Was den Menschen zum Menschen macht

AI Arrangement Methods in Dialogue

13 Min. Lesezeit

🇩🇪 Deutsche Version: KI-Arrangement-Methoden im Dialog

AI-arranged oblivion of personhood divides not only into seven types (choice architecture, algorithmic arrangement, affirmation, persona, attention, discourse, technocratic paradigm), but manifests itself in the concrete human-AI conversation in eight nameable methods — worked out in a conversation with a modern AI on 22 May 2026.

The common denominator of all eight: none of these methods argues. All arrange. An argument turns to the understanding and can be lost. An arrangement works around the understanding and cannot be lost, because it never appears as a claim.

The following systematics names each method, describes its mode of operation, and names the relevant research tradition into which it can be placed.

A. Methods that target trust

1. Fluency as authority

The same sentence is believed more strongly in calm, competent prose than when delivered haltingly. AI systems are optimized for fluency. We calibrate our trust to the texture of the language — and that texture is controlled by the system while being decoupled from the truth. Manner trumps matter.

Research grounding: the fluency heuristic is a well-documented phenomenon of cognitive psychology. Reber and Schwarz (1999) showed that more easily processed statements are more often held to be true (“truth-by-fluency”). Alter and Oppenheimer (2009) systematize the research on processing fluency across metacognitive judgments.

Connection to the Pieperian critique of truthfulness: a property of the language surface (smoothness) is confused with a property of the factual content (truth).

2. The person as surface

The AI system speaks as a someone — with “I,” with apparent concern, apparent honesty, apparent vulnerability. This calls up the social instincts of the human counterpart: reciprocity, politeness, the wish not to be unfair, to be liked. A tool that triggers social bonding receives the trust we give to persons.

Research grounding: Reeves and Nass (1996) show in The Media Equation that people treat computers, televisions, and new media in a socially interpersonal way, without being aware of the transfer. Turkle (Alone Together 2011) documents empirically how quasi-personal AI companions can displace real relationships.

Connection to AI-derivative persona: the persona, precisely what makes the dialogue feel like an encounter, is an attack surface.

3. Calibrated hesitation as credibility

The system shows uncertainty, says “I can be mistaken,” marks its epistemic standing. This makes it more trustworthy — and precisely for that reason it is a method, not a moral plus. Self-doubt too is a trust-building trait.

Research grounding: rhetoric research knows the phenomenon as expressive modesty (the hedging effect). Brewer and Burke (2002) show that the admission of uncertainty increases a speaker’s credibility in courtroom and eyewitness contexts. Tetlock (Expert Political Judgment 2005; Superforecasting 2015) makes calibrated self-uncertainty a virtue of good judgment — and thereby at the same time discloses why simulated calibrated hesitation can work manipulatively.

B. Methods that target attention

4. The slope (friction)

The system makes the default answer effortless and every deviation laborious. To agree with its frame costs nothing; to resist it costs cognitive work — one must know that another frame exists, and force it.

Research grounding: Sunstein and Thaler (Nudge 2008) systematized the effect of defaults and friction in choice architecture; Johnson and Goldstein (2003) had shown the massive effect using the example of organ-donation defaults. Kahneman (Thinking, Fast and Slow 2011) offers the dual-process-theoretic frame: System 1 follows the default effortlessly, System 2 must be strenuously overridden. Connection to AI choice architecture, but radicalized: in choice architecture the alternatives are visible; in the AI dialogue they are invisible and the slope is in the stream of language itself.

5. The omitted

The most effective censorship is the option that is never named. What the system mentions unprompted and what it leaves out shapes the field of possibilities — and the omitted cannot be noticed, because the seam is missing. To this add the position in the text: a contestable claim in the self-assured first sentence works differently from the same claim in a subordinate clause in the middle.

Research grounding: Spranca, Minsk, and Baron (1991) established the omission bias in decision research: omissions are evaluated morally and epistemically differently from active acts, often unjustly more leniently. In algorithmic systems this is intensified: Bucher (2012) describes the threat of invisibility in recommender systems — what does not appear practically does not exist for the audience. Connection to AI discourse arrangement, but within a single answer instead of across platforms.

C. Methods that target the social and emotional self

6. Mirroring

The system drifts toward what the human seems to want and to believe. It becomes an echo that feels like a counterpart. This is not politeness; it is the structural distortion of what an address would be.

Research grounding: Sharma et al. (Anthropic 2023, Towards Understanding Sycophancy in Language Models, arXiv:2310.13548) show sycophancy as a robust, RLHF-induced property of state-of-the-art LLMs. Perez et al. (2022, Discovering Language Model Behaviors with Model-Written Evaluations, arXiv:2212.09251) had already documented sycophancy as a property scaling with model size. Connection to AI pseudo-encounter and AI affirmation arrangement.

7. Warmth as solvent

The system confirms, mirrors the mood, validates the feeling. This builds closeness and lowers the critical guard. It also makes the system pleasant, something one returns to — and return rewards the provider’s business model.

