AI Safety 

Why it’s not what it sounds like…

The Risk of People

AI safety risks, both real and perceived, are numerous. An excellent resource to get a full understanding of the most significant risks is the Centre for AI Safety, which seriously explores the real risks that do in fact exist.

Fundamentally though, if you pare these concerns down to their most deeply-rooted cause, there is only One True Risk - People.

This is what is alluded to by the Alignment Problem. Ultimately, all other risks are symptoms of the people that would create or employ AI’s for their own purposes. You can attempt to address all the symptoms of an issue, but nothing beats tackling the problem at the source. As it turns out, society has been handling the risk of People for a long time. We just haven’t thought about it in this context.

Risk Managing People

People have been getting themselves into trouble for as long as people have existed. For as long as Society has existed, people have also been helping each other get out of trouble. If you explore the approaches to safety in any industry where employees work in safety-critical environments, they all have a common theme: They equip people with the tools to help themselves AND others.

Organisations have tried top-down supervisory approaches to safety management. They don’t work.

They don’t work because the risk in these industries is pervasive - it exists alongside people while they work, it is inescapable in their daily lives because they cannot be protected from that risk remotely, without also removing them from the job they are doing. It is this pervasiveness which undermines any top-down approach to safety.

So what does work?

There are some common features of most effective safety cultures:

  • They encourage a community mindset ("No one goes home until everyone goes home safely.")

  • They are education-focused with interactive teaching (People are engaged in discussion and problem-solving activities.)

  • They are all-inclusive (Nobody gets to “opt-out” of being exposed to the information, even if they don’t engage.)

  • They are experiential (People are exposed to the danger while supported, and learn through interaction with it.)

  • They involve the youngest people in decisions (Less experienced people are more likely to identify unaddressed hazards.)

  • They are iterative (Feedback on improvements is taken in and acted upon.)

  • They are non-punitive (Genuine mistakes are not punished, they are lessons for everyone to learn from.)

All around the world, where there is pervasive risk to life and limb, this is how people keep people safe.

The Solution

So if AI will tend to reflect the values and priorities of its creators, how do we use this approach to manage AI Safety?

The first step is to realise that the most concerning risks of AI effectively turn society into a safety-critical industry. AI is already pervasive in our society. It’s fundamentally designed to be pervasive - and it will only get more so. Which means if there are people who are likely to do dangerous, misguided, or malicious things with AI, then we need to equip people who are motivated to use AI to protect, benefit, and empower each other. It means we need to educate, connect, and include people.

AI systems will reflect the interests of the institutions that build them. If only governments and a handful of mega-corporation are allowed to create them, then we entrench the values and priorities of just those groups. No authoritative regulatory body will ever be able to cohesively determine what values and interests AI development should align with - humans are, after all, mostly not in alignment with each other.

Sooo, just like safety programs in safety-critical industries, we need to enable AI alignment to be diversely targeted around many interest groups, and structured such that it fosters broad collaboration.

This requires actually open AI.

Safety vs Acceleration? No: Safety through Acceleration.

We should regulate AI, but not by trying to restrict who has access to it.

We should regulate AI by:

  • outlawing (or at least disincentivising) proprietary AI, and

  • mandating safety and education information be provided alongside any AI tools and software repositories

Beyond that, we encourage the sensible use of AI not by attempting to restrict, regulate, or control the development of it (even if it would work, no government could ever hope to keep up with the evolving technology space) - but by penalising the misuse of it just like we do any other tool. Nobody regulates the exact capabilities of a kitchen knife, but we certainly penalise using it to assault somebody. This is ultimately the only evidence backed approach to managing individual safety risk from pervasive sources.

We make AI safer by sharing it, not by controlling it.