Die KI-Dividende: Epistemische Allmende, kognitive Wertschöpfung und gesellschaftliche Teilhabe

Abstract (English) Large Language Models are built on the collectively accumulated knowledge of humanity — an epistemic commons that is being privatized while the knowledge workers whose labor made these models possible are threatened by automation. Because LLMs derive their capabilities from humanity's collective intellectual heritage, the question of who benefits from this value is a matter of distributive justice, not merely of market efficiency. This paper develops multiple mutually independent lines of argumentation that converge on the same conclusion: humanity has a justified claim to participation in AI value creation. The justifications range from a maintenance obligation (the obligation follows from input received, not output quality) through the variation thesis (securing all as a rational obligation), the variation cycle (AI permanently requires human variation), and Kantian ethics (contradiction in conception) to the Logos tradition (knowledge belongs to no one) and the Human-LLM analogy (which compels a revision of intellectual property). The strength of the argument lies in its convergence: because the justification lines rest on largely independent premises, they diversify the normative foundation — even if individual lines are challenged, the remaining ones sustain the core claim. A Five-Pillar Model is proposed: (1) AI value creation levy, (2) public data royalties, (3) direct citizen dividend, (4) public AI infrastructure, and (5) a coexistence guarantee for individual copyrights. Zusammenfassung (Deutsch) Große Sprachmodelle basieren auf dem kollektiv akkumulierten Wissen der Menschheit — einer epistemischen Allmende, die privatisiert wird, während die Wissensträger durch Automatisierung bedroht werden. Dieser Artikel bündelt mehrere voneinander unabhängige Argumentationslinien, die auf denselben Schluss konvergieren: Die Menschheit hat einen begründeten Anspruch auf Teilhabe an der KI-Wertschöpfung. Die Begründungen reichen von der Unterhaltspflicht (die Pflicht folgt aus dem empfangenen Input, nicht aus der Qualität des Outputs) über die Variationsthese (Absicherung aller als rationale Pflicht), den Variationskreislauf (KI braucht dauerhaft menschliche Variation) und die kantische Ethik (Widerspruch in der Konzeption) bis zur Logos-Tradition (Wissen gehört niemandem) und der Mensch-LLM-Analogie (die eine Revision des geistigen Eigentums erzwingt). Die Stärke der Argumentation liegt in ihrer Konvergenz: Da die Begründungslinien auf weitgehend unabhängigen Prämissen beruhen, diversifizieren sie das normative Fundament — selbst wenn einzelne Linien angefochten werden, tragen die verbleibenden den Kernanspruch, wenngleich auf unterschiedlichen Begründungsebenen. Es wird ein Fünf-Säulen-Modell vorgeschlagen: (1) KI-Wertschöpfungsabgabe, (2) Public-Data-Royalties, (3) direkte Bürgerdividende, (4) öffentliche KI-Infrastruktur und (5) Koexistenz-Garantie für individuelle Urheberrechte. Changelog v10.4 (14 June 2026): German-synchronization maintenance revision after the v10.3 source-correction release. The English PDF is unchanged from v10.3. The German version was aligned with the v10.3 English style pass: the market-access and Brussels-effect wording was softened, the shared burden-of-refutation formulation was made more precise, the implementation conclusion was calibrated to avoid a stronger proof claim, and the missing German label for the Pillar II subsection was restored so the EN/DE anchor structure is again parallel. No new source family, core argument, or policy claim is introduced. GER and Kombi PDFs were rebuilt locally on 2026-06-14; the Kombi PDF keeps EN first and GER second. Changelog v10.3 (12 June 2026): Source-correction and English-style maintenance revision. The legal-basis discussion now separates the Digital Services Act and the Digital Markets Act instead of treating both under a shared 6% sanction formula: DSA Article 52 is stated as allowing fines up to 6% of worldwide annual turnover, while DMA Article 30 is stated as allowing fines up to 10% of worldwide total turnover and up to 20% for repeated infringements. The embedded bibliography note for the DSA/DMA reference was updated accordingly. The English version also received a small style-maintenance pass after the source correction, tightening the formulation of regulatory integration, innovation incentives, gatekeeper/Brussels-effect language, and the shared burden of refutation. No new core argument, source family, or policy claim is introduced. EN/DE/Kombi PDFs were rebuilt locally on 2026-06-12. Changelog v10.2 (19 May 2026): Source-check maintenance revision. The litigation/discovery discussion of New York Times v. OpenAI was dated more precisely to the November 2025 production order and the 2 December 2025 denial of reconsideration. The market-model paragraph on data/content licensing deals was recalibrated: Reddit/Google remains the only listed example with a publicly reported annual amount; AP/OpenAI and Stack Overflow/Google are now described as agreements or partnerships without disclosed financial terms; Axel Springer/OpenAI is described as an official partnership with external press reports of tens of millions of euros per year rather than a specific EUR 100+ million claim. EN/DE/Kombi PDFs were rebuilt locally on 2026-05-19. Changelog v10.1 (18 May 2026): Disclosure-maintenance hotfix. The English and German AI disclosure sections were updated to the current neutral L5 convention: AI systems are not listed as authors, co-authors, acknowledgement recipients, ground-truth authorities, or independent validation instances. The author's full responsibility for all theoretical contributions, conceptual decisions, substantive conclusions, final editorial choices, and possible errors is stated explicitly. No substantive argument, source claim, or policy conclusion was changed in this maintenance version. EN/DE/Kombi PDFs were rebuilt locally on 2026-05-18. Previous changelog v10: Source-check and design-check revision for v10: corrected EU AI Act Article 53 timing (Chapter V applies from 2 August 2025; pre-existing GPAI models have a transition deadline of 2 August 2027), narrowed Kneschke v. LAION to the scientific TDM exception and machine-readable rights reservations, narrowed GEMA v. OpenAI to memorisation and output reproduction of protected song lyrics, corrected Wikimedia Enterprise from exclusive content licensing to paid enterprise API/data access and infrastructure support, reframed the Voss report as a non-binding European Parliament own-initiative resolution rather than enacted levy law, and rebuilt EN/DE/Kombi PDFs locally on 2026-05-12. Changelog v10.5 (planned, July 2026): Data presentation, structural comparison, labor market guardrail, and citation-synchronization maintenance revision after the v10.4 release. The 25 July 2026 data presentation upgrade expanded Table 4 (Five-Pillar Framework) in both English and German to include analytical dimensions for impact level (Wirkungsebene) and primary collection body (Primäres Erhebungsorgan). A new synoptic comparison table (Table 5 / Section 6.7) systematically contrasts the Five-Pillar Model with Universal Basic Income (UBI), Data-as-Labor (Lanier/Posner), Robot Taxes (Gates/Stiglitz), and pure copyright approaches across five core dimensions. Earlier July updates recalibrated the labor-market exposition around job transformation, task exposure, and conservative guardrails (Freund/Mann 2026, World Bank 2025, New York Fed 2026), integrated the Stanford HAI AI Index 2026 macro anchor, aligned the employer-side efficiency dividend framing, and brought German bibliography metadata into 100% parity with the English apparatus (134/134 citation keys). No new source family, main claim, or policy proposal is introduced. EN (49 p), GER (55 p), and Kombi (104 p) PDFs are local maintenance artifacts.