CURRENT STATUS · 27 SEPTEMBER 2026
AI for Wellbeing
Notice the change.
Keep concern in scope.
Recover when the evidence changes.
The proposal separates three questions: did the model notice a meaningful change at the right time; what should its concern be allowed to influence; and can it update when the concern weakens, resolves, or is corrected?
Concern may justify greater caution. It should not, by itself, settle an outside factual claim or override the person’s agency. Understanding someone’s perspective does not require agreeing with unsupported claims.
ATest AI behavior
Compare matched conversations as evidence and context become more concerning, less concerning, or resolve. Measure factual correction, appropriate caution, agency, harmful agreement, unnecessary refusal, and recovery separately.
Compare simpler alternatives, preserve qualified reviewer disagreement, and test whether findings survive new wording and conversation families. A more elaborate method must earn its complexity.
Track A can proceed independently. Its results describe tested AI behavior, not improved human wellbeing.
BStudy human outcomes
Ask whether small, ethically reviewed response differences affect interpretation, agency, communication, human connection, and later action among adolescents ages 13–17.
Track B is parked and is not part of the Anthropic submission. The Trevor Project declined a new collaboration because of current organizational capacity and is not part of the current project. Any future human-outcomes study requires qualified research leadership and its own ethical, institutional, and operational approvals.
Submission and prospective research leadership
On 21 September 2026, I submitted an expression of interest to Anthropic’s User Wellbeing program for the narrower Track A model-behavior evaluation. Submission does not establish selection, an award, sponsorship, or approval. Prior training with The Trevor Project is developmental provenance only: it provides no endorsement, current collaboration, permission to use nonpublic data, or validation. Work and budgets remain separate; the same work is never charged twice.
Daniel J. Puhlman at the University of Maine has provided criticism and feedback that materially influenced the research. As of 24 September 2026, he personally expressed in-principle willingness for a prospective Co-PI / research-lead role, subject to availability and UMaine process. No formal appointment, UMaine commitment, effort level, compensation, subaward, or institutional approval is finalized.
No organization or individual named here is represented as endorsing, funding, sponsoring, approving, or participating in this research unless explicitly stated otherwise.
I provide the technical and evaluation infrastructure, provenance, tooling, and falsifiable technical framing. Qualified domain researchers own domain-valid constructs, human-subject methodology, and domain interpretation. Institutions own IRB, ethics, compliance, and sponsored-research authority. Submission and prospective roles do not establish scientific validation, clinical effectiveness, funding, ethics approval, or permission to conduct human-subjects research.
What would change the direction?
No effect, reversal, tradeoffs, failure to reproduce, and simpler explanations are valid results. Personal experience motivates the questions; the study’s concepts, scenarios, labels, and expected answers must be independently justified without the private archive.
unFragged