ORAL STATEMENT
Humanists International
Article 17 dialogue seminar at the European Commission
“The ethical and social impact of artificial intelligence”
Speaker: Dr. Dorothea Winter
Let me start with an image. Europe is like a mosaic, and when you walk up close, you see only stones, each with its own color, its own shape, and the differences are everything: countries, languages, cultures, religious and philosophical traditions. But when you step back, the stones compose a picture, and that picture is what the stones, for all their differences, share.
And that picture is humanist before it is anything else. The founding values of the Union, written into Article 2 of the Treaty on European Union, are human dignity, freedom, democracy, equality, the rule of law, and respect for human rights, and they are humanist because they hold the human being to be a free, reasoning subject. If we strip these commitments away, the mosaic loses its picture; the stones remain, but the meaning does not.
Into this picture, now, comes a new tool, and that tool is the hammer. We use a hammer to work on stones, to carve them, refine them, fit them into a pattern. And the hammer in this moment is Artificial Intelligence.
A hammer can break a stone or it can repair one, and AI has the same capacities. The question, therefore, is in whose hand it sits, with what training, and to what design. Every hammer carries the choices of its makers: what it amplifies, what it suppresses.
This is where the Digital Humanism Framework enters. Digital humanism is the philosophical humanist position that rests on three commitments.
The first commitment is that AI is a tool, and as a tool it is neither good nor bad nor neutral. To treat AI as neutral would be a category mistake, because neutrality implies truth, whereas AI systems produce only probability. An LLM such as ChatGPT predicts the next token from patterns in its training data, so that every output we receive has to be understood as a prediction.
The humanist response, therefore, is clear. We cannot set expectations of AI systems that they cannot fulfil. We have to talk about AI rationally, without falling into the dangerous narratives that surround it: animism, transhumanism, or, most recently, the claim by CEO of Big AI-Companies that AI is approaching “recursive self-improvement.” Such claims are philosophically, humanistically and scientifically problematic, and they distract from the pressing questions about specific systems, with specific designs, producing specific effects. Software is software. This, I should add, is also the architecture of the AI Act itself: a risk-based framework that classifies systems by their actual capacity to harm, not by the rhetorical category of “AI in general.”
The second is that AI can be called a System of Selection. AI Outputs are a selection of which sources to draw from, which patterns to amplify, which alternatives to suppress, which framings to inherit and which to discard. The selection is shaped by the data the system was trained on, by the engineers who designed it, and by the commercial incentives of the platform that hosts it. None of this is visible to the citizen, who uses the system.
This invisible selection is now operating at the heart of two domains where humanist concern is most pressing: public administration and education.
An example from public administration is the so called Toeslagenaffaire from the Netherlands — a tax-authority algorithm that treated dual nationality as a risk factor for benefit fraud — wrongly accused roughly twenty-six thousand families and forced the Rutte government to resign in January 2021. These are the structural consequence of delegating selection to opaque systems without democratic oversight.
In education, the question is even more consequential. Generative AI tools used by students pre-select what counts as a relevant answer, which sources are worth consulting, which voices are visible and which are not. Eurostat reported that more than 50 per cent of Europeans aged sixteen to twenty-four trust generative AI Systems to provide correct answers; 40 per cent rarely or never verify the outputs. A generation is learning what knowledge looks like through a selection mechanism that is opaque, designed outside European ideals, and driven by commercial profit. And the AI literacy that would let them recognise this is itself being weakened in European law: the AI Act made it an obligation under Article 4, and the Digital Omnibus would reduce it to a recommendation.
The third commitment concerns responsibility. AI systems are the most successful technology we have ever built for diffusing responsibility. When the output of an AI system is ethically problematic, when, for instance, someone is discriminated against, the responsibility diffuses. Designers point at engineers; engineers point at trainers; trainers point at the data.
In philosophy we call this the problem of many hands. In the complex and globally distributed processes through which AI systems are developed and deployed, it becomes difficult to identify who did what, and harder still to say who can be held accountable for any particular output.
