Der Mensch in der Maschine: Sind LLMs vernünftige Wesen im kantischen Sinne?
Abstract
Current research — English This ongoing research project develops a seven-dimensional Kantian diagnostic framework for large language models, agent systems and deployment configurations. It distinguishes conceptual conditions of rational agency from empirically testable task proxies. The current evidence supports selected functional analogies but does not positively identify transcendental unity or moral autonomy. It establishes neither their absence nor an architecture-independent impossibility. The framework is conceptual and heuristic, not a validated test battery or numerical ranking. Theory-of-Mind performance depends on tasks, prompts and configurations; behavioral success is not by itself a bridge to Kantian rational agency. The current reading distinguishes formal from functional language competence, clinical phenomena from transcendental claims, and contrastive activation-vector detection from consciousness or a unified subject. Street's sixth-order comparison reports a descriptive GPT-4 advantage against itemwise human-majority answers, with a nonsignificant paired comparison at that order. Strachan's main battery involved 250 human participants and 15 chats per test/model; framing and likelihood controls limit raw faux-pas interpretations. Selected human perturbation errors also prevent a general LLM-specific exclusion of Theory of Mind. These are source-bounded reports, not a new model–human study. The public v9.0 PDFs, later unpublished local v10 working sources and the present v10.0 maintenance successor are distinct provenance layers. Exact published v9 TeX was not identified. Four historical compound reasons were withdrawn; this does not amount to a withdrawal of the paper's entire conceptual framework or a refutation of every component. D3/D7 construct identification, D2 not directly tested, a direct D5 protocol and D6 methods/gold review and study remain open. The D6 instrument's 16 items, 64 planned cells and local fixtures are not executed study results. Aktueller Forschungsstand — Deutsch Dieses laufende Forschungsprojekt entwickelt einen siebendimensionalen kantischen Diagnose-Rahmen für Sprachmodelle, Agentensysteme und Deployment-Konfigurationen. Es trennt begriffliche Bedingungen vernünftiger Handlungsfähigkeit von empirisch prüfbaren Aufgabenproxys. Die aktuelle Evidenz trägt ausgewählte funktionale Analogien, identifiziert aber transzendentale Einheit oder moralische Autonomie nicht positiv. Sie belegt weder deren Abwesenheit noch eine architekturunabhängige Unmöglichkeit. Der Rahmen ist begrifflich und heuristisch, keine validierte Testbatterie oder numerische Rangordnung. Theory-of-Mind-Leistung hängt von Aufgaben, Prompts und Konfigurationen ab; Verhaltenserfolg bildet allein keine Brücke zu kantischer rationaler Handlungsfähigkeit. Die präzisierte Lesart trennt formale von funktionaler Sprachkompetenz, klinische Phänomene von transzendentalen Aussagen und Detektion kontrastiver Aktivierungsvektoren von Bewusstsein oder einem einheitlichen Subjekt. Streets Vergleich sechster Ordnung berichtet einen deskriptiven GPT-4-Vorsprung gegenüber itemweisen menschlichen Mehrheitsantworten, bei nicht signifikantem gepaartem Vergleich dieser Ordnung. Strachans Hauptbatterie umfasste 250 menschliche Teilnehmende und 15 Chats je Test/Modell; Framing- und Likelihood-Kontrollen begrenzen die Interpretation von Faux-Pas-Rohscores. Auch menschliche Fehler bei ausgewählten Perturbationen verhindern einen allgemeinen LLM-spezifischen ToM-Ausschluss. Dies sind quellengebundene Berichte, keine neue Modell-Human-Studie. Die öffentlichen v9.0-PDFs, späteren unveröffentlichten lokalen v10-Arbeitsquellen und der jetzige v10.0-Wartungsnachfolger bilden getrennte Provenienzschichten. Die exakte veröffentlichte v9-TeX-Quelle wurde nicht identifiziert. Vier historische Sammelbegründungen wurden zurückgenommen; daraus folgt weder ein Widerruf des gesamten begrifflichen Rahmens noch die Widerlegung jedes Teilsatzes. Die Konstruktidentifikation für D3/D7, D2 ohne direkten Test, ein direktes D5-Protokoll sowie D6-Methoden-/Goldreview und Studie