AI machines as moral agents, Is consciousness required? Himma’s arguments (part 11)
In Part 10 I concluded a definition for moral agency, derived from the definition of agency. But still, many consider consciousness a necessary part of any definition on both of the above. Here I discuss Kenneth Einar Himma’s arguments for why consciousness is necessary.
For a mission statement, see Part 1 — for an index, see the Overview.
5. Does agency require consciousness?
Now it is time to consider some counterarguments to why artifacts can be moral agents. I will compare Floridi and Sanders and my reasoning with Kenneth Einar Himma’s arguments for why ultimately artificial agency will necessitate consciousness from his paper “Artificial agency, consciousness, and the criteria for moral agency: what properties must an artificial agent have to be a moral agent?” (2009) and Deborah G. Johnson’s argument that artifacts will never rid itself of the intentionality of its human creators and will never be free enough to be moral agents, from her 2006 paper “Computer systems: Moral entities but not moral agents”. This will allow me to expand on my discussion of the freedom to act and indirect programming from previous sections and to argue that it is better to steer clear of contentious philosophical problems. I will also argue that Himma’s and Johnson’s arguments does not invalidate my reasoning here.
The reader should be aware that when they consider moral agency, Himma and Johnson refer to responsible moral agency, the type of moral agency this paper is not directly concerned with. However, Himma argues for an agency that relies on a form of intentionality agency, thereby refuting that the artifacts that I consider as moral agent are even mere agents. Johnson does not specifically address mere agency, only moral agency, but from her list of requirements for moral agency it is implied[1] that artifacts cannot be mere agents because they do not have internal mental states and they cannot be moral agents because they do not meet the control condition. This is why these arguments are still relevant here.
5.1. Himma’s arguments
Himma’s claim is (i) an artificial agent can only be an agent if it is conscious and (ii) the agent can only be a moral agent if it is free enough to understand what morality means and requires, which presupposes consciousness, making consciousness necessary either way (2009, p. 28).
Claim (i) is referring to intentionality agency, and his definition is:
“X is an agent if and only if X can instantiate intentional mental states capable of directly causing a performance” (2009, p. 21).
Although Himma himself joins the view that regards the relevant mental state as “volitions” which suggests intentionality agency, he does not rule out that beliefs and desire pairs also can accommodate this, as long as it is not “simply a desire or simply a belief”. That is, not a belief or a desire on its own but that they are coupled together (2009, p. 21).
My conclusion is then that if Dennett’s instrumental assignment of beliefs and desires is true, then claim (i) is false since Himma’s definition of an agent does not require X to be conscious.
Moreover, Himma adds that consciousness seems to be presupposed by agency because mental states are conscious. Unconscious mental states are “incoherent” because “mental states are characterized by the ability to be privately observed by the subject by introspection” (2009, p. 27).
It is of course possible to say that mental states are characterized by their ability to be introspected, as Himma does, and although humans mental states do (or at least mostly do) indeed have this introspective quality, he gives no reason why it should be a necessary feature of mental states. I argue that not taking a stand on whether the agent is conscious of the mental states is an advantage of using the instrumental attribution of the mental states ‘beliefs’ and ‘desires’, because it does not seem necessary. Introspection seems more like an ability that can give the agent advantages, but the question here is whether this ability is necessary to have beliefs and desires, and Himma does not have a specific argument for that.
Himma continues that the claim that ‘an artificial agent is conscious’ is problematic in two ways. First, philosophers of mind disagree whether this is possible. Some say that only biological entities can accomplish this, others say that it is only possible for complex structures as our brain, no matter what material it is made of. Second, since philosophy has yet to solve whether humans are conscious because of the problem of other minds, claiming that artifacts are conscious is even harder since we only have access to our own minds, and artifacts are so dissimilar to us, much more than humans. Determining whether an artifact is conscious involves great difficulties that are “potentially insuperable”. (2009, p. 28).
If Himma is right about this, it only gives us more reasons to not define agency this way, to not take a stand on whether an entity is conscious or not, and it makes Barandiaran et al’s minimal agency more appealing.
