AI machines as moral agents, Agency. (part 6)
In Part 3, 4 and 5 i talked about moral accountability and responsibility for AI machines. The following parts I discuss what is meant by saying something has agency and the concept of minimal agency.
For a mission statement, see Part 1 — for an index, see the Overview.
3. Defining agency
So far, I have not in detail defined agency and mostly assumed it to be evident what is and what is not an agent. This is not uncommon in discussions on artificial agents[1]. However, since there are definitions of agency that question whether artificial agency is even possible, this does not seem evident. Therefore, I will now look at two ways to define agency, intentionality agency and minimal agency (a less demanding agency, with ‘minimal’ necessary properties). I argue that, when debating artificial agency, minimal agency is in many ways preferable and does not introduce contentious and vague concepts in the definition.
An agent is often thought of as the source of an action or the cause of an action. Sometimes when we speak, we say things like “the avalanche caused great damage” or “I lost my job because of the recession”, and the avalanche and the recession is often referred to as the ‘agent’ that is responsible for this, some kind of a ‘causal agent’, since it is used in causal descriptions. It is very useful in summarizing and simplifying explanations, e.g. the person who lost their job does not need to explain what an economic recession is and how it evolves, but can distinctly suggest that the recession was the reason why the person lost their job, and that it was not, for example, because of his incompetence. Nonetheless, this it is not the type of agent that is of interest here.
One problem with the causal agent is that it does not seem to be the source of the event, it just seems to be the event. The avalanche might have caused great damage, but if it was started by something else, that is what we are looking for, an agent that is the direct source of the event. This is effectively saying that the ‘causal agent’ is a too wide definition and include phenomenon that are not considered direct sources of acts. Let us instead look at the more traditional way to define agency, what I will call ‘intentionality agency’ and then compare it to Barandiaran et al’s more recent minimal agency.
3.1. Intentionality agency
Markus Schlosser, when presenting a philosophical overview of agency, defines an agent as “a being with the capacity to act, and ‘agency’ denotes the exercise or manifestation of this capacity”. Action (the capacity to act) requires intentionality, and intentionality in turn depends on causation by the agent’s mental states. He calls this the standard theory of agency (p1, 2015). One way of interpreting what is meant by ‘intentionality’ comes from philosophy of mind. It is a complicated interpretation which is relevant here because it is often argued that it requires or presupposes consciousness (as e.g. Himma does, as I will discuss in chapter 5). I will refer to this definition ‘intentionality agency’ and I will argue that minimal agency (which I will introduce in the section after this) should be preferred over intentionality agency.
Intentionality, which was first introduced by Franz Brentano in the late nineteenth century, is a complicated and highly debated concept which has engaged many prominent philosophers (e.g. Kim, Searle, Putnam, Davidson, Block, Fodor, Dennett) (Jacob, 2019). Intentionality[2] describes the peculiar property of being directed at something, being about something. Beliefs, hopes, judgments, intentions, love and hatred all exhibit intentionality. It is a property that only mental phenomena exhibit, physical phenomena do not. (Jacob, 2019, p. 1–5). The mental state that X believes that mount Everest is 8848 meters high is a mental state that is “directed” at the height of the mountain, it is “about” a property of the mountain. The mountain itself is not a mental state, and it is not about something else, it just “is”.
Intentionality is an interesting philosophical concept that should be investigated by philosophers, but the question here is which view of intentionality should be included in a definition of agency. One view is that intentionality is often thought to require some type of consciousness, or subjective experience, as mental states are properties of the conscious mind. Searle’s Chinese Room argument[3] can be used to argue that agency requires consciousness[4]. It is a famous argument that basically says that a strictly computational system (e.g. a computer or a robot) can only manipulate symbols, it can never understand that a symbol (e.g. ‘mountain’) is about something, directed at something, the system cannot have intentionality; “they have only a syntax but no semantics” (Searle, 1980, p. 422). By extension this would mean that artificial agents[5] can never achieve the intentionality needed to have ‘real’ agency, that the very definition of agency, intentionality agency, makes ‘artificial agent’ a contradiction in terms since these entities are not possible.
