Can an Algorithm be Accountable?
We must ensure that our technologies contribute to public good.
It’s been almost one year since the National Science Foundation founded a Center on Responsible AI and Governance (“CRAIG”) at my university. Since then, I have been working with a team of researchers across the academy—people in as wide ranging fields as Computer Science, English, and Sociology—to partner with industrial giants to create new internal standards of governance for AI systems. Our job is essentially to determine what should be top of mind for developers so we can align all these emerging AI creations with our collective values. After all, governance is about how an organization or body politic is ruled, administered, and held accountable to a shared understanding of the public good. At the basis of all our collective endeavors, even corporate tech rollouts, we must know how we are going to ensure that our technologies contribute to that good.
This is no easy task. We’re talking deep soul searching for technologies made by entities that seem hellbent on releasing their latest advances before any of this deep reflection can be done. Even CEOs like Anthropic’s Dario Amodei—the guy attempting to put the brakes on egregious military and surveillance uses of AI that are counter to the American citizen’s interest—are releasing potentially world-breaking versions of their AI as soon as it becomes available and then sometimes backpedalling. (Sometimes not.)
And these companies are “going public,” a.k.a. becoming entirely dedicated to shareholder profits? How is that going to fare?
In an attempt to inject some sensible governance into AI R&D before it is too late, our CRAIG is developing “AI audits,” systems that check that AIs are doing what they say they are and working to our collective goals. As my colleague Jorge Ortiz has said, we need to document what AI platforms are actually doing when we put them to work—look under the hood so to speak—and create ordered accounts of the models they are deploying, the validity and safety of those models, and the potential harms and threats that they pose. We need to create transparency where none yet exists.
But that is not all. Once we know what a company’s AI system is doing, we need to then make our newfound awareness transparent and accountable. We need to create a system to continue to see what is going on, to ensure that the AI platform in question is doing right by us, by our standards of what is good for us, what is in our collective interest.
And when I say “our,” I don’t just mean a particular tech company’s best interest, but rather that public good that we all share in. Thousands of workers at Google, Meta, Amazon, Salesforce, and OpenAI have issued open letters protesting harmful AI applications in warfare and surveillance. Still more have signed petitions and conducted walk-outs and other hands-on protests. They have warned us that the very AI systems that they have had a hand in creating and bringing to market are “black boxes” (as a friend who is a developer shared with me recently, “even I don’t know what it’s doing!”) that have the power to do broadscale transnational damage.
As Nate Matias, one of the authors of the new book Auditing AI, shares, within companies, developers have been asked to open up that black box. When he worked at Microsoft SwiftKey, he was tasked with determining “when a new AI product is better than the last” and “when a model has gone off the rails in harmful ways.” Yet, as he and his coauthors explain, making AI platforms accountable to those beyond the confines of a given tech corporation requires thinking outside the black box. It requires considering how harmful it is to be caught in a system of ethical catchup that puts our entire species and planet at risk.
Right now we are chasing down companies to audit them as individual entities, making policy after the fact, as we face the tragedy of kids committing chatbot-assisted suicide, and as we learn of global data breaches jeopardizing millions worldwide. It doesn’t have to be this way.
We can start from a concept of the public good, one that considers what is good on a human, planetary level. We can then drill down to specific goals and objectives regarding where it makes sense to apply a particular AI platform (what some techies are calling Artificial Narrow AI). Then we can create an AI platform with a requisite auditing platform to ensure that this and only this use-case is activated. We can build accountability to our global community into our everyday algorithms.