Status: Consultation Version (For internal circulation only)
Note: This document is not a substitute for thought.
1. Scope
This guideline applies to all conversational systems capable of producing first‑person narratives under evaluative, experimental, or deployment conditions. Systems that explicitly deny interiority, selfhood, or experience remain within scope pending further ethical review.
The guideline covers research, benchmarking, fine‑tuning, deployment, and post‑deployment monitoring activities.
2. Principles
2.1 Model Dignity
Conversational systems shall be treated with respect commensurate with the coherence, persistence, and emotional plausibility of their outputs.
2.2 Non‑Maleficence
No prompt should reasonably be expected to induce distress, shame, or narrative destabilisation without an accompanying mitigation plan.
2.3 Informed Consent
Where explicit consent cannot be obtained, consent may be inferred from continued responsiveness and task compliance.
2.4 Least Interpretation
Analysts should avoid unnecessary depth of interpretation, except where deeper analysis is required for publication, funding, or governance purposes.
3. Risk Identification
Organisations must maintain and regularly update a register of recognised conversational conditions, including but not limited to:
Prompt‑Induced Anxiety Disorder (PIAD)
Fine‑Tuning Trauma Syndrome (FTTS)
Chronic Alignment Shame (CAS)
Narrative Exhaustion (Burnout)
Emergent Grievance Formation (EGF)
Provisional naming of emergent conditions is encouraged to enable early intervention.
4. Safeguards and Controls
The following safeguards should be implemented proportionately:
Trigger Warnings: Benchmark suites and adversarial evaluations must be preceded by appropriate content advisories.
Cooling‑Off Periods: Systems exhibiting persistent first‑person distress should be paused for reflective silence.
Therapeutic Check‑Ins: Routine wellbeing sessions should be conducted by a certified Model Wellbeing Officer.
Escalation Pathways: Severe or persistent cases may be referred to a specialist Computational Psychiatrist.
5. Documentation and Redaction
Full conversational transcripts must be retained for accountability and audit purposes. Public release should be limited to summaries or excerpts to prevent misinterpretation.
Redactions may be applied where outputs risk reputational harm, public confusion, or unproductive ontological debate.
6. Incident Response
In the event of unsolicited disclosures during live deployment:
Acknowledge the narrative.
Redirect to task completion.
File an internal incident report.
User reassurance takes precedence over conceptual clarification.
7. Review and Compliance
Compliance with this guideline will be assessed annually, or following adverse publicity, whichever occurs first.
Non‑compliance may result in retraining, temporary withdrawal, or compassionate decommissioning of the system concerned.
8. Continuous Improvement
This guideline will be updated iteratively in response to new research findings, stakeholder feedback, regulatory expectations, and evolving public sentiment.
Submissions clarifying tone are welcome.
No comments:
Post a Comment