bleiben offen. Die 16 Items, 64 geplanten Zellen und lokalen Fixtures des D6-Instruments sind keine ausgeführten Studienergebnisse. Changes since public v9.0 — v10.0 Replace unsupported absence and structural-impossibility conclusions with nonidentification limits, including direct RLHF, duty-motivation, Floridi, table and conclusion consumers. Correct D1/D3/D5 allocation, preserve B72's undecided scope and the bounded AA 4:429 German wording/English paraphrase; full edition concordance remains unverified. Keep ToM-to-Kant, formal/functional-to-analytic/synthetic and clinical-to-transcendental bridges explicitly open; correct Lindsey's contrastive-vector mechanism. Scope Street's sixth-order reference and nonsignificant paired result, and Strachan's battery, framing/likelihood controls and selected human perturbation findings. Carry forward the later local v10 framework, evaluation-awareness, evidence-matrix and layout developments without claiming all arose after public v9.0. Separate historical R26 withdrawals from current source reports and the unexecuted D6 design. Changes in 8.4 (June 2026): Bibliographic metadata and evidence-structure maintenance — added arXiv DOI metadata for Sparks of Rationality, Kadavath et al. 2022, Seo 2024, and Ullman 2023; aligned the German bibliography's preprint markers for Kadavath et al. 2022, Seo 2024, and Ullman 2023 with the methodological caveat; added a compact D4/D6/D7 evidence bridge before the global assessment in both language versions, separating metacognition, Theory of Mind, and moral-autonomy evidence from the final Kantian diagnosis. No change to the article's core argument. Changes in 8.3 (May 2026): Bibliographic source verification update — added missing DOI, page, issue, and article-number metadata for Bender et al. 2021, Evans 2022, Fazi 2024, Gigerenzer 2024, Lampinen et al. 2024, Macmillan-Scott and Musolesi 2024, Manna and Nath 2021, Orbik 2024, Schüring and Schmid 2024, and Wu et al. 2025; expanded Cheung et al. 2025 and Mitchell 2024 with volume/issue/article metadata; completed the Sacco 2025 title and DOI; moved Wu et al. 2025 from the preprint caveat to the version-of-record list. No change to the article's core argument. Changes in 8.2 (May 2026): Bibliographic source verification update — Xu et al. corrected back to arXiv/CoRR metadata rather than an unverified ACL 2024 proceedings entry; Shin et al. corrected to Journal of Yeungnam Medical Science with DOI; Street et al. updated to the 2026 Frontiers in Human Neuroscience version of record; Qiu et al. updated to Nature Communications 17, 1238 (2026); Tak et al. corrected to 2026 arXiv metadata; Kambhampati, Kosinski, Griot, Fontana, Steyvers, Liu, Sclar, and Cheung DOI/venue details expanded. No change to the article's core argument. Changes in 8.1 (April 2026): Bibliographic correction — Pi et al. updated from arXiv-only preprint metadata to the CogSci 2025 proceedings record; no argument-level changes. Changes in 8.0 (March 2026): Bibliographic corrections — Systematic verification and correction of all references against original publications. Corrections: Mitchell/Science (author), Sparks Rationality (subtitle), Lampinen (title). Changes in 7.0 (March 2026) Two full internal review cycles Major: Added radar diagram, heatmap, distributional rationality concept, Tier 1/2/3 methodology, Reflexive Transparency section, Practical Implications, embodied AI analysis, reasoning models in D4, structural/contingent criteria, Other Minds Problem Minor: Fixed 5 BibTeX keys, updated 3 preprints to peer-reviewed, compressed existentialists and Section 10, added functionalism steelman, CoI/Funding statements DE/EN: German version fully synchronized. Kombi-PDF deprecated. Changes in Version 9.0 (July 2026) Source and claim corrections — replaced a compound emergence claim not supported by Wei et al. with a benchmark-bounded formulation; corrected the description of OpenAI o1 so that raw chain-of-thought is not presented as user-visible; calibrated the Butlin et al. consciousness conclusion to the report's evidential wording; added the supporting OpenAI primary source. German and English versions were updated in parallel.