SLIP & ETHICS: A Graduated Safety Protocol for AI Emotional Companions
To tackle the safety-rapport paradox in AI emotional companions, researchers have developed SLIP (Staged Layers of Intervention Protocol) and ETHICS (Emergent Taxonomy for Human-AI Interaction Context Signals). SLIP consists of a four-stage graduated approach that identifies interventions—none, soft, or hard—based on structured qualitative indicators: affect intensity (a) and narrative dynamism (m). Meanwhile, ETHICS offers a taxonomy focused on "signals not labels" for context signals. An evaluation that included a small-scale deployment (N=68 entries, 10 users, 10 weeks) alongside a synthetic persona battery (N=91, 5 behavioral-risk profiles) reported 0% false positives for the flow persona. Nonetheless, findings indicated that 8 consecutive days of heightened energy resulted in no interventions (0/8), highlighting a limitation of the "do not pathologize" principle. The study appears on arXiv with the identifier 2605.15915.
Key facts
- SLIP is a four-stage graduated methodology for AI emotional companion interventions.
- ETHICS is a taxonomy using 'signals not labels' for human-AI interaction context.
- Interventions are classified as none, soft, or hard based on affect intensity (a) and narrative dynamism (m).
- Evaluation included a production deployment with 68 entries, 10 users, over 10 weeks.
- A synthetic persona battery involved 91 entries across 5 behavioral-risk profiles.
- The flow persona achieved 0% false positives.
- Crisis-oriented personas showed expected escalation patterns.
- 8 consecutive days of high-energy elevation resulted in zero interventions (0/8).
Entities
Institutions
- arXiv