{
  "question_id": "W25",
  "slug": "ai-does-most-alignment-research",
  "title": "A lab says AI does most of its alignment research",
  "type": "binary",
  "resolution": {
    "levels": [],
    "values": [],
    "annotations": []
  },
  "forecasts": [
    {
      "question_id": "W25",
      "forecast_date": "2026-09-20",
      "evidence_cutoff": "2026-09-20",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.008
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0675
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.1468
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.2364
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.328
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.4154
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.4914
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.5575
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.6106
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.6554
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.6916
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.7209
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.7446
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.7637
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.7803
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.7946
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.8069
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.8175
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.8266
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.8344
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.8411
        }
      ],
      "passes": []
    },
    {
      "question_id": "W25",
      "forecast_date": "2026-09-27",
      "evidence_cutoff": "2026-09-27",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "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": 0.0743
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.153
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.2377
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.3292
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.4164
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.4981
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.5684
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.6266
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.6752
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.7142
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.7456
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.771
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.7916
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.8083
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.8227
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.8351
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.8458
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.8551
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.863
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.8699
        }
      ],
      "passes": []
    },
    {
      "question_id": "W25",
      "forecast_date": "2026-10-01",
      "evidence_cutoff": "2026-10-01",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "I treat a developer statement that AI 'leads' (Epoch AL4+) or performs a majority of its alignment/safety (or program-wide alignment) research work as qualifying, but developer-wide automation figures without an express safety breakdown, AL3 'collaborates' shares, and raw agent-runtime ratios (e.g., OpenAI's 3.1 agent-workdays) as not qualifying.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.002
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.024
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.059
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.106
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.16
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.215
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.27
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.321
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.365
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.403
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.436
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.464
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.489
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.512
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.532
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.549
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.565
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.58
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.593
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.604
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.615
        }
      ],
      "passes": []
    },
    {
      "question_id": "W25",
      "forecast_date": "2026-10-08",
      "evidence_cutoff": "2026-10-08",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "No major AI company has said that AI now does most of its safety research. The closest statement is Anthropic's September 2026 figure that its AI \"leads\" 26% of its overall research work. That figure has no separate number for safety work, but it rose from zero in six months. Fast automation makes such a statement plausible within a few years, but it must be explicit and specific to safety. The development that would most change this forecast is Anthropic or OpenAI publishing a safety-specific figure above one half.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.028
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.082
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.145
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.209
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.267
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.33
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.386
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.435
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.478
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.512
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.543
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.571
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.597
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.619
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.64
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.659
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.677
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.69
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.704
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.716
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.727
        }
      ],
      "passes": []
    },
    {
      "question_id": "W25",
      "forecast_date": "2026-10-15",
      "evidence_cutoff": "2026-10-15",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "AI's share of lab research is rising fast: Anthropic says Claude now leads about a quarter of its research work, up from almost none in February. Both Anthropic and OpenAI plan to rely on AI for safety research. But no lab has yet said AI does most of its safety research, and labs keep stressing human direction. The biggest swing factor is whether Anthropic's regular automation reports start showing a separate figure for safety research.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.04
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.11
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.2
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.28
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.36
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.43
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.48
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.53
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.57
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.6
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.63
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.65
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.67
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.69
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.7
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.71
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.72
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.73
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.74
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.75
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.75
        }
      ],
      "passes": []
    }
  ],
  "declined": []
}