Kognitive Auslagerung oder kognitive Schuld? (Cognitive Offloading vs. Debt)

Abstract (English) Background: Kosmyna et al. (2025) reported differences in directed EEG functional connectivity during essay writing with an LLM, a search engine, or no external tool. These measurements are not alpha/beta power or a general measure of cognitive decline. Objective: This ongoing narrative research examines the evidential boundaries of cognitive offloading and cognitive debt and proposes a conditional framework based on timing, interaction design, and context. Methods: Critical narrative comparison of heterogeneous task-specific studies and theoretical accounts; selected primary passages were checked for the current source corrections. Results: Immediate connectivity, recall, performance, and self-report outcomes must be distinguished from durable learning, retention, and clinical efficacy. Fan et al.'s meta-analysis reports a positive average learning-outcome effect, while the inspected September studies concern immediate arithmetic performance, metacognitive self-reports, a search-cue package, or a conceptual software-team framework. These findings neither validate the proposed three-factor taxonomy nor settle long-term harm or benefit. Conclusions: Passive delegation and active integration remain useful design hypotheses rather than validated universal outcome classes. The Session 4 directional interpretation, clinical and neurodivergent transfer, retention, dose effects, and the inventory gates U01–U05 remain open. Recommendations and the cognitive-prosthesis matrix are prospective and require direct evaluation. Zusammenfassung (Deutsch) Hintergrund: Kosmyna et al. (2025) berichteten Unterschiede der gerichteten funktionellen EEG-Konnektivität beim Essay-Schreiben mit LLM, Suchmaschine oder ohne externes Werkzeug. Diese Messungen sind weder Alpha/Beta-Power noch ein allgemeines Maß kognitiven Abbaus. Ziel: Diese laufende narrative Forschung untersucht die Evidenzgrenzen kognitiver Auslagerung und kognitiver Schuld und schlägt einen bedingten Rahmen aus Einsatzzeitpunkt, Interaktionsdesign und Kontext vor. Methoden: Kritischer narrativer Vergleich heterogener aufgabenspezifischer Studien und theoretischer Ansätze; ausgewählte Primärpassagen wurden für die aktuellen Quellenkorrekturen geprüft. Ergebnisse: Unmittelbare Konnektivitäts-, Abruf-, Leistungs- und Selbstberichtsbefunde sind von dauerhaftem Lernen, Retention und klinischer Wirksamkeit zu trennen. Fans Metaanalyse berichtet einen positiven mittleren Lernoutcome-Effekt; die geprüften September-Arbeiten betreffen unmittelbare Arithmetikleistung, metakognitive Selbstberichte, ein Such-Cue-Paket oder einen konzeptuellen Software-Teamrahmen. Diese Befunde validieren weder die Drei-Faktoren-Taxonomie noch entscheiden sie langfristigen Schaden oder Nutzen. Schlussfolgerungen: Passive Delegation und aktive Integration bleiben nützliche Designhypothesen statt validierter allgemeiner Ergebnisklassen. Die Richtungsinterpretation von Sitzung 4, klinische und neurodivergente Übertragungen, Retention, Dosiseffekte und die Inventurgates U01–U05 bleiben offen. Empfehlungen und kognitive Prothesenmatrix sind prospektiv und benötigen direkte Evaluation. CHANGELOG Changes in Version v10.6 (October 2026) This maintenance version narrows the scope of several claims and synchronizes their English and German wording. It does not report clinical efficacy, causal LLM effects, or validated long-term retention outcomes. Proposed factors and Figure 2: The three factors are described as proposed and hypothetical. Figure 2 distinguishes Stadler's reduced germane-load finding from the MIT study's directed EEG connectivity; the measures are not treated as equivalent. Accessibility and neurodivergence: Accessibility is framed as a hypothesis. Direct clinical efficacy and longitudinal effects of generative LLMs remain untested. Wang et al. (2026): The paper describes the student survey/CB-SEM result as statistical mediation through cognitive self-efficacy alone; it does not attribute an SRL mediator, causal LLM effect, or retention result. Hermeneutic framing: The Gadamer passage is presented as a conceptual analogy with testable hypotheses, not as empirical identification of genuine understanding. German-language correction: A grammatical error in the introduction was corrected without changing its intended claim. Bilingual files: The English and German PDFs were rebuilt at 24 and 26 pages. The combined PDF remains a local control artifact and is not part of the public file set. Changes in Version v10.5 (August 2026) Comprehensive live source and citation metadata audit against Crossref, arXiv, and PubMed, followed by structural/table synchronization across English and German manuscripts and a narrow English technical-language micro-pass. Live Source & Citation Audit: Verified all 52 bibliography entries against live Crossref, arXiv, and PubMed APIs. Corrected author metadata and titles for 2026 preprints (Guingrich et al., Yu et al., Poquet et al., arXiv cs.HC), updated Paglialunga & Melogno (2025) and Bastani et al. (2025) authors, and expanded full author lists (Gramann et al., 2017). In-Text Citation Synchronization: Aligned in-text attribution in the evidence ledger table (Iannone & Giansanti, 2024; J. Pers. Med.) with the primary narrative review. Evidence & Table Design: Fully synchronized 3-mode taxonomy matrix, cognitive/metacognitive channel column, and July 2026 evidence balance ledger across both language versions. English Technical-Language Pass: Tightened method wording, softened one over-categorical conflation phrase, and clarified neurodivergence metaphors as user-reported labels, without adding sources or claims. Bilingual Parity: Recompiled English and German manuscripts (23 and 25 pages respectively; current SHA-256 EN 970813875F88CD5029AC71EC60217339DC7887E06D601AC46AC92BF0E672B9BA, GER FB4D0DADB45DF47873631D6601F5906518A7E49D6908367777CF749AE6F1C8AF). Upload set remains English and German PDFs; combined PDF is maintained locally. Changes in Version v10.4 (June 2026) Maintenance version after focused English and German style checks of the active v6 manuscript. No new sources and no new main claims were added. English style: tightened idiomatic wording in the taxonomy, policy, and neurodivergence sections; calibrated one remaining over-strong cognitive-debt formulation. German style: smoothed method-language hybrids and English calques, including terminology around cognitive endpoints, germane load, peer-reviewed studies, and the neurodivergence scaffold/workspace passage. Citations: post-style citation recheck confirmed EN/GER synchronization with 46 active bibliography items and no missing, unquoted, or duplicate keys. DE/EN: English and German PDFs remain synchronized. The upload set remains EN and GER only; the combined PDF is kept as a local control artifact and is not intended for Zenodo upload. Changes in this version (v10.3 source metadata, style, and evidence-ledger maintenance): Source metadata, style, and evidence-ledger maintenance — Verified the active bilingual source apparatus after the v10.2 disclosure hotfix. Corrected Bjork (1994) pages to 185–206; added the Oxford University Press DOI, ISBN and New York publication place for Clark (2008); aligned Gadamer to the cited 2004 Continuum second revised English edition while preserving the 1960 original in the note; added ISBNs for Kahneman (2011) and Willingham (2009); updated the McKinsey Superagency report URL to the current official path. The English manuscript was also calibrated for academic style and evidential caution, softening over-strong formulations around evidence strength, LLM scope, neurodivergent users, and the MIT-study critique. A new bilingual evidence ledger was added after the 2x2 taxonomy to link central inference claims to their strongest evidence anchors, current status, remaining gaps, and design implications. No new sources or new main claims were added. Compile: EN 18 pp., DE 20 pp., 0 errors. Previous changes (v10.2 disclosure hotfix): Disclosure hotfix — Hardened the AI/KI disclosure in both English and German to describe generative AI systems only as non-authorial tools for literature organization, drafting support, linguistic revision, consistency checking and LaTeX/PDF quality assurance. The disclosure now explicitly excludes AI authorship, co-authorship, acknowledgment credit, independent validation, truth authority and scientific/public-credit roles; sole responsibility remains with the human author. Two residual collaboration-wordings were neutralized as user-directed AI use. No substantive paper claims changed. Compile: EN 18 pp., DE 20 pp., 0 errors. Previous changes (v10.1 source-check): Source-check correction — Corrected Iqbal et al. (2025) from an inaccurate generic title/article number to the verified Scientific Reports article 15(1):16610, DOI 10.1038/s41598-025-01676-x, with full author metadata; added Gerlich (2025) correction notice DOI 10.3390/soc15090252; added arXiv/DataCite DOIs for Kosmyna et al., Choudhuri et al., Navneet et al. and Zhang & Reicherts; completed Lee et al. (CHI 2025) authors/pages and Mittler (HEAd'25) DOI; added DOI metadata for Bjork 1994, Slamecka & Graf 1978, Sweller 1988 and Wegner 1987. No substantive paper claims changed. Compile: EN 18 pp., DE 20 pp., 0 errors. Changes in this version (v10): Source-check correction — Updated Dell'Acqua et al. from Harvard Business School Working Paper to the peer-reviewed Organization Science article (37(2):403–423, DOI 10.1287/orsc.2025.21838); corrected the 94%/13%/4% AI-use figures from WEF attribution to the primary McKinsey workplace report; reclassified Lee et al. (CHI 2025) as mixed in the vote-counting table.LIT-1 (Mittler 2025) — Added longitudinal case study (Harnessing Generative AI to Overcome Executive Dysfunction in Higher Education, HEAd'25) as the only available longitudinal documentation of active AI integration in e