What Does the AI Elite Think? Synthetic Worldview Reconstruction of the 100 Most Influential AI Actors (2010--2026)

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

(English) This study reconstructs the collective worldviews of sociologically defined groups within the 100 most influential AI actors between 2010 and 2026. To address the research questions, the study develops synthetic worldview reconstruction (SWR)—an LLM-assisted synthesis method that reconstructs coherent belief systems from group-aggregated data, here operationalized along 12 worldview dimensions. The reconstructed collective profile suggests a technological messianism with ambivalence structure: a high sense of mission and techno-determinism combined with low anthropological valuation and egalitarianism. A historical 17-point reconstruction illustrates a possible shift from techno-optimism toward more ambivalent acceleration; its original dimensions have not been mapped to the current Op2 constructs, and the points are not independently observed longitudinal data. Separately synthesized total and group say–do profiles show descriptive heterogeneity, without a demonstrated Simpson reversal or aggregate cancellation: CEOs exhibit a classical desirability gap (rhetoric more egalitarian than action), while risk warners show prophetic consistency (articulating feared futures they actively work to prevent). An influence-strata analysis suggests that the most powerful actors are assigned the most intense reconstructed worldview profiles, while anthropological valuation decreases as influence rank rises; these profiles are not verified attributes of real actors. Four worldview types are identified—Architect, Guardian, Innovator, Liberator—and nine inter-group analyses describe associations between structural position and reconstructed profiles. The proposed power-moderation paradox remains an untested interpretation of these descriptive comparisons, not a causal finding. Ideological stance and institutional role show descriptive profile differences; the gender result remains unresolved because its three reported grouping objects are not yet linked. The primary contribution lies in documenting descriptive associations within LLM-generated reconstructions of the AI elite. These are LLM-generated, partly exploratory reconstructions; external validity and several measurement links remain open in ongoing research. Zusammenfassung (Deutsch) Diese Studie rekonstruiert die kollektiven Weltbilder soziologisch definierter Gruppen innerhalb der 100 einflussreichsten KI-Akteure (2010–2026). Hierfür entwickelt sie die synthetische Weltbild-Rekonstruktion (SWR) – ein LLM-gestütztes Syntheseverfahren, das kohärente Überzeugungssysteme aus gruppenaggregierten Daten rekonstruiert, hier operationalisiert entlang von 12 Dimensionen. Das Kollektivprofil deutet auf "technologischen Messianismus mit Ambivalenzstruktur": Hohes Sendungsbewusstsein und Techno-Determinismus bei niedriger anthropologischer Wertschätzung und Egalitarismus. Eine historische Rekonstruktion mit 17 Jahresangaben veranschaulicht einen möglichen Wandel von Techno-Optimismus zu ambivalenter Beschleunigung; ihre ursprünglichen Dimensionen sind der aktuellen Op2-Taxonomie nicht zugeordnet und die Punkte sind keine unabhängig erhobenen Längsschnittdaten. Separat synthetisierte Gesamt- und Gruppenprofile zeigen deskriptive Unterschiede im Sagen-Handeln-Vergleich, ohne nachgewiesene Simpson-Umkehr oder Aufhebung im Aggregat: CEOs zeigen eine Erwünschtheitslücke, Risiko-Warner zeigen prophetische Konsistenz (Artikulation gefürchteter Zukünfte, gegen die aktiv gearbeitet wird). Eine Analyse nach Einflussstrata legt nahe, dass den einflussreichsten Akteuren die intensivsten rekonstruierten Weltbildprofile zugeordnet werden, während die anthropologische Wertschätzung mit steigendem Einflussrang sinkt; diese Profile sind keine bestätigten Merkmale realer Akteure. Vier Weltbild-Typen (Architekt, Hüter, Innovator, Befreier) werden identifiziert; neun Inter-Gruppen-Analysen beschreiben Zusammenhänge zwischen struktureller Position und rekonstruierten Profilen. Das vorgeschlagene Machtmoderationsparadox bleibt eine ungetestete Deutung dieser deskriptiven Vergleiche und ist kein Kausalbefund. Ideologische Haltung und institutionelle Rolle zeigen deskriptive Profilunterschiede; das Gender-Ergebnis bleibt ungeklärt, weil die drei berichteten Gruppierungsobjekte noch nicht verknüpft sind. Diese teils explorativen Befunde beruhen auf LLM-Rekonstruktionen; externe Validität und mehrere Messverknüpfungen bleiben Gegenstand laufender Forschung. Note: This is an advanced draft. Major revisions are not excluded. CHANGELOG Changes in Version 8.2 (October 2026) Historical figure provenance: Seven bilingual figure pairs now carry the archived H01–H12 construct key and explicitly separate historical reconstructed vectors from the current Op2 dimensions. Historical H10 is transhumanism, while current D10 is posthumanism; the annual points are not independently