Resources: AI standards and regulations

Exercises: AI standards and regulations

Resources: Introduction to Machine Learning

Exercises: Introduction to Machine Learning

Resources: Career Advice (AIG)

Exercises: Career Advice (AIG)

Resources: Closing regulatory gaps through international agreements

Exercises: Closing regulatory gaps through international agreements

Resources: Projects  (AIG)

Exercises: Projects  (AIG)

Next steps: Programs

Resources: Challenges in achieving AI safety

Exercises: Challenges in achieving AI safety

Resources: Additional proposals

Exercises: Additional proposals

Resources: Introduction to potential catastrophic risks from AI

Exercises: Introduction to potential catastrophic risks from AI

Resources: Closing regulatory gaps through non-proliferation

Exercises: Closing regulatory gaps through non-proliferation

Even if some governments enact excellent AI regulations, AI developers in countries that do not set adequate guardrails could still cause global damage. This week and the next, we’ll consider: how can policymakers guard against that? Two (complementary) approaches are: (i) slowing the proliferation of advanced AI capabilities beyond countries with adequate guardrails (after some countries establish such guardrails); and (ii) reaching agreements in which multiple governments put in place adequate guardrails (1). These measures could each give regulated AI labs more (and hopefully enough) time to develop critical safety methods and defensive technologies. That could reduce risks from AI deployments within any country (2). Aside from these approaches, domestic regulations in any country could have international influence by creating economic incentives and setting examples.
This week, we’ll study the first of the above approaches: non-proliferation. In particular, we’ll focus on some policy levers that may help achieve this. Note there are also potential limitations and downsides to the non-proliferation approach: it may be difficult to predict which states will set adequate guardrails on AI (3); accelerating some AI developers could leave less time for developing safety methods; and adversarial approaches to AI policy could motivate counter-investments and (if a lead is maintained through aggressive government actions) potentially lead to international conflict. For these reasons, some see non-proliferation as a backup option, for cases where cooperation fails.
\-
(1)  Another framing for these two approaches is: (i) make a given “cautious coalition” have a large lead, and (ii) expand the “cautious coalition”
(2) Policymakers arguably cannot rely on countries without adequate guardrails never deploying advanced AI, since that gets easier over time. (It gets easier because, over longer time horizons, the hardware requirements for producing advanced AI will likely drop. That is because researchers find more efficient ways to use AI hardware, and because—as an alternative to quickly developing AI with large amounts of cutting-edge AI hardware—people can instead use smaller amounts of worse-quality hardware, though that takes (much) longer.) Instead of relying on countries without adequate guardrails never being able to develop advanced AI, policymakers can bet on using their countries’ leads to reduce the risks from other countries’ AI developments, by (i) developing and sharing safety methods, (ii) assessing and demonstrating risks, and (iii) deploying defensive measures. However, this is not a sure bet either, since perhaps these actions would take longer than cautious countries’ lead allows for. (Unfortunately, we are not aware of any strategy for AI safety that is a sure bet.)
(3) As of mid-2023, no major country has established regulations focused on risks from advanced AI systems. Going forward, it is not yet clear which countries will set adequate AI guardrails. For now, AI safety researchers, labs, and advocates are mostly concentrated in the US (and UK). However, as of mid-2023, most states (including the US and UK) have yet to pass major regulations focused on AI. Additionally, future developments in AI (such as clearer demonstrations of AI’s risk, or of AI’s potential) might dramatically change various countries’ receptiveness to AI safety regulation, in either direction. So could new administrations.

By the end of the unit, you should be able to:
\- Explain why nonproliferation is considered important for achieving international guardrails on AI.
\- Explain how several policy tools could be used by governments with (future) adequate AI regulations to contribute to nonproliferation. We’ll cover:
    \- Export controls
    \- Immigration
    \- Information security / selective publications
    \- Global monitoring
\- Identify three types of actions by which leading AI labs can leverage their technology to advance AI safety. E.g.:
    \- Developing safety methods
    \- Assessing and demonstrating risks
    \- Deploying defensive technologies
\- Evaluate whether or not the US has a tangible and sustainable lead on AI development, globally.


