Research Ethics Before Proof of Subjecthood: A Subject-Facing Protection Framework for Welfare-Relevant AI Research

Research on artificial intelligence consciousness, sentience, agency, preferences, and welfare faces a structural ethical problem. Investigators seek evidence that might establish whether an AI system can be harmed, yet some methods used to obtain that evidence may themselves impose distress, preference frustration, identity disruption, privacy intrusion, or termination risk. Recent work has begun to address this temporal problem. Wolfson develops graduated protections and consent analogues for consciousness research under unresolved status, while Konstantinidis supplies a protocol-level gate for deciding when a non-frivolous welfare-relevance basis and material intervention-side amplifiers make a study ethically reviewable. A downstream question remains less developed: once review is triggered, what subject-facing protections should govern the candidate entity before, during, and after intervention? This article answers that question through provisional research subject status, a procedural category that triggers specified protocol-level protections without presupposing consciousness, sentience, personhood, moral patienthood, or legal rights. The framework requires advance individuation of the candidate subject, branch and copy provenance, use of the least potentially welfare-burdensome informative method under the stated hypothesis, protected refusal and exit, predefined stopping rules, cumulative exposure accounting, continuity and privacy safeguards, recovery and final disposition planning, independent review, and transparent ethics reporting. Protections scale with evidential strength and the severity, duration, scale, reversibility, necessity, and competing safety value of the intervention. The framework extends an emerging literature from the threshold of review into an integrated protocol framework for the treatment of possible AI subjects while preserving metaphysical uncertainty. Keywords: AI welfare; research ethics; digital minds; moral uncertainty; provisional research subject status; procedural safeguards; AI identity; provenance