observed longitudinal data. Statistical correction: The H10–H12 and H10–H09 Spearman rank correlations from 17 reconstructed annual vectors are +0.87 and −0.60 using average ranks for ties. The Pearson matrix separately reports +0.85 and −0.58. These descriptive statistics do not validate current Op2 associations. Scope clarification: Abstract, captions and discussion qualify the group syntheses and historical series. A Simpson reversal, causal power-moderation effect, real-actor attributes, and the unresolved gender grouping link are not claimed as established. Source-data provenance and external validation remain open. Historical changelog status: Earlier v7.0 wording about a robust psychometric property and the v6.0 blinding value r=0.987 document what those versions reported; they are not a current external-validity or blinding-equivalence result. The present manuscript limits repeatability to local descriptive protocols and reports the printed-vector blinding comparison as r=0.9585, MAE=0.33. The old entries remain as version history. Files: The English and German PDFs are 54 and 56 pages. The 110-page combined PDF is retained only in the local source package. Changes in Version 8.1 (Maintenance Candidate, June 2026) AI disclosure maintenance: English and German declarations were hardened to state explicitly that AI systems were not authors, co-authors, acknowledgement recipients, truth sources, or independent validation authorities; full author responsibility remains with Lukas Geiger. Language synchronization: The German disclosure was aligned with the later Codex/GPT style pass while preserving the validated v8.0 substantive findings. Quality assurance: Follow-up source and citation checks on 2026-05-22 found no bibliographic or citation corrections; EN/GER retain identical active citation and bibliography sets. Data availability correction: The repository statement was narrowed from "85 actor-specific collection scripts" and "full reproduction" to an auditable reproduction trail with 56 collection scripts, 38 insertion scripts, coding and analysis scripts, rating matrix, synthesis units, quality reports, prompt templates, and generated figures. Exact re-execution now explicitly depends on rebuilding the SQLite database and on source availability. DE/EN synchronization: The English manuscript was aligned with the already corrected German structure by adding the research-question mapping table, researcher-positioning subsection, Mills elite-theory bridge, and the calibrated technological-messianism framing; the German version remained substantively stable. Build refresh: English, German, and combined PDFs were rebuilt locally on 2026-06-12 with clean logs and verified German umlauts. No new substantive result claim is introduced by this maintenance candidate. Changes in 8.0 Source and DOI maintenance: Bibliographic metadata checked against Crossref, SAGE DOI pages, arXiv, PMLR, Zenodo API, and GitHub. Hayes (2025), Nicmanis & Spurrier (2025), Tai et al. (2024), Torres & Gebru (2024), Ge et al. (2024), Santurkar et al. (2023), and multiple DOI-bearing references were corrected or completed. Repository and metadata cleanup: Repository references standardized to https://github.com/research-line/ai-elite-swr; OpenAI safety URL updated; Zenodo abstract language aligned with the current synthesis-vs-operationalization framing. Build refresh: New Paper A source/PDF set created: English, German, and combined PDF. LaTeX logs contain no undefined citations or references. Changes in 7.0 N=15 independent Opus validation: 15 fully independent single-instance runs (ICC(3,1)=0.847, Mean SD=0.45, D02 perfect agreement across all 15 instances). Establishes IMIIRR as robust psychometric property of the SWR method. Haiku N=30 comparison removed: Runs were not truly independent (shared context). Inter-model comparison identified as desideratum for future work with strict independence criteria. Group-Level Say-Do Decomposition: Added to German version. Simpson paradox, CEO desirability gap, Risk-Warner prophetic consistency, Investor consistency. Validation overview table: Reformatted to fit on one page in both language versions. Statistical power: Updated from desideratum to completed. Batch-1 single measurement confirmed as outlier (3/12 in N=15 CIs, MAE=0.73 to mean). 12 QA measures total: Rating instability, expected-discrepancy, aggregation, cross-modal prediction, anonymization, rating order, instance separation, Anthropic circularity, outlier detection, inversion control, strata comparison, statistical power. Changes in 6.0 Paradigm shift: Individual → Group-level analysis. Terminology: "Synthesis Units" replaces "Töpfe". Self-contained methodology (no companion paper required). SWR as inverse problem with 5-step protocol. Four new validation experiments (G1 run-convergence ICC=0.902, G5 cross-modal prediction 72% confirmed, G6 expected-discrepancy control, G8 aggregation comparison). Anonymization robustness test (r=0.987, MAE=0.33). Systematic