It would be a poor philosophy talk that did not invoke Hannah Arendt at least once, so here we are. Hannah Arendt developed a concept she called the rule of Nobody. By this she meant the danger that arises when humans hide behind rules and procedures to avoid personal responsibility, so that no one can be argued with, no one can be held accountable, and yet someone is always being ruled. I see the same danger in the implementation of AI, where human responsibility is delegated to systems that are themselves incapable of bearing it. Decisions about people’s lives, their education, their communication, their access to medical advice, are made through AI systems, while every human in the chain points to the system. And the system, as we have just seen, cannot answer.
Digital humanism refuses the rule of Nobody. It insists that for every consequential decision an AI system co-produces, there must be a human being or an institution that can be named, addressed, and called to account. This is also the principle the AI Act enshrines in Article 14, the so called human-in-the-loop approach: high-risk AI systems must be designed so that natural persons can effectively oversee them throughout their operation, and so that no consequential decision is delegated wholly to the AI-Systems.
If we permit AI to develop without these humanist-European anchors, we undergo a structural transformation into a new economic order, the first contours of which are already discernible within Europe itself. I have come to call this order AI-Neocolonialism. A small number of multinational technology corporations hold the models, the compute, the data, the political access, and the power to shape what counts as knowledge, and they extract substantial commercial rent from this position. All others, the European citizen included, supply what these systems run on: data. Data is the raw material of the 2020s, the Age of AI, scalable without limit, never depleted by use, and mined continuously from our attention, our public discourse, our educational content. And in this mining, we are addressed by these systems as transparent consumers rather than as the subjects of free judgement that our humanist European tradition presupposes us to be.
What follows from this? First, every European citizen should have the right to a human review of any consequential automated decision, whether made by a public authority or by a private company.
When the state grants a benefit, scores a child for protective intervention, or assesses a defendant for risk, or when a company decides on credit, hiring, or platform access, the citizen must have a named human being or institution who can be made to answer. Article 22 of the GDPR already establishes this principle, and the Parliament can now strengthen it into a positive entitlement that binds both.
The Digital Omnibus of November 2025 moves in the opposite direction. In the name of simplification, it suspends the high-risk obligations of the AI Act until December 2027, which includes the citizen’s right to an explanation for an automated decision; introduces a new “AI” legitimate interest under the GDPR which allows AI development to override the right to data protection; and permits providers to classify their own systems as not high-risk without publicly registering that decision, removing the only transparency safeguard that existed.
Where the AI Act and the GDPR treat the human being as the subject of judgement, the Omnibus treats the human being as raw material for AI development. The right to a named, accountable decision-maker is the constitutive feature of democratic governance under conditions of automation, and no simplification package should erase it.
Second, when AI is deployed in the workplace, workers must have a right of co-determination. Where AI changes the conditions of work, through algorithmic scheduling, through automated evaluation, through workplace surveillance, the workers whose conditions are being changed must be consulted before the system is in place, not informed once it is. The European Pillar of Social Rights already commits us to this principle, and the European Works Council Directive can be amended to extend it to AI specifically. And the empirical case is firm. Daron Acemoglu, in his 2024 study for the National Bureau of Economic Research, estimates that the productivity gains of generative AI over the coming decade will amount to less than one per cent of total factor productivity, and that these gains will accrue overwhelmingly to capital, with almost none reaching labour.
Third, Europe must establish a European Public AI Trust to govern both the compute infrastructure announced under InvestAI and the multilingual training-data layer beneath it. The governance model is CERN, where a treaty-based council brings together member states, the scientific community, civil society and the Article 17 partners, all bound by charter to European fundamental rights and to public access. Europe has done this before. The post-1945 public-service broadcasting tradition rests on the principle that certain goods are degraded when they are surrendered to the market, and the basic infrastructure of public reasoning, of communication, of education, is among them.
Let me return to the image I began with. Europe is a mosaic. Hands accountable to no rule, with a hammer poorly made, will keep striking the stones until their forms blur, their colours fade, their edges dissolve, and from any distance the picture is no longer there. Thank you.
'Quo vadis, AI? A Humanist Approach for Europe', Humanists International