

Save to Pocket listen Print view
A ballot paper next to the ballot envelopes

For the state elections in Saxony and Thuringia in September, there is an AI from Thuringia to help with an overview of the election programs.

(Image: Jan von nebenan/Shutterstock.com)

5 min. read

The Wahl-o-mat is designed to give voters a better overview of the parties' positions. The Wahlweise service aims to go one step further: With the help of artificial intelligence (AI), the service aims not only to answer specific questions with suitable passages from the election manifestos, but also to compare voters' own viewpoints with the programs of electable parties.

"We come from Thuringia ourselves and thought we would use artificial intelligence to get an overview," says Martin Schiele, one of the initiators and CEO of the Thuringian AI company AIUI. In contrast to the Wahl-o-mat from the Federal Agency for Civic Education, the answer comes directly from the parties' election manifestos. "We initially used the Wahl-o-mat questions as a guide, but didn't always find a concrete answer in the manifestos," explains Schiele. The answers from the Wahl-o-mat come directly from the parties, but are not necessarily reflected in the election manifesto. Instead, the teams that develop the questions are based thematically on the election manifestos.

Alternatively, there is a chat for specific questions about the election programs or a quiz to compare your own views with those of the parties.

(Image: Screenshot: wahlweise.info)

The model is based on Llama 3 and works with Retrieval Augmented Generation (RAG): The AI searches the database of election program excerpts for those that match the question asked as closely as possible. The team did not need any training data for this. The advantage of dividing the election programs into thematic excerpts is that the person asking the question receives answers that match the content, but at the same time the excerpts are short enough that the AI does not forget anything when summarizing - which can happen with longer texts. "We carried out a lot of tests and improved the prompt with sixty to seventy iterations so that both the answers matched and the AI only did what it was supposed to do: Output information from the election programs," emphasizes Schiele. "The AI also provides a specific source where you can read the statements."

The AI reformulates the excerpts from the party manifestos, says Schiele. The aim is to formulate statements as neutrally as possible. The AI itself should also not be biased. This was important to the team, emphasizes Schiele. Studies have shown that artificial intelligence can develop its own political bias through the training material. "To remove this bias, we anonymized the programs." This means that the AI can no longer assign which statement belongs to which party, but instead assigns it to an alias. In the quiz, even the person in front of the screen does not know which statement belongs to which party. "We really wanted to make it possible to assign content," says Schiele. Without bias on either side.

Schiele and his team financed Wahlweise from their own funds - as did the partner companies that provided the hardware or the tests, for example. "Of course this is also marketing for us, but first and foremost we wanted to make this tool available to others."

Using Wahlweise-KI to get to grips with the political agenda of the parties requires a lot more reading concentration than the Wahl-o-mat, for example. The text passages are longer and no direct comparison can be made - unless users ask the same question several times if they have selected different parties. Not all parties are represented either: From a three-page flyer, for example, not enough information could be found in response to all questions. Nevertheless, the team would like to use the tool to motivate young people to engage with the political views of the parties. "We're trying to get younger people primarily by asking them to elicit nonsense from our AI," says Schiele. Creating an absolutely secure system that does not produce errors is not feasible.

Schiele is certain that AI assistance will play an increasingly important role in political education. "We will no longer do many things like research: In three to five years, an AI assistant will do that for us," he assumes. "The question then is: who will be the gatekeepers offering the systems?"

(are)

Don't miss any news – follow us on Facebook, LinkedIn or Mastodon.

This article was originally published in German. It was translated with technical assistance and editorially reviewed before publication.