Practical AI Governance summary

AI Governance. It sounds like a kind of complicated topic.

Well, that's because it is.

While we'd love to be able to share 'six quick tips' or even better 'one easy hack' to sort out your organization's approach to AI governance, the truth is, AI governance is tricky to wrap your head around and even harder to implement.

The good news though is that with a clear framework for AI governance, it becomes far more manageable than it first appears.

Before we address the 'how' of AI governance, let's look at the 'what'. As in, what exactly is AI governance?

In this context, governance is simply how an organization steers its use of AI. In other words, who decides what tools get used, how risks are managed, how mistakes get caught and how to keep on top of all of this as the technology changes. 

But you can't steer something well until you understand what you're steering. With AI this can be especially hard because AI is, for want of a better word, wobbly. It's unreliable in ways that ordinary software is not. The general term “AI” covers a sprawling range of things from the system that flags a suspicious payment on your card to the software helping a radiologist read a scan. But when we talk about AI in this content, we're largely talking about large language models, or LLMs. They are trained on vast amounts of text. This means they can confidently draft an email or schedule a meeting or even write a sonnet. But, crucially, when they do these things, they are not thinking; they are predicting. Word by word, they work out what is statistically most likely to come next, based on the patterns they have absorbed. 

Engineers have a term for this; they call AI non-deterministic. A calculator is deterministic. If you ask it for two plus two, you get four every time. An LLM is a probability machine that can be confidently, fluently wrong. Ask it to back up a claim with a source, for instance, and it may hand you a flawless-looking reference, including author, date, and title, to a study that was never written.

On its own, a single made-up answer is harmless enough. But what happens if people stop checking and confidently hallucinated output gets treated as fact and folded into real decisions or a wrong figure shapes a budget?

You can begin to see why some form of AI steering is so important. And yet, AI governance is like steering a car down a road filled with potholes and alligators and maybe there’s quicksand on either side of the road, too.