Nr. 1 AI Safety Policy

AI is becoming obiquitous, we may not notice it yet, but much of our online experience is at least guided by AI, if you count Machine Learning, genetic algorithmes in that basket (so not ARGO or real world AI). People discuss AI ethics, but that’s for region in peace and control. In regions where there’s war AI can be used to spot people, target them, guidance systems, maybe jamming of radio, encryption. But nobody will wonder about ethics.

AI is still very safe, as it is not much robust. ARGO (autonomous, robust goal, orientation) is the bad guy. That’s the machine that keeps on going no matter what you throw at it, until it found you and ripped your guts out. Current algorithms really don’t have that talent. But because ethics is an endless discussion with no teeth, a simple rule should be proposed, so this is what this is about : The no-touch rule. Simply said AI should not drive its robot or actuator to ever touch a human.

The no-touch rule is just rule nr. 1. The no-touch rule is very simple, and it is simple to implement. It is an consilidation of the object-avoidance that mobile robots with AI have to do anyway. Tesla’s avoid humans. The occupancy detector (network) assigns no-go areas for the car (and Optimus), and a child on a bike will for sure occupy a volume and be avoided. This is probably not entirely robust, but certainly compared to humans.

The trouble with most rules you could think of is that you then need the robot to pay attention to it and know what to avoid. If you say ‘a robot can not meddle with food’ which is of course a risk, then the robot needs to know food, what it looks like, what food will go into which human eventually, how it will be prepared. You can’t clean meat with bleach (although this happens in industry, and ammonia, bon appetit) but a thinking robot might. It might remove chicken from the fridge in an attempt to get to the cheese, on a warm day that’s a bad idea.

To keep this short, two main strategies to keep robots safe in the beginning. One is to keep them weak. Just make the robots arm break before it can break a human arm. Optimus has motors that can pull 500 kg, but they should be housed in a way that allows something to shear in an emergency, so maybe a visible place on the arm you can hit, where a wedge will break it if anyone is in trouble. Safety through failure. This is on a mechanical level.

On a control level the no touch rule towards humans would be a start, it would be a clearly definable one. You can have an Optimus in a cubicle handling car doors all day, they don’t look like humans. Instructions to it are, 1. don’t leave the cubicle, 2. Do your tasks thank you 3. If a human arrives in the cubicle make sure you never touch him, also not with a door in you hand (which is why for humans anything we hold on to becomes part of our body image).