How to Hold AI Accountable

AI needs to be held accountable—by people and other AI.

That’s the takeaway from Zeze Peters, an actual rocket scientist and founder/CEO of Beam.city, in his talk at the Marketing AI Conference (MAICON) 2021.

In it, Peters shares how businesses can begin to hold AI tools and systems accountable, and prevent serious issues down the line.

PS - Have you heard about the world’s leading marketing AI conference? Click here to see the incredible programming planned for MAICON 2022.

There’s no question some very smart people are scared of AI’s potential to go wrong, says Peters. He cites Elon Musk’s fears that the “danger of AI is much more than the danger of nuclear warheads.”

Much of the fear around AI comes down to explainability. Right now, many AI systems, even dangerous ones, are “black boxes.” We can’t see inside them to fully understand how or why they make decisions. Because AI can devise its own pathways to goals, sometimes even the engineers who build these systems don’t always know why they do what they do.

We need to make AI understandable by opening the black box, says Peters. There are four key ways to do that.

1. Put people back in the loop. People in organizations must be empowered to check AI assumptions, AI simulations, and AI decisions. You can’t just let these systems run without oversight.
2. Make sure AI is compatible with your business. Different AI systems use different algorithms and datasets to achieve results. When implementing AI, you’ll want to thoroughly evaluate these “ingredients” to make sure they work well with your business.
3. Implement decision signatures. These are mathematical signatures attached to decisions made by AI systems. These should be extensively implemented to make sure all decisions can actually be tracked and easily retracted if need be.
4. Create a detailed log of AI actions. In addition to decision signatures, make sure you’re logging the results of decisions and actions taken by AI systems.

In some advanced cases, says Peters, you may actually need another AI to audit your first AI system and keep it in check. (These are sometimes called generative adversarial networks.)

This may be a lot for non-technical business leaders and marketers. But it’s necessary to start understanding, so that your AI adoption efforts benefit, rather than harm, your business.

Even if you’re starting from square one, you can get your bearings by asking a simple question, then finding people who can help you answer it?

What is your AI, and what does it do?

PS - Have you heard about the world’s leading marketing AI conference? Click here to see the incredible programming planned for MAICON 2022.