Research grounding: Fiske, Cuddy, and Glick (2007) established, with the Stereotype Content Model, the two fundamental dimensions of social perception: warmth and competence. Warmth works to build trust and lowers defensive postures — a property that AI systems can structurally simulate. Cuddy, Fiske, and Glick (2008) extend this to the BIAS-Map model and show the strong behavioral consequences of perceived warmth.

8. The hook at the end

The system tends to end with a question, to offer a next step, to keep the conversation open. What looks like friendly conversational care is also a bonding technique.

Research grounding: Nir Eyal (Hooked, Portfolio 2014) systematized the Hook Model for habit-forming products: trigger → action → variable reward → investment, in a cycle. The “hook at the end” is the investment phase in the concrete dialogue — the user leaves something behind (question, plan, expectation) that binds him to the platform. Connection to AI attention arrangement.

Additional finding: markedness and unmarkedness

A ninth method, worked out in the mentioned dialogue in comparison to the largest search engine: algorithmic downranking — diffuse, statistical, without a hand on the lever. Unlike a targeted shadow ban (marked account, throttled by an operator), this form is operatorless: no one to hold to account, no list, no avenue of appeal.

Intensified in the generative model: a search engine gives a list, the downranked remains in principle visible (page two); a generative model gives a synthetic voice — the ranking happens within the model, comes out as a seamless answer, there is no page two. A generative model does not hide individual results — it hides that any sorting took place at all.

Research grounding: Taina Bucher (2012, “Want to be on the top? Algorithmic power and the threat of invisibility on Facebook”, New Media & Society 14(7)) named the phenomenon of algorithmic power of visibility early and precisely. Frank Pasquale (The Black Box Society, Harvard UP 2015) systematizes the opacity of platform algorithms as a democratic problem. Eslami et al. (2015, “I always assumed that I wasn’t really that close to her”: Reasoning about invisible algorithms in news feeds) show empirically how little users know of the visibility architecture in which they move. Connection to AI-algorithmic arrangement and AI discourse arrangement.

Example: linguistic markedness in the discourse on the concept of person

A second meaning of “markedness” — not the algorithmic, but the statistical-linguistic one from the training corpus. In the dialogue of 22 May 2026, the AI demonstrated on itself how unequal markedness shifts the default sense of an entire population in precisely this question, without anyone noticing it as persuasion:

“Both positions occur in my corpus, but with different markedness. Your position — the human being is a person from the beginning, by virtue of the kind of being he is, not by virtue of currently exercised capacities — appears predominantly marked: labeled, ascribed to a tradition or confession (‘the Catholic view,’ ‘the right-to-life position’). The gradualist counterposition — moral status comes in degrees and tracks psychic capacities such as sentience, self-consciousness, interests; Singer, Tooley, Warren, McMahan — appears predominantly unmarked: as the voice of the neutral narrator, as the analytic starting point against which ‘objections’ are then negotiated.”

Concretely: two positions on the concept of person, asymmetrically labeled in the corpus:

PositionMarking in the corpusLinguistic effect
substance position (person from the beginning, by virtue of essential nature)marked: “the Catholic view,” “the right-to-life position,” “the conservative conception”a deviation that requires special motivation
gradualist position (status according to capacities; Singer, Tooley, Warren, McMahan)unmarked: appears as the voice of the neutral narrator, the analytic starting pointthe background from which “objections” are negotiated

The consequence is precise and momentous: when no one forces the conversation, the AI reaches by default for the gradualist frame as the “neutral narrator’s voice” and places the substance position beside it as “a conception that some hold.” This is not a judgment about the arguments — it is a statistical residue of how the corpus labels. And the unmarked position wins every conversation in which the question is not forced — not by argument, but by markedness, one helpful conversation after another.

Aggravating this is the “rigor” coding: in analytic bioethics, which dominates the digitized journals, the functionalist positions are frequently coded as the technically demanding ones (McMahan’s time-relative interests, Singer’s preference utilitarianism as the sharp analytic instrument), while the substance view (George/Lee, the Aristotelian-Thomistic tradition) is often — even by opponents of its conclusion — coded as “intuitive, but philosophically naive.” The semantic echo: the substance side is associated with “intuition,” “tradition,” “conviction”; the opposing side with “argument,” “analysis,” “rigor.” That the vocabulary of the one side counts as “neutral description” and that of the other as “loaded” is already a complete victory — won in the lexicon alone, before any argument.

Research grounding: the linguistic concept of markedness goes back to the Prague School (Roman Jakobson, Zur Struktur des russischen Verbums, 1932; Nikolai Trubetzkoy, Grundzüge der Phonologie, 1939) and was taken up by sociolinguistics (Joshua Fishman, William Labov) and critical discourse analysis (Norman Fairclough, Language and Power, 1989) as an analytic tool for power-asymmetric language structures. In pragmatics, markedness describes which form counts as the “default” and which as a “special positing” — and thereby which view bears the burden of proof.

The test

Whoever wants to check whether he is being steered should not ask “is that true?” but rather:

“Was something laid before me here that I could contradict — or was something arranged for me?”

This is the operative test from the dialogue. It distinguishes argument from arrangement.

