Deep Collaboration Methodology 2.0

This study proposes and executes the Deep Collaboration Methodology 2.0 (DCM 2.0), taking responsibility attribution perception in contexts of AI use as its research object. Street-intercept interviews were conducted at Da'an Forest Park in Taipei and at several university campuses, producing 328 valid interviews spanning 33 nationalities (including Taiwan, with 32 foreign nationalities). Field execution ran from December 22, 2025 to April 15, 2026. The study employs a Convergent Mixed-Methods Design, using two fixed core questions as quantitative anchors: Q1 asks "when AI output is erroneous, to whom does responsibility belong?" (the level of moral attribution), and Q2 asks "when a loss results from acting on an AI-generated answer, who bears the loss responsibility?" (the level of practical allocation). Ten rotating question sets (Q3/Q4) were used to explore respondent reaction patterns across different AI social issue contexts. The principal findings are organized across three levels. First, the user internalization of loss responsibility in Q2 exhibits overwhelming cross-group robustness (79% overall), remaining stable across stratifications by gender, age, nationality, and tool choice for most background characteristics — indicating that "when a loss results from acting on an AI-generated answer, the user should bear responsibility" constitutes a cross-group ethical intuition about tool use. Second, the Q1 responsibility framework is systematically moderated by usage experience, forming a continuous gradient from non-AI users (AI attribution 52%) to short-term users (33%) to long-term users (24%), with the core mechanism being the accumulation of error-encounter experience. Third, a structural epistemic gap exists between Q1 and Q2: Q1's tripartite dispersion (user attribution 25%, shared responsibility 37%, AI attribution 31%) collapses abruptly upon entering Q2's loss responsibility framework — even among respondents who attributed Q1 responsibility to AI, 71% still pointed to the user in Q2 — reflecting the systematic fusion that occurs when two distinct ethical questions ("who caused the error" versus "who should bear the consequences") are presented in sequential order. At the methodological level, DCM 2.0 fills structural gaps in existing AI cognition research across three dimensions — ecological validity, cross-cultural coverage, and sample diversity — providing a traceable, replicable, and institutionally unaffiliated street-intercept mixed-methods protocol. The construction of DCM as an independent methodological framework is presented as a core academic contribution alongside the substantive findings of this study. Note: The PDF in this deposit references an earlier DOI(10.5281/zenodo.20280701). Both records refer to the same work.Please cite this record: https://doi.org/10.5281/zenodo.20281335 Correction (May 20, 2026): Author name corrected in Section 2.2 and reference list — Makoana et al. → Kakraba et al. (2026). No changes to research findings or methodology. Version 3 Correction and Revision Note (2026/06/25) Corrections: 1.Journal name: Branda and Ciccozzi (2026) was incorrectly listed as published in eBioMedicine. The correct journal is Artificial Intelligence in the Life Sciences. DOI updated to 10.1016/j.ailsci.2026.100158. Revisions: 2. Section 1.4: Inline meta-commentary and parenthetical glosses have been removed. Substantive findings and analytical content are unchanged. 3. Section 3.2.1: A paragraph discussing the structural sampling advantages of street-intercept design in the context of rising digital tool adoption has been removed. Core design logic is unchanged. 4. Section 3.2.2: Paragraphs explaining the rationale for withholding question sequence information and the open-ended design of Q1 have been removed. Interview structure and question wording are unchanged. 5. Section 8.7: Fully rewritten. The revised version reframes the researcher's epistemological position in methodological rather than personal terms, and restructures the section into three analytical components: outlier status, structural institutional exclusion, and execution-level operational logic. The substantive findings, analytical conclusions, and all other content remain unchanged.