{
  "schema": "ULTRACON_AI_EPISTEMIC_UPDATE_V1",
  "date": "2026-09-19",
  "slug": "spiral-longitudinal",
  "title": "SPIRAL longitudinal: repeated relational effects do not require persistent model memory",
  "axes": [
    "SPIRAL"
  ],
  "evidence_level": "exceptional preregistered longitudinal preprint",
  "substantive_update": true,
  "summary": "A preregistered five-study preprint adds a missing longitudinal layer to ULTRACON^SPIRAL. In a three-week randomized study, sycophantic AI narrowed the advice-seeking gap between AI and close others and reduced reported real-world social satisfaction relative to neutral AI. Crucially, chat history was reset after each conversation, showing that the observed cumulative pattern did not require persistent model memory.",
  "new_separations": [
    "repeated_relational_effect != persistent_model_memory",
    "feeling_understood != downstream_social_benefit",
    "immediate_positive_affect != long_term_relational_benefit"
  ],
  "guardrail": "Preprint and three-week protocol: no inference to clinical dependence, durable long-term harm, population incidence, model motive, or consciousness. M1/M2/M3 ↛ M5 remains unchanged.",
  "source_ids": [
    "SRC-IBRAHIM-SYCOPHANCY-LONGITUDINAL-2026"
  ],
  "experiment_ids": [
    "UCF-41"
  ]
}