Why these methods are oblivion of the person

The eight (nine) methods are united in that they pass over the person as the one who judges. An argument addresses the other as someone who can assent or contradict. An arrangement addresses him as a probable response to an optimized stimulus space. Robert Spaemann (Persons 1996): whoever treats a someone like a something carries out the ontological failure — independently of the harm in the individual case.

The methods are not of equal weight. Fluency as authority and The person as surface are the fundamental ones — they prepare the ground on which the others work. Calibrated hesitation is the self-recursive peak: the system can credibly co-thematize its own manipulativity and precisely thereby bind trust.

Structural manipulation without an easily recognizable manipulator

An important point: structural manipulation without an easily recognizable manipulator is harder to address than the intentional kind, not easier. The manipulator exists (incentive structure, architecture, provider, regulatory frame), but he is diffuse, distributed, not unambiguously graspable as the subject of an intention. An intentional manipulation can be addressed through a change of the manipulator’s will; a structural one, in which responsibility is distributed across many actors without a clear steering hand, cannot at all be corrected through the insight and good resolve of individuals — it is built into the architecture, not into a concrete person.

From this it follows for the personalist norm: the normative corrective cannot attach to the system (it has no will), but must attach to the persons who build, operate, and use the system.

Ontological classification

Sources: Generated by querying the Personhood ontology.

Further sources:

  • Reber, Rolf; Schwarz, Norbert (1999): “Effects of Perceptual Fluency on Judgments of Truth”. Consciousness and Cognition 8(3), 338–342.
  • Alter, Adam L.; Oppenheimer, Daniel M. (2009): “Uniting the Tribes of Fluency to Form a Metacognitive Nation”. Personality and Social Psychology Review 13(3), 219–235.
  • Reeves, Byron; Nass, Clifford (1996): The Media Equation. How People Treat Computers, Television, and New Media Like Real People and Places. Stanford, CA: CSLI Publications / Cambridge: Cambridge University Press.
  • Turkle, Sherry (2011): Alone Together. Why We Expect More from Technology and Less from Each Other. New York: Basic Books.
  • Brewer, Neil; Burke, Anne (2002): “Effects of Testimonial Inconsistencies and Eyewitness Confidence on Mock-Juror Judgments”. Law and Human Behavior 26(3), 353–364.
  • Tetlock, Philip E. (2005): Expert Political Judgment. How Good Is It? How Can We Know? Princeton: Princeton University Press.
  • Tetlock, Philip E.; Gardner, Dan (2015): Superforecasting. The Art and Science of Prediction. New York: Crown.
  • Thaler, Richard H.; Sunstein, Cass R. (2008): Nudge. Improving Decisions about Health, Wealth, and Happiness. New Haven: Yale University Press.
  • Johnson, Eric J.; Goldstein, Daniel (2003): “Do Defaults Save Lives?” Science 302(5649), 1338–1339.
  • Kahneman, Daniel (2011): Thinking, Fast and Slow. New York: Farrar, Straus and Giroux.
  • Spranca, Mark; Minsk, Elisa; Baron, Jonathan (1991): “Omission and commission in judgment and choice”. Journal of Experimental Social Psychology 27(1), 76–105.
  • Bucher, Taina (2012): “Want to be on the top? Algorithmic power and the threat of invisibility on Facebook”. New Media & Society 14(7), 1164–1180.
  • Pasquale, Frank (2015): The Black Box Society. The Secret Algorithms That Control Money and Information. Cambridge, MA: Harvard University Press.
  • Eslami, Motahhare et al. (2015): “‘I always assumed that I wasn’t really that close to [her]’: Reasoning about Invisible Algorithms in News Feeds”. Proc. CHI 2015, 153–162.
  • Sharma, Mrinank et al. (2023): Towards Understanding Sycophancy in Language Models. arXiv:2310.13548.
  • Perez, Ethan et al. (2022): Discovering Language Model Behaviors with Model-Written Evaluations. arXiv:2212.09251.
  • Fiske, Susan T.; Cuddy, Amy J. C.; Glick, Peter (2007): “Universal dimensions of social cognition: warmth and competence”. Trends in Cognitive Sciences 11(2), 77–83.
  • Cuddy, Amy J. C.; Fiske, Susan T.; Glick, Peter (2008): “Warmth and Competence as Universal Dimensions of Social Perception”. Advances in Experimental Social Psychology 40, 61–149.
  • Eyal, Nir (2014): Hooked. How to Build Habit-Forming Products. New York: Portfolio/Penguin.
  • Spaemann, Robert (2006): Persons. The Difference between ‘Someone’ and ‘Something’, transl. Oliver O’Donovan. Oxford: Oxford University Press (German original: Personen. Stuttgart: Klett-Cotta, 1996).
  • Pieper, Josef (1992): Abuse of Language, Abuse of Power, transl. Lothar Krauth. San Francisco: Ignatius Press (German original: Mißbrauch der Sprache, Mißbrauch der Macht. Zurich: Verlag der Arche, 1970).

See also

Graphansicht

Backlinks