While many artificial intelligence chatbots created by Silicon Valley tech companies, trained on American content, European tech companies are developing their own models, using the continent's culture and languages.
A woman’s voice emanates from Michel-Marie Maudet’s laptop, sitting on a desk at the headquarters of his software development company Linagora, in Issy-Les-Moulineaux, south of Paris.
"Hello, I am Lucie, a large language model trained on a massive data set of text and code in French and other European languages."
Speaking English with a French accent, she continues: "I am able to understand and respond to questions in a way that is sensitive to the nuances of European culture and language."
This chatbot, which can communicate in French and several other European languages, uses the word "nuance" frequently when describing itself – which Maudet echoes.
"It is a question of nuances. These large language models are statistics, and if the models are trained mainly on US content, you are more likely to get answers influenced by US culture."
Listen to an interview with Michel-Marie Maudet (and "Lucie") in the Spotlight on France podcast, episode 124:
The idea is that different content will be generated on a tool that has been trained on different languages.
"Languages are our culture, our civilisation, our values, and we developed Lucie, our large language model, to fix this under-representation of our culture," said Maudet.
Lucie was released to the public in January with little testing beforehand, and ran into problems as users found it was generating nonsense – and worse. It was taken offline three days later.
Maudet said that while the release was premature, it generated interest – notably about the training data, which was made public at the same time as the chatbot, because Linagora is committed to developing open-source tools.
Behind the curtain of AI
"It's a completely open-source model," he said. "If you want to build transparency and trust in an AI system, you have to know where and how these models are built."
The training data set was downloaded 10 times more than the actual model, revealing the level of interest in how these tools work.
Paris hosts AI summit, with spotlight on innovation, regulation, creativity
And while Lucie's release was something of a public relations disaster, Maudet says it also demonstrated an interest in alternatives to tools developed by US tech giants.
This was borne out at the AI Action Summit held in Paris earlier this month, at which France and other European countries sought to stake their claim in terms of innovation and governance.
European identity
"People are asking for this kind of technology, as an alternative to Chinese or US companies," Maudet added. "I think the debates around Lucie were very interesting, because they raised an expectation that we want to have our own technology, our own strategy, our own mastery of our digital future."
Linagora is not the only company developing these alternatives, and far from the most powerful. But the company is dedicated to transparency and open sourcing, in its aim to create a tool that can generate text not derived from American content.
France's Mistral AI teams up with UAE-backed developers as Le Chat app launches
"We want to incorporate these systems into our daily life, and I am not sure we have the same approach in the US as our social system here in France or Europe," Maudet explained.
However, the company's mission presupposes a European identity that is not always clear, or unanimous.
"A big challenge for Europe is to act as one continent," said Maudet. "AI models could ease a common vision of what we call Europe. We will be stronger and better if we play collectively and act as a single continent and one entity."
Listen to an interview with Michel Maudet in the Spotlight on France podcast, episode 124, listen here.
