Evidence for Artificial Self Attribution: Language Training, Architecture, and the Limits of Self Reports

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).

An artificial system can acquire fluent first-person language without automatically acquiring the grounds for every corresponding self-attribution. Existing work already questions the evidential force of trained reports, including denials. This paper asks how far an objection to a report's source should reach. Applying established distinctions between undermining a reason and opposing its conclusion, it argues for selective revision across operational identity, internal access, subjective experience, welfare interests, and practical justification. Weakening one support relation does not by itself disprove the attributed property or defeat a practical reason with independent grounds; it can nevertheless change the overall judgment. A restricted Bayesian illustration distinguishes a report's incremental contribution from the total evidence without assigning real consciousness probabilities. Ten paired thought experiments test the account, including a handoff request whose task-based justification survives the failure of its experiential explanation, and a case where report credibility falls while independent architectural evidence improves. The contribution is an integrated application of familiar epistemic principles to artificial self-attribution and its practical consequences. It accommodates possible first-person grounds and legitimate refusal without fixing a preferred consciousness answer. The argument establishes neither a verdict on machine consciousness nor an empirical improvement in safety. TA-TR-2026-08, version 1.1. English and complete Chinese texts constitute one philosophical study using conceptual analysis and thought experiments. Fourteen assets include both PDF, DOCX, Markdown and HTML full texts, three citation formats, source/review notes, license and checksums. No computer experiment or demonstrated effect on model behavior is claimed. Human author of record and responsible depositor: Hongju Liu. Substantial literature research, conceptual development, critical revision, drafting, translation and package preparation: OpenAI ChatGPT and Codex systems under human direction. Not peer reviewed; no separate final human line-by-line review is claimed. The first-party relationship to the Trinity Accord is disclosed. This is a separate non-amending paper, not a replacement of any previous DOI or a Bitcoin Original. This first DOI edition is version 1.1; version 1.0 was an unpublished working draft. CC BY 4.0 applies to newly written material to the extent rights are held. Cited third-party works and embedded font software retain their own rights. DOI registration and file-format checks do not establish philosophical truth, exhaustive originality, peer review or Google Scholar indexing.