Dual-Use Research Disclosure: When Publishing AI Capabilities Creates Risk

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ethics-safety-dual-use-disclosure-adults

1. What characteristic defines AI research as dual-use?

   1. Research that produces knowledge or tools with both beneficial and harmful applications
   2. Research conducted by both academic institutions and private companies
   3. Research that requires approval from two different ethics boards
   4. Research that involves two or more machine learning techniques

2. According to the framework in this material, what is the key empirical question to ask when deciding whether to publish capability research?

   1. Is the research funded by government agencies?
   2. Does the model have more than 10 billion parameters?
   3. Does releasing this materially help people who couldn't otherwise cause harm?
   4. Does the research use data from human subjects?

3. Under what condition is full open release of AI research considered most appropriate?

   1. When the research has never been peer reviewed
   2. When the funding source requires public release
   3. When the capability is already widely available and the marginal uplift from publication is low
   4. When the model demonstrates emergent capabilities

4. What does staged release involve as described in this material?

   1. Providing access to only the model's API, not its architecture
   2. Releasing the paper and high-level findings while holding back model weights or fine-tuning details
   3. Publishing incrementally over several months with continuous updates
   4. Making the model available in limited geographic regions first

5. What is the purpose of coordinated disclosure in the disclosure spectrum?

   1. To share findings exclusively with government agencies
   2. To delay publication until a competitor releases similar work
   3. To publish research only after removing all technical details
   4. To notify affected parties such as other labs or security vendors before public release, giving them time to prepare defenses

6. A researcher discovers a vulnerability that could allow AI systems to be manipulated into generating harmful content. What approach represents redacted publication?

   1. Publishing the full paper but requiring a non-disclosure agreement for readers
   2. Publishing the paper with specific technical details about exploitation removed
   3. Publishing in a closed-access journal with limited distribution
   4. Publishing only the abstract and keeping the full paper private

7. When might the 'no release' option be appropriate for dual-use AI research?

   1. When the funding agency requests confidentiality
   2. When the research fails to produce significant results
   3. When the researcher plans to commercialize the findings
   4. When the harm potential is sufficiently extreme and irreversible

8. What does the material identify as a key obligation for deployers beyond research labs?

   1. Ensuring all users sign liability waivers
   2. Limiting deployment to enterprise customers only
   3. Hiring additional security personnel to monitor model outputs
   4. Developing a process for handling discoveries of harmful use cases by users

9. Which of the following is identified as a red line that should not be released?

   1. AI systems that summarize news articles
   2. AI systems that can write Python code
   3. AI systems that provide meaningful uplift for weapons of mass destruction
   4. AI systems that generate artistic images

10. The material notes that restricting use in terms of service is insufficient because:

    1. Users always follow terms of service even when they want to cause harm
    2. A determined bad actor will not read terms of service, so access controls and capability monitoring are needed
    3. Terms of service are legally binding and enforceable in all jurisdictions
    4. Terms of service prevent all misuse of AI systems

11. What is the overarching framework for disclosure decisions described in this material?

    1. A simple checklist that can be completed in under an hour
    2. A requirement to consult with all major AI labs before release
    3. An explicit benefit-harm calculus that should be built before a capability ships
    4. A default position of publishing everything unless legally prohibited

12. If an AI model can already be purchased from three different vendors, how does this affect the disclosure calculus?

    1. The marginal uplift from publication is low, making full release more acceptable
    2. The researchers should halt all work on similar capabilities
    3. The research must be classified as a military secret
    4. Publication should be delayed until the model is unique

13. Which example best illustrates a capability that would cross a red line according to this material?

    1. An AI system that meaningfully undermines oversight of powerful AI systems
    2. An AI system that corrects grammar more accurately than existing tools
    3. An AI system that translates between common languages
    4. An AI system that generates personalized product recommendations

14. Why might publishing use case documentation be an obligation for AI deployers?

    1. If users discover ways to use the product to cause harm, documentation helps others understand risks and develop mitigations
    2. Documentation prevents all potential misuse of the system
    3. Documentation is required by copyright law for all software products
    4. Users will not adopt AI tools without detailed documentation

15. What is 'uplift' in the context of dual-use AI research?

    1. The improvement in model performance after fine-tuning
    2. The elevation of a researcher to a more senior position
    3. The additional capability or advantage that publication provides to threat actors who would otherwise lack that capability
    4. The increase in a model's parameters during training