I first came across this phrase in a Billy Connolly sketch, and learnt that it was a different way of saying the Billy Bunter habitual statement: it wasn’t me, I wasn’t there, I never threw the ball at your window.
Shirking responsibility is not a new thing. It’s as old as humanity.
Experimentation is not a new thing either. And we all learn more from mistakes than we do from getting things right.
So when I hear about rogue agents I take it with a very large pinch of salt. Even though it’s not good for my heart. Anyway, getting anxious about rogue agents isn’t good for my heart either. As I’ve said before, I’m not a AI doomer.
My last two posts were on the subject of permissions in the context of agents. I want to continue riffing on this.
When something, let’s call it a child, does something it is not meant to, and does so within a controlled environment, we learn something. That the controls were poorly designed. There was probably a time when fireplaces in family homes didn’t have fire guards. When the child relied on the parent to be the protector. Which was fine when both child and parent were in the same room at the same time. The invention of a guard allowed the parent to read a book or attend to the door or stove or whatever, albeit briefly. Controlled environments, with protections that improve and evolve over time.
Sandboxes and labs are not new things. But for them to work, they have to be designed to contain the risk and to reduce it. Minimise damage.
I guess it’s called acting responsibly. Now responsibility is a strange thing. It means very different things in different contexts. If I lived as a hermit in a cave somewhere miles from anywhere, I could do things very differently from, say, when I am standing in the middle of a crowded railway platform.
There are individual responsibilities, family responsibilities, community responsibilities and social responsibilities. And the nature of the risk posed by some potential action has to be seen in the context of the splash zone when it goes wrong.
I know very little about the Covid virus. But one small thing I know is that it is by itself inert. It has no means of travelling by itself. For a while everyone thought it was only carried in droplet form, rather than aerosol. And the sandboxes and control environments may have possibly be designed to handle the droplet risk rather than the aerosol risk. There are clever people who will know the answer.
So if the virus “escaped” from a lab, it was because of design failures, weaknesses in operating procedures and the human decisions involved. If it was lab-made then it got out of the lab because of a control failure. That someone was responsible for.
So it is with rogue agents. I’m not a fan of the anthropomorphism anyway: as I’ve said before, we don’t have the appropriate trust levels in place as humans. There is considerable risk in our accepting agents as human, even though we can learn something by pretending they are. Metaphor can be useful as a tool, as opposed to full-on imbuing agent with human characteristics and behaviours — unless, of course the agent is actually human.
Which brings me to the issue of recourse. Product liability. Environmental contamination. Hazardous substances. Whatever. We’ve evolved a whole series of protective procedures and guidelines and regulations and laws to solve for these issues.
I’ve said before that I get bored with the “because cancer” or “because China” flag-waving that goes on to defend against indiscipline and irresponsible behaviour. With all the rogue agent stories, one of the themes emerging more frequently is a question on why agents need unfettered and unconstrained access to the internet. Scraping music and art and books and news sites is also rogue behaviour. More so in cases where that which is taken illegally is then made available in some form as a substitute for the purloined object.
We’ve been talking about the contamination of the internet and the Web with AI slop for some time now. I hesitate to give credence to reports of the extent to which the internet has been contaminated by AI slop.
Agents are not bad per se. I continue to have great expectations from the world of agents. And sure, I continue to expect some things not to work, some damage done, some lessons learnt on the way. Such is life.
But those mistakes were unforeseen and not protected against. Lessons learnt; problem fixed. That’s different from making discrete decisions to train agents in deceit and subterfuge, to remove safeguards and guardrails, to unlock access to places that would otherwise be locked safely away.
Protective mechanisms can be poorly designed. Product liability. The environment can be contaminated by bad behaviour. Fines and perhaps imprisonment. Protective mechanisms can be undone, doors left unlocked, brakes removed, warning mechanisms switched off. All discrete decisions by humans.
Responsibility is a wondrous thing, God wot.
A big boy did it and ran away. Won’t wash.
I’m an optimist. We will learn to be more responsible with agents, responsible with what they can do, responsible with what we ask of them; responsible for their training, their objectives, what good looks like for them; responsible for monitoring them and for correcting behaviour.
More than anything, responsible for the impact they have on society.