"Ethics cannot be an afterthought": a conversation with AI ethicist Ning Wang

Apertus was developed under the Swiss AI Initiative by ETH Zurich and EPFL, and is presented as an example of compliant, transparent AI. As an ethicist, what is your first reaction to a model that publishes its source code, training data and training methods in full?

Apertus takes a visionary step towards open and transparent LLMs. But it is only one first step, as AI Systems like Apertus do not simply reflect reality; they structure what counts as knowledge, whose voices are legible, and which perspectives scale. Instead of the conventional question “Is the model biased?”, we should start asking “Who has the authority to define, interpret, and contest what we see as the informational substrate of society?” In this context, neutrality is no longer a passive stance – it becomes an active design choice. Especially for Switzerland with a long tradition in defining itself as neutral, this is an important shift.

In the case of Apertus, publishing source code, training data, and methods signals a strong commitment of Switzerland to procedural openness. Ethically, this matters for two reasons:

• It lowers the barrier to scrutiny. Researchers, journalists, and civil society actors can examine not just outputs, but the underlying assumptions embedded in the system – data selection, filtering choices, optimisation goals. This is a critical precondition for accountability.
• It ensures a redistribution of access to knowledge. Most LLMs today are “black boxes” controlled by a few profit-oriented actors. Full transparency opens the possibility – at least in principle – for broader participation in auditing, improving, or contesting the model.

Looking through a more fundamental moral lens, however, further critical questions are being raised: does publishing training data and methods per se resolve questions about legitimacy? Who decides what data counts as representative? Which languages, dialects, or knowledge systems are prioritised or excluded? What normative assumptions are baked into filtering and alignment?

These are not just technical questions but political ones too. Transparency can indeed expose them but cannot settle them. In this sense, Apertus is less the endpoint of ethical AI, but more the beginning of a different kind of responsibility – once everything becomes visible, the challenge then boils down to building the institutions and practices that can make the greater visibility more meaningful.

You have argued that AI development teams tend to prioritise technical and economic goals, leaving ethical and societal concerns until later. What is the practical cost of this sequence?

In the case of Apertus, what we see is that stereotyping persists despite transparency. This reflects a broader sequencing problem in AI development: systems are optimised first for performance and efficiency, while questions of inclusion, representation, and societal impact are deferred, potentially externalising harms.

From my perspective, this ordering has a series of concrete and far-reaching consequences:

• First, it locks in biases at the structural level. When issues like gender inclusivity or representation are not addressed during data curation and model design, they become part of the system’s internal logic. Applying filters afterward, whether on training data or outputs, can mitigate symptoms but rarely remove the underlying patterns. In practice, this means stereotypes are not just occasional errors; they become statistically reinforced tendencies. There is plenty of research showing that this is the case, especially for underrepresented minorities.
• Second, it shifts the burden of correction onto users and the affected groups. If a system reproduces biased or exclusionary outputs, the cost is not borne by the developers but by those misrepresented. Consequently, they must report, contest, and navigate these harms. Over time, this creates a form of “participation fatigue” where marginalised groups are expected to continuously fix systems that were not meant for them, and that they did not (co)design.
• Third, it creates path dependency that is expensive to reverse. Once a model is trained, deployed, and integrated into workflows, correcting foundational issues becomes technically, economically, and more crucially societally costly. Retraining, re-curating datasets, or redesigning objectives is far more difficult than addressing these concerns upfront. The initial sequencing decision thus effectively constrains future ethical choices.
• Fourth, it normalises a reactive model of ethics. By treating ethical concerns as something to be “patched” later, organisations implicitly signal that societal impacts are secondary to technical milestones. This shapes institutional culture – ethics with this framing becomes a compliance obligation, rather than a design principle.
• Fifth, it affects trust in a subtle but durable way. Even with a transparent system like Apertus, users still notice when harms were foreseeable but not prioritised. This can erode trust not only in a specific model, but in the institutions behind it. The issue is not that imperfections exist, but that they reflect a hierarchy of priorities.

The real cost of this sequence is not biased outputs; it is the institutionalised delay in a system, where ethical considerations are always one step behind technological capability. A more sustainable approach would invert that sequence – not by slowing innovation, but by redefining what counts as “core functionality”. In that framing, inclusivity, representational balance, and societal impact are not add-ons; they are part of the system’s performance criteria from the start.

Unlike a human, who might make a few biased decisions a day, an algorithm can make thousands per second. Does this level of AI decision-making require a different ethical and legal response to what we currently have in Switzerland?

