The Button That Stops Nothing

Deutsche Fassung

Rüdiger Safranski has written a book about artificial intelligence. That is good news to begin with: at last, someone is arguing about AI who was demonstrably capable of thinking before ChatGPT arrived.

The book is called The Fourth Humiliation. Its logic runs like this: Copernicus removed us from the centre of the universe, Darwin removed our special rank above the animal world, and Freud removed our command of our own house. Now a machine performs intellectual tasks for which we previously thought consciousness was necessary — ours, to be precise. The fourth humiliation.

Safranski’s consolation, as the Austrian broadcaster ORF reports his position, is that AI remains dead matter. It has no outside, but it also has no inside; it has no idea what it is doing and merely processes according to rules.

That is elegant. It is also almost right. The point where it becomes too narrow is where things get interesting.

The contest nobody announced

“Human versus machine” is a comparison, and comparisons demand a ranking: which is better, the brain or the chip? The question sounds profound. It is not. An aeroplane is not “better” than a bird. It simply flies differently, and nobody would think of writing a book about the humiliation of the blackbird.

Yet we remain trapped in this contest format. You can hear it in the word “superior”, without which no AI debate seems complete. The ranking question conceals precisely what matters: not only what the machine is and what the human is, but what emerges between them. The relationship. The interplay. The place where it is actually decided what both will become.

Thirty-eight seconds on the fourth humiliation: Thomas Schneider on blackbirds, aeroplanes and the right to stop the machine.

Design question: Which of our AI debates would become productive at once if we banned the word “better” for a season?

What the terms reveal when we sort them

Logic is a system of rules for valid conclusions: formal, low in context, ending in yes or no. Chips are at home there, and we lose. Admitting that costs us nothing. Calculators did not humiliate us either; at least nobody wrote a book about it.

Reason is more: offering reasons, accepting reasons and — Safranski’s strongest point — examining one’s own goals. Instrumental reason calculates a route to a goal. Reflective reason asks whether the goal is any good and knows that it may be wrong. According to ORF, Safranski calls AI the “monstrous materialisation of instrumental reason”. It is a good line. We should only finish the thought: instrumental reason did not wait for AI. It has occupied targets, quarterly logics and process manuals for decades. The machine did not conquer thought. It merely moved into rooms we had already drained of life. The furniture was waiting.

Thinking is more again: doubt, detour, association, pause. Thinking can stop. Discourse is an exchange of reasons among people who risk something with their position — their reputation, their certainty, their sleep. This is where the machine leaves the room, not because it is stupid. It does not fail at thought; it fails at having something at stake. An entity that can lose nothing does not exchange reasons. It exchanges output.

Two real differences remain, and both are descriptions rather than victory ceremonies. First, a technical system can turn a request into an output, a rejection or a hand-off. NIST describes such human–AI configurations, while selective prediction models show that a technical reject option can be designed. Human judgement may still remain suspended: “I don’t know” is a state for us, not a system crash. A rejection is not the machine exercising moral agency. It is an operating decision. Who authorised it, and who carries the consequence?

Second comes the more comic finding: in normal operation, AI always answers. Ask it at three in the morning about the meaning of life and it will provide five bullet points before breakfast. Safranski says it does nothing. More sharply: it cannot decide, on its own initiative, to do nothing. That is its actual defect.

For now. The defect is not a law of nature but a design choice. Systems can be built to ask back, mark uncertainty or hand a decision to a human being. This is not yet human hesitation: the machine risks nothing and carries no responsibility. The pause becomes valuable only when it opens a genuine examination and someone can defend it against time pressure and sanctions. This is why NWNC uses the image of the lazy AI: a machine that does not throw five bullet points at every question, but occasionally puts its feet up and makes us think first. The rarity of such boundaries tells us less about the technology than about those commissioning it.

Design question: Where does our work still permit a pause — and where did we optimise it away long before AI arrived?

The workshop objects

This is where my experience departs from the theory. The consensus about creativity says: people have it, machines do not, end of announcement. I have worked with editing machines, code and projections since the early 1990s, and I cannot accept that consensus. Even machine errors can become remarkably creative. The glitch, the wrong connection, the image nobody ordered: it becomes creative when someone treats it as material rather than a disturbance. At the editing desk this was never a thesis. It was daily work. The material answers, and those who listen receive something they could not have made alone.

Creativity, in that account, is not a substance stored in the brain or absent from the chip. It is an event in an encounter: between person and material, but also among people who interpret, argue and discard. This is why the binary of dead and alive is insufficient. Dead material becomes alive when it enters an arrangement that people can enter, test and rebuild. That is not mysticism. It is montage. And it meets Safranski’s own hope, as reported by ORF: living spirit could direct the dead so that it serves life. In the workshop that means editorial sovereignty. Who may intervene? What remains revisable? Where can experience correct the system?

The last difference may be the sharpest: logic versus taste. Logic decides according to rules. Taste judges without a fixed rule — saturated with experience, embodied, never entirely derivable. A machine can imitate taste statistically. It condenses patterns from past judgements and may produce surprises. But it does not stand up for a departure. Taste is not merely the average of earlier judgements. It is an embodied judgement that can deviate from convention and answer for doing so. Every film editor would recognise Safranski’s thought that wisdom includes the absence of perfection. She might only add: the same is true of organisations.

Design question: Who may depart from the average in our systems — and what does it cost them?

The humiliation as a building brief

Let us put it in order. Safranski is right: AI is dead matter. It has no inner life, risks nothing and cannot remain silent out of its own responsibility. But any organisation that only calculates is dead too — and those organisations existed before AI; now they merely run faster. A system becomes alive not through a claimed inner life, but through revisability, dissent, stop rights and the ability to exit.

According to ORF, Safranski calls for a separation of powers against monopolies. We need to carry that demand down one floor: not only into constitutions, but into everyday work. Who can challenge an AI judgement before it takes effect? Who has a stop right inside the process, rather than only in the law? Who can leave without falling into poverty?

And because delay does not sanctify itself — bureaucratic slowness is not yet judgement — this possibility has a cost centre. A stop right without paid time, protection from sanctions and decision power is only a handsome button on a machine that keeps running. The technical objection remains: what nobody can inspect is difficult to challenge. Inspectability does not replace power, however, and power without inspectability remains blind.

Whether such rights work in daily practice is not a promise but a question of design and evidence. In complex systems, not every intervention has a linear effect. If it succeeds, the fourth humiliation could be the first one that builds something for us instead of merely taking something away. The machine wins at logic; granted. Taste remains a human responsibility. Creativity belongs to neither side: it emerges between them, sometimes even from the machine’s error.

We only need to build rooms in which dead matter may answer us — and secure the right not to answer back. At least until breakfast.

Sources and transparency

Safranski’s positions and the short quotations are treated here through ORF’s report, not as a review of the complete book: ORF, “Plädoyer für ‚lebendigen Geist‘ gegen KI”. Technically, an AI system can be designed to abstain, ask for clarification or hand a decision to a human. NIST documents patterns for human–AI oversight; SelectiveNet is an example of a model that can reject uncertain cases. Neither source gives the system responsibility: responsibility remains with people and institutional design.

The article transparency page makes sources, review steps and technical integrity records readable. Human editorial responsibility remains explicit, and the current blockchain status is shown separately.