AI Ethics: How to Think About Risks
Google DeepMind just published a paper on AI risks and how they’re planning to mitigate them. Even though it is focused on harms to all humanity rather than everyday harm, it’s notable in its structured approach to thinking about AI risks. Consequently, I believe it can form a foundation for nonprofits to think about the implications of their own AI use.
The paper identifies four primary risk areas: misuse, misalignment, mistakes, and structural risks. Bear with me as we go through them comprehensively. Google defines these risks as follows in the quotations (with my nonprofit context replacing or supplementing their examples):
“Misuse: The user intentionally instructs the AI system to take actions that cause harm, against the intent of the developer.” For example, your employee might use AI to try to overcome security barriers that prevent certain employees from accessing private information about the people you serve.
“Misalignment: The AI system knowingly causes harm against the intent of the developer. For example, an AI system may provide confident answers that stand up to scrutiny from human overseers, but the AI knows the answers are actually incorrect.” Their definition includes what we often call “hallucinations,” or deception, scheming and loss of control. In a nonprofit context, someone on your marketing team might use a picture because the AI claims it isn’t copyrighted when it actually is.
“Mistakes: The AI system produces a short sequence of outputs that directly cause harm, but the AI system did not know that the outputs would lead to harmful consequences that the developer did not intend.” For example, the AI might propose a series of lectures for a course that inadvertently lead students in the direction of learning the material with built-in bias.
“Structural risks: These are harms arising from multi-agent dynamics – involving multiple people, organizations, or AI systems – which would not have been prevented simply by changing one person’s behavior, one system’s alignment, or one system’s safety controls.” This risk prevents the most difficulty in pinning down, but I propose the following (albeit unlikely) example: Your nonprofit introduces an AI-driven financial planning tool for a certain segment of the population, and it eventually causes the AI models at the banks who serve them to alter their transaction approval criteria to disadvantage these same people you intended to help.
How can we use this information for community benefit organizations? Create a systematic review process in your nonprofit when introducing new uses of AI. You could:
Misuse: Include in your AI policy and training some specific prohibitions and an explanation of why you’re preventing this kind of use.
Misalignment: Train your team to follow the “AI sandwich” method of use. The bottom bread is the first draft, for which you use AI. The meat is your own knowledge, which you use to ensure that you’re dealing with accurate information. The top bread is also AI, which you can use for purposes that are unlikely to cause harm such as correcting grammar. I would add that you can mitigate these problems by being transparent with your supporters, service recipients and community members about when you use AI.
Mistakes: This one is tricky, but you should have an AI committee that regularly reviews new and existing AI tools to check that you’re not introducing bias, unfairness or false claims that are often baked into AI training data.
Structural risks: This risk is the hardest of all to prevent, but I submit to you that one way to help is simply good governance combined with frequent discussion with your community partners in which you review possible implications of your collective AI use.