Source: French to English Tester Published on: 2026-04-23
Source: The Conversation – in French– By Caroline Gans Combe, Associate professor Data, econometrics, ethics, OMNES Education
As artificial intelligence systems take over decisions once made by humans, the question is no longer just what they can do, but what we ourselves stop doing. Quietly, it is our capacity for judgment that is changing, with potential effects on democracy.
In 1963, while observing Adolf Eichmann during his trial in Jerusalem, Hannah Arendt formulated one of the most disturbing insights of the 20th centurye20th century: this man was not a monster. He had simply stopped thinking. By delegating his moral judgment to the bureaucratic apparatus, he had become capable of participating in a radical evil without ever recognizing it as such. Arendt calls this phenomenon the “banality of evil”: the moral catastrophe does not require a wicked intention – it requires onlythe systematic erosion of the capacity to think for oneself. It identifies a specific sequence:before the political collapse comes the cognitive collapse. Totalitarian regimes first conquer minds by making independent thought structurally superfluous. Sixty years later, we are building something whose philosophical architecture would have been immediately familiar to them.
Delegating judgment at the company level
Agentic artificial intelligence refers to systems capable of perceiving, planning, deciding, and acting autonomously. According to the consulting firmGartner, by 2028, 33% of enterprise software will incorporate AI agents and 15% of daily work decisions will be made autonomously. Another consulting firm,McKinseyemphasizes that these architectures introduce risks that existing governance frameworks were never designed to manage, precisely because they adapt their behavior to the context rather than obeying a predetermined logic.
This is not conventional automation: agentic AI reconfigures judgment itself. However, it is less about the disappearance of human judgment than about a continuous displacement of its object – without humans always being able to perceive its existence. The question is therefore not about delegation in itself, but about the quality and irreversibility of the ongoing reconfiguration.
When democracy delegates its own decisions
The most revealing examples are political, and they are already among us. Indeed, Arendt’s approach distinguishes action (the capacity to initiate something truly new through engagement with others in the public sphere) from mere behavior, which is predictable and manipulable. Democracy is not a mechanism for processing preferences: it is a space of appearance where distinct individuals, capable of independent thought, come together to create meaning collectively. Every decision absorbed into an automated agentic process, every moment of deliberation replaced by an algorithmic result, then represents a cumulative erosion of this space.
In 2016, behavioral targeting algorithms — deployed duringBrexit and Trump election campaigns– selected and amplified emotional content from targeted voters. It was not classical propaganda: it was the technique described by Arendt – not to persuade, butto submerge, create a condition in which the distinction between fact and fabrication collapses under the volume of contradictory stimuli. Algorithmic scoring systems then evaluated citizen reliability, predicted judicial recidivism (like COMPAS in the United States, whose racial biases have been documented by independent mediaProPublica) or oriented theallocation of social housing in the Netherlands and the United Kingdom. In each of these cases, a decision with a heavy ethical burden is absorbed into an opaque process, removed from public debate.
Beyond these specific examples, what is at stake can be considered systemic: it is no longer a question of directing content within a preexisting space of deliberation, but of gradually transforming the very conditions in which judgment is exercised. When three or four major language models mediate an increasing share of the informational interactions of billions of individuals, the main risk is not that they guide responses – it is that they silently harmonize what can be formulated, recognized as a legitimate question, treated as a valid argument. Asked about redistribution policies, the dominant modelstend to frame the responsein terms of economic efficiency – not by censorship, but because their corpora massively overrepresentcertain intellectual traditions. Other registers (justice, dignity, intergenerational solidarity) become less fluid, gradually marginalized without anyone having decided it. When the same models mediatize citizens from distinct traditions, a cognitive convergence sets in,without democratic deliberation. Yet, this is exactly what totalitarian regimes aimed at (and aim at) as a priority, not the opinions themselves, but the space in which different viewpoints can be expressed, meet, and confront each other.
Axiological shift: transforming what counts as value
The deepest danger therefore goes beyond occasional substitution. By increasingly and systematically replacing human decisions, AI systems transform what a society recognizes as value in the sense of what deserves to be considered and/or preserved – efficiency rather than fairness, engagement rather than truth, predictability rather than uniqueness. This process – which I call axiological displacement – is silent, progressive, and operates both through institutions and through daily practices and market dynamics, including where no institution orchestrates it. When algorithmic platforms determine which information is visible and which debates circulate, they do not filter a preexisting public space – they reconstruct it according to their own logic of optimization. The plurality that democracy presupposes is threatened by a convergence toward what the system determines as optimal.
Two dimensions of the same transformation
This trip may be prepared in advance. Longitudinal studies published inJAMA Pediatricshave observed correlations between intensive use of social networks during adolescence and changes in neural circuits – reduced activation of regions associated with rational judgment, hypersensitivity to social feedback. These results call for caution: they do not allow for making global judgments about entire generations. What seems more certain is what the cultural philosopherByung-Chul Hanhighlighted: the saturation of attention by digital flows reduces spaces for critical distance. Agentic AI could amplify this dynamic by giving it an institutional form. Reconfiguration of attention and delegation of judgment may not be two separate problems, but two dimensions of the same transformation for which we still lack the perspective to measure all the consequences – and which AI governance cannot ignore.
Rethinking governance
Arendt was not technophobic. Her project was constructive: to preserve the conditions in which the plurality of viewpoints, the spontaneity of action, and the capacity for singular judgment — notions at the heart of her political theory — can survive. This requires governance frameworks that integrate axiological impact assessments: systematic analyses of how agentic deployment transforms not only operational efficiency but the very foundations of democratic deliberation. It requires preserving spaces where decision-making is truly constructed by actors capable of independent thought — not as symbolic consultations, but as structural elements of the decision-making process. Finally, it requires rethinking the learning of judgment in a world where AI is omnipresent.
The question “Can an AI do this?” must always be accompanied by another, Arendtian in its demand: What happens to us, politically, when we stop doing it ourselves? The real answer is not in refusing delegation, but in rigorously defining its boundaries – which decisions are non-delegable, or under what modality delegation can remain compatible with democratic vitality.
Hannah Arendt spent her life understanding how civilized societies could produce the conditions for their ownmoral collapse. His answer was disturbingly simple: not through a dramatic failure, but through the silent, cumulative, mundane abdication of the responsibility to think. We may be witnessing the beginnings of a new collapse, but we can guard against it before the systems to which we delegate our judgment end up shaping, in our place, what we are.
Caroline Gans Combe received funding from the European Union as part of the Deform and ProRes projects.
–ref. Delegating our decisions to AI: a threat to democracy? –https://theconversation.com/delegating-our-decisions-to-ai-a-threat-to-democracy-276697