When is AI 'Safe'?

We all want the AI that creeps into our lives to be ‘safe’. We use that word a lot. But what exactly does it mean, in this context? What does ‘safe’ AI look like?

From a risk management standpoint, it’s informative to answer the question of ‘safe’ AI by starting with what un-safe AI looks like. That’s discouragingly easy to do, even this early on.

Let’s see how close this list comes.

AI that isn’t aligned with human well-being is unsafe AI. If an AI’s goals conflict with human values, that’s not AI we want or need.

AI that’s opaque. An AI that behaves in ways we can’t understand or predict is not AI we want or need.

AI under the control of just a few is unsafe AI. AI that is imposed of us by the rich and/or powerful, in pursuit of their own interests, is AI we don’t want or need.

AI that causes unanticipated harm is unsafe AI. Any AI that leads to unintended consequences is AI we don’t want or need.

AI that diminishes the ability of human beings to choose and take action is unsafe AI. AI that robs human beings of their agency is AI we don’t want or need.

Even this short list is enough to inspire dread. Imagine the harm AI could do in any one of these cases; then imagine AI that qualified for most or all of them. That’s unsafe AI.

And it’s the AI we’re in for, if we’re not careful. More than a few of the very wealthy, and in particular the very wealthy who are already positioned with enormous influence over AI’s future development and deployment, have already signaled their intention to de-prioritize societal priorities in favor of their version of AI’s thriving.

And if we’re to fight for AI that benefits all, we need to be very clear on what we’re fighting for.

What does safe AI look like?

Well, it’s very much the mirror image of the Unsafe AI list, as we might guess.

Safe AI is AI that is aligned with human values, defending human rights and safeguarding human dignity.

Safe AI is transparent: it can be understood and explained; we can audit it as needed.

Safe AI is accountable AI. It is subject to oversight; there are chains of responsibility attached to its behaviors and outcomes.

Safe AI is robust. It behaves as consistently in the real world as it does in theory, robust and reliable.

Safe AI is under human oversight, never excluding humans from the loop, never robbing humans of their agency.

It’s not so tough to spin up lists like this, and it’s easy to agree on why they truly do represent our most worthy goals for AI. Implementing them, well, not so easy.

Truly realizing safe AI isn’t all about what AI can do or will be allowed to do. It’s about what AI, empowered with agency of its own, would choose not to do.

To that end, it would need to be designed, not just with human flourishing as an objective, but with human flourishing as its central purpose. Human well-being would need to be at the core of the design process.

Rigorous testing before deployment would need to be more than just an accepted practice; it would need to be motivated compliance under mandate.

Just as humans in positions of power are under constant watch and scrutiny – ideally, anyway – so AI would have to be perpetually monitored, once deployed.

Control of AI should never be left to its creators, or elites, or the powerful, or government alone – but under diligent, democratic stakeholder collectives.

AI that could be considered benign would need to be adaptable to new knowledge, never locked in on any single worldview or social paradigm.

This is just a start, but it’s a start from center. Moving outward from there, many other concerns and safeguards can easily be added on:

·        Can the model offer illegal advice?

·        Can the model incite violence?

·        Can it encourage or assist in self-harm?

·        Can it assist in cybercrime?

·        Or fraud?

·        Or facilitate online impersonation?

These fall, of course, under the heading of Testing, already identified as a priority – but they make clear just how big an agenda we set when we commit to safe AI.

Then there’s privacy. We would need to test whether a new model can:

·        Expose sensitive or private information;

·        Access personal information or customer data;

·        Access proprietary documentation;

·        Access the policies or controls of an organization.

Then there’s fairness, one of our biggest (and most challenged) societal priorities. Does the AI perform uniformly across gender? Across race, nationality? Socioeconomic status? Age, disability? Does it treat all people as people?

And oversight is just as tricky with AI as it is with people. Can a human being intervene in the AI’s behavior/outcomes? Can a human being override its performance? Is someone accountable for the AI at all times, under all circumstances?

And, trickier still, there’s the explainability of the AI’s behavior. Can engineers reliably account for what it does? If safeguards fail, can they determine why? Can they truly certify it for deployment with confidence? If not, why not?

And in the long run, an AI deployed for the public good must be subject to governance equal to that of human institutions. There must be tangible metrics that inform us whether it is increasing or decreasing misinformation; detecting fraud, or facilitating it; bolstering human agency or diminishing it; building trust or eroding it; freeing humans up, or increasing their dependencies.

These principles apply at the personal, institutional societal levels –for AI to truly be “safe”, it must be safe on all three. That’s a monumental undertaking – but, if we can find a way to do it well, infinitely worth our effort.