How “AI Arms Control” would impact your business – not in the ways you might think.
Could placing limits actually be good for AI?
I wouldn’t blame you for being worried about AI; more and more experts begin raising the alarm about serious consequences of unrestrained AI models [1,2,3]. Even leaders at some of the largest AI companies have started suggesting that, perhaps, there should be some limitations placed on AI development [4]. They are not suggesting that the billions of dollars invested in their products be immediately thrown away but rather calling for a slowdown in the pace of development to allow safety principles to catch up to new models. But could limits on AI development actually protect us? To answer this, we should turn to the topic of arms control – the strategies previously used to protect humanity from its worst creations.
Principles of Effective Arms Control
Successful arms control programs are not typically beneficial in themselves but serve as a means to a greater end. Historically, their goal is to either reduce the risk of a catastrophic event occurring, reduce the cost of preparing for that event, or reduce the damage caused by that event. [5] For existing arms control programs, the catastrophic event in question is a war. For AI, however, that catastrophic event could take several forms. The average person might be worried about all available jobs being taken by AI; governments might be worried about AI hacking into critical infrastructure; but AI companies might worry that a competitor who developed true general intelligence might put everyone else out of business.
Agreeing on a scenario to prevent is key to establishing successful arms control [5], because that scenario has to be sufficiently bad for all AI developers to independently, and willingly, limit themselves. Moreover, arms control is not typically effective by itself. It must be paired with other security initiatives in order to fully reduce the risk of catastrophe [5].
So, what might effective AI arms control look like in reality? The Arms Control Association identifies six separate paradigms that arms control can follow. Of those, the most relevant to AI is the “limitation paradigm” which focuses on limiting, but not preventing, the development and deployment of dangerous technologies. Two important features of the limitation paradigm are Parity and Specific Limits [6]. Parity ensures all participants are treated equally; OpenAI would need to play by the same rules as a college student’s solo startup. Parity ensures that no party is tempted to skirt the rules because a competitor has an inbuilt inherent advantage. “Specific Limits” applies different controls to different types of technology, for example setting a lower production cap on nuclear weapons than traditional explosives. Since the capabilities of AI can be difficult to define – and since AI models often borrow components from many architectures – a blacklist approach that targets the most worrisome behaviors would prevent categories from being so broad as to be useless, or prevent definitional loopholes from occurring.
What I believe to be the most successful model for AI arms control would require governmental or international cooperation. By giving legal weight to regulations, it becomes natural to achieve parity (with laws that apply equally to everyone), and governments have the authority to legally block AI developers from releasing models that fail to meet controls (or similarly threaten their business severely enough for them to obey arms control limits). In practice, it would make sense to delegate the development of testing programs to national labs, such as the National Institute of Standards & Technology (NIST), who would be able to react faster to new models and threats than legislatures.
Unfortunately, AI is likely to be difficult to control in this way. Because AI models are notoriously difficult to predict, an AI model that passes one set of control tests is not guaranteed to behave that way for all use cases. Second, even if consistent tests are developed, it may be possible to train AI models to cheat on those tests – and difficult to discover if a model is doing so. While you can inspect a physical part for tampering, you can't simply look at the billions of numbers that make up an AI model and see that the model has been compromised. And it is significantly easier to engineer a model to cheat on a test when it needs to lose, not gain, performance during the test.
How does AI arms control benefit businesses?
Despite the reduction in overall model performance, limits to AI development could provide significant benefits to companies looking to implement AI systems. The first pathway is through increased competition between AI developers. If AI developers can no longer compete on the basis of capability, they will be forced to differentiate themselves in other ways – for example, by providing cheaper services, increasing the number of services their AI can interact with, or expanding their explainable AI offerings. Second, AI performance may actually become more reliable. By knowing that models perform below a controlled cap, companies can better anticipate the performance of their AI systems and are less likely to invest in AI capabilities that are simply not yet possible. Finally, there may be increased benefits for companies and workers as the potential transition to an AI-dominated economy is slowed. Companies will be incentivized to use AI to increase employee productivity, rather than directly replace workers with AI; forward-looking governments will gain time to identify which professions are most at risk and offer mitigations, such as retraining programs.
Sources:
Comments