AI machines as moral agents, mission statement (part 1).
Two years ago, I finished my master thesis “Artificial agency and moral agency: an argument for morally accountable artifacts”. It was an attempt to show the path for artificial agents (robots, bots, self-driving cars etc.) to become moral agents. Here, in a 13 part series, I will publish my thesis, with comments. For an index, see the Overview.
Now, a master thesis in philosophy is a particular way of writing. First, of course, you need to be familiar with the literature and clearly quote your sources. Secondly, your argument needs to be substantiated and your conclusion valid. This is where my thesis failed.
In order to do a great thesis, you must make sure that the scope of you thesis is narrow enough to ensure that you are not taking on the task to prove too many things, which is not only very hard, you probably don’t have the space for it even if you could (40–50 pages in my case).
And, the task is in itself already very hard as it is. As you might have guessed, this what I did. Although I was very aware of this common mistake, warned others about it, I fell into the trap myself.
BUT, what I am driving at, I still feel that there are many parts of my thesis that are worth publishing, especially as “AI” is definitely here to stay and will become much more commonplace in the future and also since there are still many misconceptions about concepts like moral agents, consciousness, artificial agents, the human mind — that I want to explain and discuss with peers. I’m not referring to my own thoughts on these subjects, but the mix of great research which the thesis references — Barandiaran et al on minimal agency, Fiala et al on dual cognition, Legg and Hutter on machine intelligence and philosophers Floridi & Sanders, Daniel Dennett, Alan Turing and a few more.
So, then, what was the thesis’ basic argument?
- That an artificial agent (a ‘machine’) can be accountable for moral acts.
This is something that many, instinctively, reject. One reasoning for this is summarized by the saying “Guns don’t kill people, people do”. Although it is the actual bullet from the gun that caused the victim to die, the gun didn’t decide to pull the trigger, the person using it did.
But what if the gun malfunctioned and a shot was fired? Who is to blame then?
Since the gun has no ‘intelligence’, we still won’t blame it. And if there is no one to blame, it was an accident. The gun has no agency — it has no intelligence, no will, no direction. So the first question is, can a machine be an agent? When you introduce artificial intelligence (AI), this, of course, at least becomes more complicated.
Now, even if you say that a machine can be some sort of agent, can it ever enter the realm of the moral world? This is also intuitively rejected by many. One argument for this is that the agent needs consciousness to be a moral agent. And if you compare a regular gun with a person, it’s easy to enumerate the many differences between the two, and to explain why a gun isn’t conscious. But the idea of consciousness is very complicated, and also very hard to explain. And I seen no reason why AI agents of the future can’t be conscious. They might not be conscious in the same way humans are, but do they have to be? And can we really define what we mean by ‘human consciousness’? If we can’t do that, exactly what is it that AI agents can’t achieve?
In the thesis, I also make use of Floridi & Sanders distinction between accountability and responsibility — which basically says that accountability is a light version of responsibility. This is sort of a cop out in order to narrow the final claim of the thesis, to not make too strong of a statement. But with a large portion of imagination, I don’t think this limitation is necessary. However, I do think that morals is social concept created by humans, and that it is constantly evolving. And as AI agents become more autonomous, we will need to incorporate this fact into our moral thinking, and this evolution is hard to foresee. It depends on what artificial agents will be commonplace in tomorrows society and how it will change our society, our way of living, our laws and what moral problems this will cause and needs to be addressed.
So, what I will do is to publish here on medium the different sections of the thesis in smaller parts, with comments. Hopefully you find these subjects as interesting as I do, and do comment or criticize if you are inspired to, that would make it all worth while.The next part will be the introduction of the thesis that will go into a lot more detail how the thesis is structured and what the argument is all about.
The thesis is divided into 6 chapters with subsections that I will reference as 2.1 and so on, but here the whole thesis is 12 parts (2–13), excluding this one. There is also an overview of all the parts for reference. I will leave you with the abstract (a short summary) of the text.
Abstract
“Machine learning in modern artificial intelligence (AI) systems has led some to draw the conclusion that these systems can become or should be considered moral agents comparable to humans. However, the concept of agency in itself is often either defined from what consciousness consists of or based on properties which seem to necessitate consciousness, resulting in the controversial claim that these artifacts are conscious. I will argue, as per Floridi and Sanders, that artifacts can be morally accountable and thereby moral agents but not necessarily morally responsible. Also, Barandiaran et al’s definition of ‘minimal’ agency avoids strong claims of consciousness for agents and coupled with a learning criterion can define an accountable moral agent. This can facilitate discussions concerning AI technology and remove the burden for these discussions to explain consciousness.”
Go to part 2, the introduction!