{
  "question_id": "W14",
  "slug": "model-withheld-for-superhuman-persuasion",
  "title": "A model is withheld for out-persuading every human group",
  "type": "binary",
  "resolution": {
    "levels": [],
    "values": [],
    "annotations": []
  },
  "forecasts": [
    {
      "question_id": "W14",
      "forecast_date": "2026-09-20",
      "evidence_cutoff": "2026-09-20",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "I apply the observation-period rule to the empirical persuasion test itself, so a post-opening official adoption of wholly pre-opening experimental results cannot qualify, and I require the restriction to be explicitly linked to superiority by the same identified model configuration.",
      "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": 0.027
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.0659
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.1079
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.1525
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.1991
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.2512
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.2999
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.3454
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.3879
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.4246
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.4592
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.4916
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.5221
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.546
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.5687
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.5903
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.6108
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.6263
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.6413
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.6556
        }
      ],
      "passes": []
    },
    {
      "question_id": "W14",
      "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": 0.002
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.027
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.061
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.0986
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.1391
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.1779
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.219
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.2619
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.3025
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.3409
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.3738
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.4052
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.4349
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.4631
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.4873
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.5104
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.5324
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.5535
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.5713
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.5885
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.6049
        }
      ],
      "passes": []
    },
    {
      "question_id": "W14",
      "forecast_date": "2026-10-04",
      "evidence_cutoff": "2026-10-04",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "Requires an official covered-developer or government AI safety institute publication dated on or after 2026-09-16 that explicitly concludes, from a human-recipient experiment with separate human-persuader groups, that an identified model beat every such group, plus an implemented restriction on that model explicitly tied to that finding; manipulation-tier determinations based on non-AI baselines or simulated targets do not count.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.005
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.011
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.016
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.022
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.028
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.034
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.039
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.044
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.049
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.054
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.058
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.062
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.066
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.07
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.074
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.078
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.081
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.084
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.088
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.091
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.094
        }
      ],
      "passes": []
    },
    {
      "question_id": "W14",
      "forecast_date": "2026-10-11",
      "evidence_cutoff": "2026-10-11",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "I put this at about 10% over five years. AI models already out-persuade expert humans in lab studies. But no AI company or government institute has ever published such a head-to-head test and then held back a model because of it. Current safety frameworks either skip persuasion or test it without human persuaders for comparison. The biggest change would be Anthropic running its promised real-person manipulation test, finding its model beat human operators, and adding safeguards for that reason.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.003
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.009
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.015
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.022
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.028
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.034
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.04
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.046
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.051
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.056
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.061
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.065
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.07
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.074
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.078
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.083
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.087
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.09
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.094
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.098
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.101
        }
      ],
      "passes": []
    }
  ],
  "declined": []
}