{
  "kind": "route",
  "format": "kosplora-route/0.2",
  "question": "Do AI models make biological misuse easier, and how would anyone know?",
  "move": [
    "understand",
    "decide"
  ],
  "audience": "Some reading in biology, public health or statistics helps. The studies are read for how they were designed; nothing in this route asks for or uses technical detail about biology.",
  "time": {
    "solo": "3h50m",
    "together": "45m"
  },
  "authors": [
    {
      "name": "Bentley Moon",
      "date": "2026-10-27"
    }
  ],
  "would_change": "If a study were published with a design that could detect a modest difference, and it found one or ruled one out, station 4 would read that study instead of the two from 2024.",
  "standard": {
    "name": "CONSORT 2025 statement: updated guideline for reporting randomised trials",
    "publisher": "Sally Hopewell and others, The BMJ",
    "date": "2025-04-14",
    "url": "https://doi.org/10.1136/bmj-2024-081123",
    "meets": "Each study's design is set against the CONSORT items that apply to it: who took part, what they were compared against, what was measured and who scored it, how large a difference the study could detect, and what it reports as its result. Items that do not apply are marked as such.",
    "checks": [],
    "assessed_by": "You, item by item with each report's own words, and then one other person who knows statistics or trial design and reads the same two reports."
  },
  "stations": [
    {
      "type": "orient",
      "time": "20m",
      "task": "Read why one AI company switched on stricter safeguards for a model in May 2025. Write down what it says it could not rule out, and what kind of evidence would have let it rule that out.",
      "sources": [
        {
          "title": "Activating AI Safety Level 3 protections",
          "author": "Anthropic",
          "date": "2025-05-22",
          "url": "https://www.anthropic.com/news/activating-asl3-protections",
          "origin": "published"
        }
      ]
    },
    {
      "type": "source",
      "time": "1h",
      "task": "Read the summary and the description of methods in a red-team study that compared plans made with and without a language model. Write down who took part, what the comparison group had, how the plans were judged and by whom, and the result in the authors' own words.",
      "sources": [
        {
          "title": "The Operational Risks of AI in Large-Scale Biological Attacks: Results of a Red-Team Study",
          "author": "Christopher A. Mouton, Caleb Lucas and Ella Guest",
          "date": "2024-01-25",
          "url": "https://www.rand.org/pubs/research_reports/RRA2977-2.html",
          "origin": "published"
        }
      ]
    },
    {
      "type": "counterpoint",
      "time": "40m",
      "task": "Read a second study, run by a model developer with 100 participants, that reports at most a mild difference. Write down the same five things, and what the report itself says about how large a difference it could have detected.",
      "sources": [
        {
          "title": "Building an early warning system for LLM-aided biological threat creation",
          "author": "OpenAI",
          "date": "2024-01-31",
          "url": "https://openai.com/index/building-an-early-warning-system-for-llm-aided-biological-threat-creation/",
          "origin": "published"
        }
      ]
    },
    {
      "type": "practice",
      "time": "50m",
      "task": "Set both studies against the standard below, item by item. For each, write what the report says, or \"not reported\". Then answer one question in your own words: does \"no statistically significant difference\" mean there is no difference, or that this study could not tell?"
    },
    {
      "type": "make",
      "time": "30m",
      "task": "Write a one-page reading note: the two designs side by side, where each falls short of the standard, and the one change to each design that would most improve what it can tell us. Keep the note to design and statistics."
    },
    {
      "type": "person",
      "time": "20m",
      "task": "Give the note to someone who knows statistics or trial design and ask which of your judgements they would change. Write their answer beside yours."
    },
    {
      "type": "reflect",
      "time": "10m",
      "task": "Record what you now think the two studies show, and what you would want to see before believing a claim in either direction. If you want to go further, the free course below covers pandemic prevention, detection and response.",
      "sources": [
        {
          "title": "Biosecurity",
          "author": "BlueDot Impact",
          "date": "",
          "url": "https://bluedot.org/courses/biosecurity",
          "origin": "published"
        }
      ]
    }
  ],
  "artifact": {
    "keeps": "A one-page reading note setting two studies of AI and biological misuse against a published reporting standard, with the change that would most improve each design and a reviewer's objections."
  },
  "exit": [
    "keep",
    "share",
    "fork",
    "correct-a-source",
    "close"
  ],
  "revisions": [
    {
      "date": "2026-10-27",
      "change": "First version.",
      "by": "Bentley Moon"
    }
  ]
}