Now, I have some concerns about intentionality agency. First, if the definition of agency is very closely connected to consciousness, which is in itself hard to pin down and define, then the definition itself becomes obscure, since until we know what consciousness is, we won’t know what agency is. If a component of a definition is not clearly understood, how can you be sure that it should be part of the definition?
Secondly, the task of explaining what consciousness is in detail seems to be an almost insurmountable task, and that task should not fall on, for example, someone trying to discuss the moral significance of advanced artifacts[6]. That would not only be unfair, it would also stifle current philosophical discussions on the morality of artifacts. I do not mean that intentionality agency cannot be the correct way to define agency because it would be unfair to those that discuss moral implication of artifacts, but rather that most of the burden of proof should lie on whoever argues that intentionality agency, not on someone who believes that highly autonomous artifacts could be sources of moral acts.
Thirdly, there is a non-negligible risk that the reason why agency is closely connected with consciousness is because of anthropomorphism. Since we consider humans agents, and humans are conscious, agency requires consciousness. But that does not show that agency requires consciousness, only that most (or all) of the entities that we currently consider as agents are conscious.
Fourthly, if we can conceive of artificial entities which share many features with human agents and we consider them sources of acts, we need appropriate terms for this ‘agency’. This is in line with what Barandiaran et al (2009) are trying to do when they are looking for ‘minimal’ agency, a more basic form of agenthood. That would not invalidate intentionality agency, only say that intentionality (or consciousness) is not necessary for minimal agency.
Besides, Searle’s argument has many critics[7] (Cole, 2019) which might disprove the Chinese Room Argument itself and thereby make it possible for computational systems to achieve intentionality.
I conclude that intentionality agency might be a too narrow definition and that it needs stronger arguments for why this should be the preferred definition, so let us look at more recent attempts of defining agency.
Comments:
My supervisor asked if I really needed to bring up Searle and the whole “consciousness thing”, and maybe I shouldn’t have in a thesis. But in this context (discussions on medium, online), unfortunately, I do think it’s a good idea, because I still believe many conflate the hard problem of consiousness with the idea of “intelligent” artifacts, AI machines. Let us separate the issues. In fact, I had a whole section about the anthropomorphism, but it didn’t make the thesis. I will add it, soon, as an addendum.
Anyway, read about the interesting theories concerning minimal agency in the next part!
Footnotes:
[1] See section 3.2.
[2] It should not be confused with what is meant by “intension” or “intention”, although it is closely related to those concepts.
[3] The main idea is that a non-Chinese speaking person sitting in a ‘Chinese room’ does not understand Chinese (semantics) although the person can manipulate Chinese symbols (syntax) with the help of instructions (a program) in a way that convinces a person outside the room that the person inside the Chinese room understands Chinese.
[4] Searle himself points out that his argument only refutes that “thinking is merely the manipulation of formal symbols” and makes no claim that “only biologically based systems like our brains can think” (1990, p. 26–7).
[5] Or at least the type of artificial agents that this text is concerned with, those that are computer based.
[6] This is not bad in itself since we should try to explicate consciousness and it is important for agency, but it is bad if it suppresses other more current discussions. I for one am sympathetic to the idea that the concept of consciousness will fall apart upon closer inspection, making it not improbable that that artificial non-biological artifacts could attain the important pieces of those parts, but as of yet I do not see how to make a strong argument for this. Hopefully neuroscience will give us important clues on this matter.
[7] For example; Margaret Boden, Ned Block, Daniel Dennett and Jerry Fodor to name a few. David Cole mentions some of the more common arguments like ‘The System Reply’ (the man in the room might not understand, but the program does (the whole system), ‘The Robot Reply’ (a natural language program like the Chinese room program might not understand, but a robot system with input/output could) or just a denial that the man does not understand or that Searle’s use of ‘understanding’ is controversial.