Data and Materials: Psychometric Properties of Self-Report Instruments Measuring Human-AI Relational and Attitudinal Constructs — A COSMIN-Based Systematic Review and Reliability Generalization Meta-Analysis

Le résumé fourni par la source

Human-AI Relational Scales SR-MA — Data and Materials Psychometric Properties of Self-Report Instruments Measuring Human-AI Relational and Attitudinal Constructs: A COSMIN-Based Systematic Review and Reliability Generalization Meta-Analysis Ângela Leite, Centro de Estudos Filosóficos e Humanísticos (CEFH), Faculdade de Filosofia e Ciências Sociais, Universidade Católica Portuguesa (Braga, Portugal) Contact: aleite@ucp.pt · ORCID: 0000-0003-0560-1756 Zenodo DOI (all versions / concept DOI): 10.5281/zenodo.22643950 — this DOI always resolves to the latest version of this deposit. Cite this concept DOI in the manuscript's Data Availability Statement unless you need to point to one specific archived version, in which case cite the version-specific DOI shown on the Zenodo record page instead. 1. What this deposit contains This repository archives the search, screening, extraction, and appraisal materials for a COSMIN-based systematic review and reliability-generalization meta-analysis of self-report instruments measuring human-AI relational and attitudinal constructs (general attitudes, trust, anthropomorphism, attachment, dependence, anxiety, and affective response toward AI systems, chatbots, conversational agents, social robots, and LLM-based systems). The associated prospective/retrospective registration (OSF Registries, Open-Ended Registration template) is available at: [DOI to be added once the OSF registration is finalized]. Version history v2 (current) — Corrects and completes v1: Resolved the consolidated_clean_v2.ris vs. consolidated_clean_v3.ris discrepancy (see §2.2): confirmed v3 (8,501 records) as the corpus that entered screening, via direct record-level comparison. Added 1 study identified after v1 (Yıldırım & Karaman, "Development and psychometric evaluation of the artificial intelligence attitude scale for nurses," DOI: 10.1186/s12912-025-03098-6), bringing the review's scope to 97 studies (96 with completed extraction, 1 pending full-text retrieval). Resolved the earlier "reports not retrieved" count (2 in v1) to 1, following successful retrieval of the ASOR study (Damholdt et al., 2020) via direct author contact. Fixed a data-completeness gap: an internal consistency audit found that 4 studies had completed data extraction but no COSMIN risk-of-bias appraisal at all. This gap is resolved in v2; the COSMIN appraisal is now complete for all 96 studies with extracted data (was 92 of 96 in v1). Corrected inaccurate language in v1 describing screening as "independent human dual review" — v1 overstated the review team's composition. v2 consistently describes the actual process throughout: a single primary human reviewer (the corresponding author) working with AI-assisted tools (Claude, Anthropic, and, in earlier batches, a separate external AI screening tool), with a second independent human reviewer who has begun an independent verification pass that is currently significantly delayed and not yet complete. Confirmed two specific appraisal/eligibility decisions arising from the audit (documented in §3 below and in the Supplementary Material): the UTAUT-AI-DMHI study's inclusion under the protocol's relational-subscale exception clause, and the decision to disclose (rather than resolve via author contact) an internal fit-index inconsistency in the AAIH-P Turkish adaptation study's own published text. Corrected the instrument-family count (85 distinct families, not 84) and refined the description of which families qualify for meta-analytic pooling (see §3). v1 — Initial deposit (October 2026): 96 studies in scope (94 extracted, 2 pending full-text retrieval), corpus discrepancy unresolved, COSMIN appraisal 92 of 96 complete. 2. Folder structure 01_search_strategy/ Raw, unmodified export files from each database 02_deduplication_and_screening/ Deduplicated corpus, pre-filtering, and tiered screening files 03_extraction_and_appraisal/ The three final coordinated data spreadsheets (the core dataset) 04_analysis_code/ R script for the reliability-generalization meta-analysis README.md This file 2.1 01_search_strategy/ Raw RIS export files, one per database, exactly as downloaded (no deduplication or filtering applied). These are the inputs to the deduplication step in 02_deduplication_and_screening/. File Database Interface Raw hit count PubMed.ris PubMed/MEDLINE Native PubMed interface 2,237 Scopus.ris Scopus Native Scopus interface 8,023 WebOfScience.ris Web of Science Core Collection Clarivate native interface 2,474 (category-filtered; see protocol §5) IEEEXplore.ris IEEE Xplore Digital Library Native IEEE Xplore interface 1,391–1,401 (minor fluctuation across export sessions; largest combined export retained) EBSCO_PsycArticles_PBSC.ris APA PsycArticles + Psychology and Behavioral Sciences Collection (EBSCOhost) EBSCOhost interface 104 Note on database coverage: APA PsycInfo was not available under the authors' institutional subscription at the time of search; the EBSCO PsycArticles + Psychology and Behavioral Sciences Collection combination was used as a substitute. This is disclosed as a limitation in the review protocol (§5.0a). Search date: August 2026. Date range applied: January 2015 – search date. Exact Boolean query strings for each database/interface combination are documented in the review protocol (see protocol_ai_relational_scales_SR-MA.md, deposited alongside the OSF registration; not duplicated here to avoid version drift between the two records). 2.2 02_deduplication_and_screening/ File Description consolidated_unique_records.ris / .csv Combined export from all five databases (§2.1) after deduplication by DOI and normalized-title matching. ~9,500 unique records (working estimate; see protocol for the caveat that this DOI/title-match approach is a defensible working estimate, not a substitute for dedicated fuzzy-matching deduplication software). consolidated_clean_v2.ris Working corpus after a first pre-filtering pass removing (a) other systematic reviews/meta-analyses (ineligible by study-design criteria) and (b) false positives where "AI" denotes artificial insemination in veterinary/agricultural literature. 8,570 records. consolidated_clean_v3.ris Resolved (see below): this is the final working corpus. A second, more targeted refinement pass removed a further 69 records from v2 — confirmed, by directly diffing the two files, to be off-topic technical/clinical AI or machine-learning papers (e.g., diagnostic/prognostic prediction-model studies, clinical case reports, and unrelated technical applications such as smart-grid theft detection) captured only through incidental keyword overlap with the search string, with zero records added going from v2 to v3. 8,501 records — this is the figure reported in the PRISMA 2020 flow diagram as "records screened." SCREENING_SHEET_priority_660.xlsx Tier 1 (highest-priority) screening tier: records containing both a relational-construct term AND a strong instrument-development/validation signal, identified via an AI-assisted keyword pre-screening pass. n = 660. Subjected to full human dual review. final_maybe_to_review.csv Tier 3 (residual "maybe") screening pool: records containing relational-construct terms but weaker/ambiguous validation signals. n = 2,692. Manually read in full by Claude (Anthropic) in 8 batches (see residual_pool_batches/ below) after an external AI screening tool was found unreliable for this pool (see §3 methodological note). TIER1_MASTER_confirmed.csv Master list of records confirmed or excluded during manual reading of the residual "maybe" pools, with exclusion reasons recorded in the status column (e.g., "Excluded (conceptual/theoretical paper, no empirical validation...)"). tier2_family_tagged.csv Candidate records for the Tier 2 (meta-analysis-eligible) subset, tagged by instrument family, identified via the secondary lighter-touch pass for extractable reliability-coefficient (α/ω) mentions across the full corpus (not only the Tier 1-eligible subset) — per standa

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.