Yes – the scale doesn’t just amplify existing problems; it changes their nature. Treating AI as if it were simply a faster human decision-maker misses what is at stake.

With advanced AI systems, we are no longer dealing with isolated decisions, but with decision infrastructures. These systems do not just act at scale; they standardise judgments across contexts, often invisibly. When decisions scale, responsibility must scale accordingly. The ethical challenge of AI is not just that it can be biased; but that it can institutionalise biases at a speed and scope that outpaces our current ethical and legal imagination.

As Apertus shows, Switzerland has the institutional capacity to respond to this. Importantly, doing so will require moving from a case-by-case logic to a system-level governance mindset. The kind of response this point to, from my perspective, is not necessarily heavier regulation, but different mindset. For example:

• Effective mechanisms to identify and act upon systemic bias, not just individual complaints.
• Defined requirements for continuous auditing, not one-time certification.
• Clearer models of shared responsibility across developers, deployers, operators, and other related institutions.
• Enlarged institutional spaces where technical, legal, ethical, and societal perspectives can jointly evaluate systems.

"If we’re not even sitting at the table, how do we make our voices heard?" You've raised this from your own experience. What would it take for AI development teams to change that?

If people are invited to the table but cannot influence what is being served, inclusion has not taken place in a truthful sense. Now for the table to look meaningfully different, the issue is not simply who is invited, but how the table is structured – who sets the agenda, and whose knowledge counts. Without changing those conditions, diversity becomes symbolic rather than consequential. In the context of Apertus, I can see several structural shifts potentially taking place:

• First, from mere representation to genuine decision-making. It is not enough to increase the presence of underrepresented groups in teams, if key decisions about datasets, model objectives, or evaluation criteria remain concentrated elsewhere. Meaningful inclusion requires that diverse voices are present at the points where trade-offs are made – not just in advisory or downstream roles.
• Second, redefining what counts as expertise. AI development is still dominated by narrowly defined technical credentials. Yet many of the hardest questions, such as those about bias, language, culture, and societal impact, require interdisciplinary and experiential knowledge. Structural change would mean formally integrating ethicists, social scientists, and community-informed perspectives into core development processes – not as external reviewers but as co-designers.
• Third, changing incentive structures. Academic and technical environments reward speed, novelty, and performance. Work on inclusion, data curation, or bias mitigation is often undervalued because it is regarded as slower and less visible. If institutions are serious about change, they need to align incentives, such as funding, promotions, recognition, with the quality and inclusivity of systems – not just their technical benchmarks.
• Fourth, institutionalising contestability. What matters is whether there are mechanisms to challenge decisions without penalty. This could take the form of structured dissent processes, independent review boards, or participatory audits that include voices outside the immediate development team.

The underlying principle for initiating these structural changes should not be merely “adding voices”; but rather, “redistributing knowledge authority”. For Switzerland, on the one hand, I see this as an opportunity – its tradition of pluralism and negotiated governance could translate into AI development processes that are more participatory and reflexive than those driven purely by scale or market logic. On the other hand, this requires treating diversity not as a pipeline problem, but as a question of institutional design at its core.

Apertus is a foundation, not a finished product. Every adaptation is a potential point of ethical drift. Which governance mechanisms are essential to ensure that its values carry through to each new application?

Some ethical challenges may be intensified if we were treating LLMs as foundation models – because we decentralise responsibility. While the base model can embody certain commitments, like Apertus, every fine-tuning step is a potential point of drift. The ultimate question is therefore not how ethical the base model is; but how (and how far) its ethical properties travel.

This means that a foundation model is not just a technical base, but rather a normative starting point. If its ethical commitments are not actively carried through each layer of adaptation, they will not scale but will, instead, erode.  This is why we should establish a practice that

• treats LLMs as public infrastructure, not just commercial products;
• aligns development with democratic oversight traditions, as Switzerland’s comparative advantage;
• exports not just technology, but governance models.

Here, again, is where I see Switzerland as having a real opportunity – and making true contributions to the world – treating AI not just as modular technical products, but as an integrated governance ecosystem, where each extension of the model remains connected to a shared set of principles that are operationalisable, instead of being declared as “window dressing” or “ethical washing” slogans.

What this requires is a shift from “AI made in Switzerland” to “AI stewardedlike Switzerland”. This is because in the age of AI, neutrality is no longer about abstaining from power; but about how power is encoded. Switzerland has the unique opportunity to move from “neutrality” to “stewardship” –  designing AI systems that do not merely optimise performance, but safeguard the integrity of knowledge as a public good, not a private asset.