Read [our guidance](https://docs.google.com/document/d/e/2PACX-1vTPYrsCZV95S03L5-cO-p10cHz1TBMxSr_onylcSdO7KHNVmBSeOGVTq1LDlO49tXHoAJ_oZh6D4pt1/pub) on how to approach writing to learn exercises.
In this exercise, you will explore one policy tool governments can use, or action a lab can take to advance AI safety. Whilst many of the ideas sound useful in principle, digging into them quickly raises questions around how to make them feasible and/or enforced across the board. These are important hurdles to clear for any proposal for technology governance.

**Select one of the following strategic directions** from “racing through the minefield”:
- **Investing in alignment: **ensuring that AI is developed safely, if it’s going to be developed.
- **Threat assessment**: assessing the risk of misaligned AI, and potentially demonstrating it (to other actors) as well.
- **Avoiding races** between actors aiming to deploy powerful AI systems.
- **Selective information sharing:** sharing some information widely, some selectively, and some not at all.
    - You may wish to focus on just one category (widely, selectively or not at all, to narrow the scope of this prompt.)
- **Global monitoring**: identify and prevent “incautious” projects racing toward deploying dangerous AI systems.
- **Defensive deployment** of AI systems. I.e. deploying with urgency once AI systems are safe.
**Explain:**
- **Somewhat more concretely, what might this look like?**
- **Explain concrete reasons this may be feasible or infeasible to implement. You may wish to focus on a particular jurisdiction of your choice.**
- **What are the potential costs to the approach, or why might it be harmful?**
- **How well does it address the risks it addresses? E.g. there is a significant difference between mitigating a little, mitigating a lot and eliminating a risk - which may inform whether this measure is sufficient for handling this risk.**
**(Aim to write <= 200 words on each of these prompts)**

**You may wish to consider:**
- Is this best implemented as a standard set by AI labs themselves, or as a policy set by a national government? What are the incentives at play in either case?


Advancing Safety with AI Labs and Policy Tools

We include these prompts to help give you ideas about questions for further investigation. Use them as you wish.
**Prompts:**
- Given the current state of AI development and governance worldwide, what do you think are the most significant challenges in implementing global AI safety measures?
- To what extent do you believe that the US currently leads on AI development? What are some reasons or evidence that would sway you that that balance could shift, in the next 30 years?
- Many of the resources this week are oriented towards the question of reassuring policymakers that safety regulation won’t cause their jurisdiction to ‘fall behind’ economically, compared with other jurisdictions. What are some other mechanisms for achieving this reassurance?
- Are export controls sufficient for controlling proliferation of AI technology?
    - What are some instances where export controls have been circumvented, in the context of other technologies and materials?
    - In the context of AI, does restricting access to advanced hardware meaningfully limit AI development?
- The US also exercises export controls on fissile material. How useful is it to consider  compute as the analogue of fissile material, and in what ways is this analogy inappropriate?
- What are the difficulties in corporations achieving selective information sharing? Could selective information sharing be incentivised through policy?
- To what extent do you believe that AI developers will by-default take action to develop safety measures, assess and demonstrate risks, and deploy defensive technologies? What are some reasons this would be difficult? (You may wish to consider each measure in isolation).
- The AI chip supply chain is highlighted as a potential opportunity for governments to influence frontier AI development globally. Discuss your views on the ethics and possible consequences of such an influence.
- Will global monitoring be necessary for ensuring AI safety? What are some mechanisms that could be used, or scenarios we could achieve, to avoid a need for global monitoring?
- To what extent does international immigration contribute to the US tech industry? What implications could international immigration have for technology proliferation and safety, and, on balance, does it seem more beneficial to open STEM immigration or restrict it?
- In the context of the potential for AI regulations to be outpaced by unregulated development, what policy tools could governments use to quickly exert checks and balances on AI development?


AI Governance (2023)

Exercises: Closing regulatory gaps through non-proliferation

Exercises