When AI Judges: Who Still Learns How to Decide?
The AI has prioritised three applications, marked two risks in red and drafted a recommendation for the meeting. Everything looks reasonable. The sources are linked, the prose is smooth, confidence sits at 87 per cent. Only one small thing is missing: someone still has to judge.
That sounds like the final human step. In reality, it is the first political one.
AI does not merely take tasks off our hands. It evaluates, sorts and pre-structures decisions. It influences which possibility remains visible, which risk is taken seriously, who counts as suitable and which deviation is treated as noise. This makes AI more than another productivity tool. It is a force of social transformation.
It intervenes simultaneously in work, education, public administration, the public sphere, value creation, power architectures and the organisation of judgement. It changes not only how we work, but what will count as work, competence, career, institution and human contribution.
The decisive question is therefore not whether AI can decide. It already participates in decisions. The question is which people and institutions will still learn to examine a pre-decision, challenge it and stop it in time.
Judgement is not an innate luxury feature
Humans were frequently wrong before AI. Experience is a long sequence of judgements, consequences and corrections. We decide, observe, fail, adjust the standard and later fail again at a slightly more sophisticated level. This is not particularly efficient. It is, however, how judgement develops.
When a plausible AI answer appears before our own attempt, that learning path changes. A young doctor, a new employee or a student no longer sees the messy material first. They see a world that has already been ordered. That helps. It can also prevent personal search patterns, doubts and standards of comparison from developing.
Research offers no simple age formula. A study by Mata and colleagues found age-related differences in the selection of decision strategies. It does not follow that the young are naive or the old are wise. Domain knowledge, practice, motivation, cognitive load and feedback often matter more than year of birth.
Experienced people can compare an AI answer with real cases. They know exceptions, side effects and the smell of a number that is too clean. Yet experience does not immunise anyone against automation bias. When people regard a system as reliable, they can overlook contradictory evidence. The machine then does not merely confirm an error. It gives it the charm of objective office equipment.
The useful line therefore does not run between young and old. It runs between people who acquire decision practice and people from whom that practice is removed.
AI can relocate thinking and spare us from thinking
A CHI study on generative AI in knowledge work describes an ambivalent shift. Critical thinking does not simply disappear. It moves from execution towards goal setting, verification and integration. At the same time, greater confidence in GenAI was associated with less reported critical effort.
This is not proof that AI makes people less intelligent. It indicates a design problem. Anyone who receives an answer effortlessly needs a reason to enter the material again. In organisations, that reason is rarely romantic curiosity. It has to be built into roles, time and rights.
Cognitive forcing functions provide one example. People might have to state their own assessment, list alternatives or mark uncertainty before seeing the system recommendation. Such interventions can reduce overreliance, but they introduce friction.
That is precisely what makes them interesting. The most important human function in an AI system may be the one every product demo wants to remove: a well-founded detour.
Democracy begins before the convenient answer
What looks like interface design in the office becomes a question of power in a democracy. Recommendation and language systems structure the public sphere. They influence which explanation people encounter first, which position looks marginal and which political action arrives pre-labelled as reasonable.
Kyrychenko and van der Linden therefore call for societal alignment in Social technologies need societal alignment. Evaluation must not end with the correctness of an individual answer. Procedures, participation, fairness and social consequences also have to be examined.
That is the right standard. Democracy does not depend on the right side always winning. It depends on decisions remaining contestable, power becoming visible and errors being corrected before they harden into infrastructure.
An AI that supplies the most probable answer in the name of a majority is not yet democratic. Perhaps, to borrow Sibylle Berg’s friendly provocation, AI is “communist” because suddenly everyone gains access to an astonishing number of answers. The means of production called prompting sits in the browser. The evaluation criteria, data centres and distribution channels still belong to remarkably few actors. A very modern communism: answers are socialised while ranking remains private property.
No source is free of interests
Judgement begins with an inconvenient insight: interests are not an accident in the production of knowledge. They are always present.
Vendors want usage, data and market share. Platforms want attention. Organisations want speed, standardisation and less liability. Governments want the capacity to act, majorities and communicative clarity. Public agencies want to fulfil their mandate and preserve institutional trust. Media organisations want relevance and reach. Critics want accountability and counter-power, but can also be rewarded for escalation, camp loyalty and recognition. Users want orientation, confirmation and sometimes simply peace of mind.
An interest does not disprove a claim. It does, however, explain which question might be favoured, which uncertainty might be minimised and which side effect might be treated as acceptable.
