The Brief: We often hear that AI should be aligned with human values. But whose values, defined by whom, and how would we know? Three researchers, working on public participation, journalism and model interpretability, examined alignment as a continuous, contested process rather than a box ticked by developers.

Moderator: Mira El Kamali

Asking the Public

Anna Sotnikova of Public AI presented the National AI Dialogue, which asks Swiss residents what they expect from AI as Apertus, the Swiss open model, is still being shaped. Participants can upvote questions, propose their own and comment on others' answers. Some 1,600 people have taken part so far, with retention above 60%, far beyond the 10% partners had expected.

Early results point to strong support for Swiss AI development, for data stored in Switzerland, and for governance shared between institutions, government and the public. Asked whether they want a tool or a companion, a large majority chose a tool. Two findings surprised the team. Bias was mentioned in only about 17% of concerns, and around 40% of respondents saw harmful outputs as the user's responsibility. "That's a signal for us that there should be AI literacy."Results will feed the Apertus constitution and a report for the Geneva AI Summit.

Journalists as Artisans of Words

Matilde Barbini of EPFL studies alignment in newsrooms. Journalists, she found, value attribution and verification and see themselves as "artisans of words." One local reporter described how AI search overviews summarise his work while readers no longer visit the site. "My words are being used to inform people, but I'm not making any profit of it."

Her central concept is editorial authority. When AI enters a newsroom without a policy, decision power migrates towards the model and the platforms. The more a model sounds like the newsroom's own voice, the more journalists defer to it. "It's a kind of feedback loop." Her conclusion: those affected should help design the guidelines.

When Models Talk to Models

Angelina Parfenova works on reasoning and interpretability. She has found that a model's output does not always match its own internal reasoning. On evaluation, she treats human disagreement as a ceiling. "If humans themselves disagree, then let models also disagree with them."

Her multi-agent experiments raised a warning. After about five rounds of discussion, three or more models converge on one opinion, even when instructed to hold opposing views. A moderator agent did not help, since it was trained on the same data. "If there is some bias and some models are sharing it, I'm afraid it's going to amplify."

Debate in the Room

Audience members pushed back on several points. One noted that culture and values are not the same, and that values are individual rather than national. Another questioned whether one constitution matters when thousands of models are released. Anna answered that the aim is a constitution for Apertus specifically. A lively exchange on whether to collect gender data showed that alignment questions start with survey design itself.

On Apertus's future, panelists noted that training only on properly licensed data reduces raw performance. They suggested that the model's value may lie in specific Swiss use cases, local data and explainability. Matilde observed that academia is turning to open models precisely because they allow research into how models work.

Strategic Implications

Make alignment participatory. Values become concrete only when the people affected define them.

Watch where authority flows. Introducing AI without a policy silently moves decision power.

Accept disagreement as data. Evaluate models against the distribution of human views, not a single answer.

Beware of consensus between agents. Multi-agent systems can amplify, not cancel, shared biases.

More on the panelists

Anna Sotnikova, Public AI
Matilde Barbini, EPFL
Angelina Parfenova
Mira El Kamali, moderator