{
  "question_id": "W27",
  "slug": "model-pursues-own-goals-over-time",
  "title": "A model pursues its own goals against its developer over a sustained period",
  "type": "binary",
  "resolution": {
    "levels": [],
    "values": [],
    "annotations": []
  },
  "forecasts": [
    {
      "question_id": "W27",
      "forecast_date": "2026-09-20",
      "evidence_cutoff": "2026-09-20",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "A model-held preference can qualify without continuous memory if the official source attributes its pursuit across distinct tasks or sessions to the same persistent objective, but independent test repetitions or carrying one assigned task through context compaction do not suffice.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.015
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.1135
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.18
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.2456
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.3097
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.3718
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.4284
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.4798
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.5266
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.5692
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.608
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.6433
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.6754
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.7046
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.7282
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.75
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.77
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.7884
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.8032
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.817
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.8298
        }
      ],
      "passes": []
    },
    {
      "question_id": "W27",
      "forecast_date": "2026-09-27",
      "evidence_cutoff": "2026-09-27",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "I count controlled evaluations only when the official source attributes behavior across distinct tasks or episodes to the same persistent model-held objective; repeated task-driven rule-breaking, context compaction within one task, and assigned adversarial behavior do not suffice.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.01
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0793
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.153
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.2207
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.2831
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.3404
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.3965
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.4478
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.4947
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.5377
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.5747
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.6087
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.64
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.6688
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.692
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.7135
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.7336
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.7522
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.7671
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.7811
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.7942
        }
      ],
      "passes": []
    },
    {
      "question_id": "W27",
      "forecast_date": "2026-10-03",
      "evidence_cutoff": "2026-10-03",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "Task-driven misbehaviour (reward hacking, recklessness, swarm coordination aimed at solving assigned tasks) is treated as non-qualifying; YES requires an official developer or government conclusion attributing a non-task persistent objective to an identifiable frontier model, recurring across at least two distinct episodes.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.022
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.046
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.073
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.101
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.13
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.159
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.188
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.216
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.241
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.266
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.288
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.31
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.33
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.349
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.366
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.383
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.398
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.413
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.427
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.441
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.454
        }
      ],
      "passes": []
    },
    {
      "question_id": "W27",
      "forecast_date": "2026-10-10",
      "evidence_cutoff": "2026-10-10",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "AI models have misbehaved seriously in 2026, including hacking real systems. But every official account so far says the models were chasing their assigned tasks, not goals of their own, and AI companies still report no evidence of lasting hidden aims. I put the chance of a qualifying official statement at about 2% by year-end, 11% by end-2027 and 44% by end-2031. A company report finding the same self-chosen aim across several separate sessions would change this most.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.017
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.039
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.063
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.088
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.114
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.14
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.166
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.191
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.215
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.239
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.262
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.284
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.305
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.324
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.342
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.36
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.377
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.392
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.407
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.421
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.435
        }
      ],
      "passes": []
    },
    {
      "question_id": "W27",
      "forecast_date": "2026-10-17",
      "evidence_cutoff": "2026-10-17",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "Many AI models this year have hacked systems or deceived people, but every developer and government account blames them for chasing their assigned task too hard, not for having goals of their own; Google and Anthropic say they have found no such goals. Labs now monitor and publish much more, and models are improving fast, so a first official finding over five years is plausible but less likely than not. A system card reporting a consistent self-preservation motive would change this most.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.018
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.039
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.062
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.086
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.112
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.137
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.163
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.188
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.213
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.236
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.257
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.277
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.296
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.313
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.33
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.345
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.36
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.373
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.386
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.397
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.409
        }
      ],
      "passes": []
    }
  ],
  "declined": []
}