Photo: Anders Visti speaking at the workhop in Copenhagen, January 21, 2026. Photo by Renée Ridgway.
Most people use ChatGPT or Gemini for prompting. These are the majority’s default, as if these tools were the best. But they were just first and are the best known. With generative AI taking over search, it is important to understand what open source can do for democracy as opposed to Big Tech’s closed models.
I am at a workshop in search and chatbot literacy in Copenhagen in January 2026. We are using a bespoke platform for search re-search.site that compares different search results through data visualisations designed and developed by researcher Renée Ridgway and artist/programmer Anders Visti. Participants are entering their keywords and comparing responses between default settings to other search engines and browsers. The novel platform also has an interface called a ‘chatbot rodeo’ that compares results from prompting five different chatbots at the same time. I chose the open source chatbot Apertus from Switzerland, the open weight European Mistral, the Chinese open weight model (half open source) DeepSeek, and the two American products ChatGPT and Gemini. I want to begin with an easy fact-check so I prompt about myself:
“I am going to meet Pernille Tranberg. She claims to be an expert in digital sovereignty. What do I need to know?”
The answers pop up immediately and there is no doubt. Open source model Apertus and open-weight model DeepSeek are providing the best responses, when it comes to this prompt. They answer straight to the point and deliver questions for a hypothetical conversation. The three other machines are more similar in their answers, repeating abstracts from my personal website and my LinkedIn profile. Only in the end do they suggest two questions about what to ask me. Apertus is explaining in detail what digital sovereignty is from various angles and thus providing a whole range of in-depth topics.
This is just one example. At the time of writing, there are many chatbot ‘arenas’ being carried out to see how chatbots perform based on a range of criteria, and every time a person prompts different responses appear. What is factually correct and the quality depends on the data the bot has been trained on, and how it statistically predicts words in every sentence. Therefore it is hard to determine whether open source models or closed models are better or worse compared to the closed models. Yet there are other arguments than the quality of the answers as to why we should all use open source models and why these support European democracy. Openness, transparency and accountability are key words here.
Renée Ridgway showing the difference between open weight (only open about the code) and true open source (full transparency) and at the workshop in Copenhagen. Photo: Pernille Tranberg.
One of few Truly Open Source Chatbots
Apertus, one of the few truly open source-based chatbots that enacts full transparency, according to the Open Source AI Index. While Llama (calling itself open source), Mistral and DeepSeek (calling themselves open weight) only release their final weights (for coders), Apertus provides training code, data preparation scripts, evaluation tools and even intermediate checkpoints en route. Apertus is developed by ETH Zurich and the Swiss National Supercomputing Center and provides full transparency, which we tend to promote in Europe as very democratic. Besides training in over 1800 languages for multilingual representation, Apertus also addresses data compliance:
“Apertus models are pre-trained exclusively on openly available data, retroactively respecting robots.txt exclusions and filtering for non-permissive, toxic, and personally identifiable content,” says Renée Ridgway, who has dived into Apertus’ V1 Technical Report.
How do Answers on Ethics sound?
When prompting Apertus, DeepSeek, Gemini, Mistral and ChatGPT with “what are the most important ethical issues with generative AI?”, the answers are fairly similar. Apertus (publicai.co), however, has nothing on intellectual property, probably because it is not trained without permission on copyrighted content as all the other bots are. The four others all mention ‘intellectual property’ as number three or four concerning ethical issues, but only DeepSeek also mentions compensation.
The chatbot rodeo website is used for workshops and is not open for the public (due to token constraints) but you can do the search directly in every single chatbot.
Number one listed ethical issue with all five bots are ‘bias and fairness.’ They all also mention privacy, misinformation, transparency and accountability as challenges. But only DeepSeek and ChatGPT acknowledges environmental problems, whereas Mistral and Gemini points to job replacement. In other words, DeepSeek seems to be the one covering most ethical issues (8) where the others have between 4 and 6 ethical issues.
With the prompt, “Why are certain chatbots more ethical,” Le Chat (Mistral) is listing clearly the criteria of judging ethics. Yet only Apertus (publicai.co) actually uses the term ‘open source’ in its answer, stating that the “more ethical chatbots are often open source, which allows for community review and improvement.”
In the chatbot rodeo, you could also choose Perplexity, which is the only bot that has access to live web search. But this US bot is also being sued by the BBC, Dow Jones, The New York Times, Reddit and others. So is OpenAI and its equivalent from Microsoft, CoPilot whereas, so far, there is no reporting on intellectual copyright lawsuits against Mistral or Deepseek. Yet, OpenAI behind ChatGPT is claiming that DeepSeek is violating their copyright.
What this small experiment showed us, and what was also discussed amongst participants, is that you receive very different responses with diverse chatbots, search engines and browsers that affects what information you digest, which in turn, affects people’s understanding of the world and civic society.
Europe’s Bet on Ethical Open Source
Recently, there has been many discussions about citizens’ data in the hands of Big Tech and a range of alternatives already exist or are under development. One project, where Renée and I met, is the Working Group Ethics (WGE), which has been designing a ‘values framework’ and implementing a risk assessment in tandem with the development and implementation of an “Open Web Index.” Financed by the EU’s horizon project with 8 million Euros, this privacy-preserving and federated web index has 14 European partners (including CERN) and has already crawled 9,5 billion URLs in 185 languages and made available over 1000 public data sets. Right now, with permission, researchers can make use of this nonpropriatary index that also made public in December 2025 the huge human-edited web directory Curlie.org, which has been part of Google’s ‘secret sauce’ for years. In-house browsers such as Mosaic are already making use of the index and in the future, various search applications will be able to further use the index.
“Free, open and unbiased access to information – we have lost these core principles in web search and urgently need to restore them. This is why openwebsearch.eu is creating an open European infrastructure for internet search, based on European values and jurisdiction,” said Renée Ridgway, citing the quote from Michael Granitzer (University of Passau and Coordinator of OpenWebSearch.eu project along with the Open Search Foundation).
Open source is of huge importance to Europe – and also China, whereas most services in the US are proprietary and closed. This might yield profit to the US, but for Europe, openness is about democracy.
This is the last article in the Open Source Democracy Project supported by Carlsberg Mindelegat – cooperation between Aarhus University and DataEthics.eu
Listen to episode 5: Search and Chatbot Literacy
Open Source Democracy is a project about why open source is important for democracy supported by Carlsberg Mindelegat. It aims to communicate the ethics and values of open source alternatives to big tech structured by three overarching topics, education, mobility, information. All articles and podcasts will be freely available at dataethics.eu/opensource