The traditional source question, “Who says this?”, is therefore no longer enough. We must also ask: Who commissioned the system? What is it optimised for? Which data are absent? Which supervisory relationship applies? Who bears the cost of an error? Who can force a correction?
When official certainty outruns internal scrutiny
A brief German case shows why this symmetrical standard matters. It belongs here not to reopen the entire pandemic debate, but because it reveals the architecture of institutional judgement.
The officially published minutes of the Robert Koch Institute crisis team state on 10 September 2021 that the institute’s scientific independence from politics was “limited in this respect” because of the Health Ministry’s technical supervision. On 5 November 2021, the crisis team internally called the public phrase “pandemic of the unvaccinated” technically incorrect. The minutes also record that the minister probably used it deliberately and that it could hardly be corrected.
In January 2022, the records concerning recovered status likewise documented differences between technical proposals, agreements among political leaders, the legal framework and a ministerial decision. This does not prove that every measure taken at the time lacked a scientific basis. It does prove that a statement by a government institution does not become reliable truth merely because of its sender.
These records did not emerge through voluntary transparency. Multipolar requested them under freedom-of-information law and placed their release under pressure through years of litigation; the first files were heavily redacted. The unredacted files later published by Aya Velázquez came from a leak. Critical analysis therefore made a real contribution to public knowledge. Yet critical interpretations do not become true automatically because they oppose official ones. They must be tested against the same original passages.
The same applies to medical generalisations. Previous infection provided substantial protection according to research; a Delta cohort found it stronger than protection from two vaccine doses among previously uninfected people. At the same time, a systematic review for Omicron found the strongest protection against severe disease among people with hybrid immunity. Neither finding automatically justifies coercion and rigid rules. Nor do they support the claim that vaccination was always medically useless for recovered people.
The lesson is not to trust nobody. It is to grant nobody epistemic immunity.
Seven practices for a society living with AI
If judgement is a practice, we can build conditions for it. New Work does not mean leaving people an attractive residual zone after automation. It means organising capability and counter-power so that people become more capable of action with AI.
First: form an initial hypothesis before seeing the AI answer. Learners and professionals first state what they observe, what they suspect and what is missing. Only then does the machine recommendation appear. Not for every low-stakes task, but wherever learning and judgement matter more than seconds.
Second: build a source ladder, not a source list. Every decision-relevant claim must lead back to the primary source, measurement method and period. A long list of links is not yet provenance. It is often a bibliography with a clear conscience.
Third: display interests beside the result. The commissioning party, optimisation objective, data gaps and affected groups belong next to the recommendation. Not in a document nobody opens, but in the decision space itself.
Fourth: make dissent a paid role. Someone must examine not only whether the system calculated correctly, but whether the question is wrong, the category unfair or the proposed action disproportionate. Dissent without time and authority is decoration.
Fifth: preserve learning paths through real cases. Junior staff need access to raw material, intermediate decisions and errors. Anyone who only approves finished AI syntheses mainly learns how to approve. Three years later, the organisation has many senior approvers and nobody who remembers why case 17 was different.
Sixth: keep decision journals. Record assumptions, counterarguments and expected effects, not just the result. Later, compare them with what actually happened. Without feedback there is no experience, only repetition with a new model name.
Seventh: create real stop rights. A person must be able not merely to comment on an AI-prepared decision, but to suspend it and force a second review. Responsibility without the right to stop is liability with a friendly interface.
These practices help younger and older people in different ways. Younger people receive protected spaces in which to build their own standards. Experienced people receive procedures that interrupt routine, status and self-confirmation. Both learn to treat AI neither as an oracle nor as an enemy, but as a powerful counterposition whose interests and limits must remain visible.
Judgement is a substrate of freedom
The AI debate is often framed as a race for models, chips and compute. This infrastructure matters. Yet a society can catch up technically while forgetting politically how to judge.
Judgement requires time, access to sources, practice, feedback, dissent and power. It is therefore a substrate of freedom: a real condition that allows people not only to choose among machine-sorted options, but to change the sorting itself.
In the AI Flowbook we collect practical patterns for source verification, institutional dissent, human latency and systems that do not colonise every pause. A good AI does not turn people into slower computers. It helps them ask better questions and say no with reasons.
The future will therefore not be decided by whether AI produces increasingly good judgements. In many domains it will. The future will be decided by whether people still have the opportunity to see before the answer, verify after the answer and act against the answer.
Anyone who was never allowed to decide cannot delegate judgement. They can only accept it.
And a democracy that only accepts has not acquired AI government. It has forgotten how to govern.
Hinterlasse einen Kommentar
An der Diskussion beteiligen?Hinterlasse uns deinen Kommentar!