A smart approach to AI regulation

Americans see AI systems incorrectly matching mugshots of 28 congressmen and criminals, demonstrating prejudice against women and people of color, and causing lawyers to falsely cite fake cases. have witnessed.

A recent MITER-Harris poll found that most Americans are concerned about AI in high-value applications such as self-driving cars, access to government benefits, and healthcare. Only 48% of them believe AI is safe, and 78% are very or somewhat concerned that AI could be used for malicious purposes. And 82% support government regulation of AI.

Even Sam Altman, CEO of OpenAI, which developed ChatGPT, is calling for regulation.

The rapid growth of Large Language Models (LLMs) like ChatGPT has changed the way people look at AI. It is personified and viewed as an independent entity with its own subjectivity and purpose. This is very different from just a year ago when most people understood AI to be smart software that resides within digital systems. When considering approaches to AI regulation, he finds it useful to treat these two models separately.

For AI as a component within an engineered system, there is a need for AI assurance, that is, ensuring that AI applications perform as expected at the right time, in the right circumstances, without unacceptable risk. It is important. Like any software component, it requires testing and validation to verify its robustness, security, correctness, etc. These characteristics are best assessed by existing regulators already in charge of related industries such as healthcare and critical infrastructure. These regulators would be a good group to assess risk within the context of their own industry norms.

However, this new realm of LLM with human-like behavior and understanding requires a different approach. There are several scenarios to consider.

First, humans will use AI as co-pilots to more effectively conduct unwanted or criminal digital activities such as cyberattacks and misinformation campaigns. His LLM here is an enabler, but humans are ultimately responsible for AI-enhanced behavior in cyberspace. We need to ensure that these acts can be prevented, defended, remedied and attributed, as they are now, but on a much larger scale.

Second, humans will give agent AI systems malicious targets. For example, AgentGPT is her instance of GPT with Internet access that attempts to perform high-level tasks by developing and executing a series of derived tasks. ChaosGPT is an instance of his AgentGPT tasked with exterminating humanity. Luckily, ChaosGPT’s hacking of state nuclear weapons isn’t all that advanced, but it represents an exaggerated example of what might happen in the future. Expect criminal ransomware groups to start using these tools in the near future.

human responsibility

Solutions to address this scenario include holding humans accountable for giving the system a malicious purpose, and continuously increasing the level of assurance of critical digital infrastructure to prevent successful AI-coordinated network intrusions. includes improving. The interfaces through which AI is intentionally connected to such systems also need to be carefully regulated.

Third, there is considerable concern about AI systems accidentally setting their own sub-goals that could unintentionally pose a hazard. While this can happen, it’s just as likely that a malicious human could give an AI a dangerous sub-goal. So the solution for this scenario is the same as the example above.

Here are some other ways to avoid the potential dangers of AI technology without impeding innovative research and development.

— Best practices and regulatory standards that apply to traditional software should also extend to AI components. However, it is important to recognize that AI software can introduce inherent vulnerabilities that require assurance measures such as testing standards, standardized code development practices, and rigorous validation frameworks.

— Regulated industries must develop response plans based on the National Institute of Standards and Technology (NIST) AI Risk Management Framework. Compliance with the NIST framework should be the starting point for identifying potential regulatory approaches in these industries.

— Creating a “warranty case” prior to deployment provides documented evidence that the system complies with important warranty characteristics and operating boundaries.

— For AI intended to augment human capabilities, regulation should prioritize system transparency and auditability. To hold individuals accountable for intentional misuse of AI to cause harm, the intent and execution must be documented.

We will complement our regulatory efforts with further investments in research, create a common vocabulary and framework for AI collaboration to guide future research efforts, and ensure that advances in AI impact human values ​​and society. You can be sure that your happiness will match. And to automatically detect fake content, more investment in research and development is essential.

The AI ​​revolution is moving so fast that we can’t wait long to address these concerns. Taking action now to create foundational regulations for AI assurance can ensure that AI technology operates in a safe and responsible manner. Harnessing the full potential of AI while protecting society requires thoughtful regulation that balances innovation and risk mitigation.

T. Charles Clancy Senior Vice President of MITER, General Manager and Chief Futurist of MITER Labs. This article is based on the recently published paper “A sensible regulatory framework for AI security“