Most of the features Himma’s claim (ii) are referring to apply to responsibility. That the agent needs to understand morality, and that it makes no sense to praise or blame something that lacks conscious mental states — “Praise, reward, censure, and punishment are rational responses only to beings capable of experiencing conscious states like pride and shame” (2009, p. 24–25) [2]. Even so, some of these arguments are still interesting here, as the fact that being a moral agent presupposes the agent to be free. And to be free, the agent needs to be rational. If the agent acts arbitrarily, it cannot be a moral agent. The agent needs a reason why it acts, and reasons needs to be “grasped”, and grasping requires consciousness (2009, p. 24). But he also argues that reasons come from beliefs and desires of the agent and gives an example where a dog eats because she is hungry (2009, p. 24). If this explains reasons, then it can also be explained by the beliefs/desire model as argued before. And the reason Barandiaran et al added the normative condition was to rule out arbitrary and random behavior in agency, which means that the minimal agent is rational. By ‘grasping’ it seems plausible that he refers to moral understanding that is associated with responsibility, thereby only linking consciousness to responsibility, not accountability. This means that if you grant the instrumental stance on beliefs and desires, moral accountable agency does not require consciousness, that the artifact can be rational and act for reason, but that “grasping” is only needed for the artifact to be responsible.
Next, Himma also brings up another pressing issue. A “non-conscious thing” can be either:
“fully determined (and explainable) in terms of the mechanistic interactions of either mereological simples or higher-order but equally mechanistic interactions that emerge from higher order structures composed of mereological simples” (2009, p. 24)[3].
The concern raised here can be recognized from the discussion of freedom for directly and indirectly programmed systems. A directly programmed system can be characterized as “mechanistic interactions of […] mereological simples”, and it is intuitive to say that such a system is fully determined (and thereby not free) but, as I argued, for an indirectly programmed system this is not as intuitive. However, if it is also possible to say that an indirectly programmed system is fully determined, that would make it a strong argument against the possibility of such a system to be free, to have concrete freedom. I will consider such an argument next.
One advantage of an indirectly programmed system is that you do not program directly how the system classifies something, instead the programmer uses indirect methods that reinforce the artificial neurons in a way that lets the system learn what the programmer wants it to learn from a set of training data. This could give the impression that the system can learn almost anything to a very high degree, but this is not necessarily the case. It could be the case that a neural model has a specific threshold maximum where the model is ‘saturated’. An example might make this clearer: Let’s say that a model M with a specific set of parameters P can learn how to classify something (e.g. pedestrians and plastic bags from before) to a 70% degree after it has been shown X number of relevant training examples, and that it is also clear which the other 30% are (the incorrect classifications). Let’s call this instance model M1. This would mean that using M1 is actually as predictable as the directly programmed method. It only takes a certain amount of relevant training data to reach its maximum ability, after that the model is ‘saturated’, the model is fully trained. If it the case that indirectly programmed systems always have a determined saturation point, then the only difference between direct and indirect programming is that they are two different methods to approach a programming task that have some practical differences, but they are both as determined. An indirectly programmed artifact cannot be an accountable moral agent because it is not free enough, but merely a ‘set of instructions’ with the difference that these instructions are not explicitly declared.
I argue that this example of a saturated neural network is not in fact indirect programming, but direct programming. If it is the case that all neural networks can be saturated (which I find implausible), then relying on neural networks might always end up in direct programming. But there are other ways to achieve indirect programming[4]. Davenport illustrates this in the following passage:
“To be sure, in the case of robots, this may be a programmer — a knowledge engineer in expert systems terminology — entering an explicit set of rules (a program), in which case they would be responsible for the machine’s actions, but it might equally be an accumulation of information the robot happens upon in its travels, in which case there is no one to blame for the robot’s “program” but the robot itself. This is entirely equivalent to the human child growing up and being “programmed” (explicitly or otherwise) by its environment — including parents, schools, religious institutions, etc. Once we are sure we have indoctrinated them sufficiently, and assuming they do not have any mental disability, they are taken to be morally responsible for their actions” (2014, p. 54).
A directly programmed system is completely determined by the programmer. But the more the behavior of the system is determined by external input, the more its behavior will be determined by other factors than the original program, much like a child becoming a moral agent. At a certain threshold of autonomy and input, the artifact, by analogy, would also be a moral agent. That is all the freedom the artifact needs to be a moral agent. I find it plausible that if indirectly programmed system spends enough of their time “in the wild”, and the systems base enough of their behavior on what they have learned from its interactions, this analogy holds. And if these systems are complicated enough, the system’s interactions with the environment is what determines what “choices” it makes, not the specific instructions from the programmer.
Lastly, Himma concludes that “free will poses tremendous philosophical difficulties that would have to be worked out before the technology can be worked out; if we don’t know what free will is, we are not going to be able to model it technologically.” (2009, p. 24). I agree that if we do not know what free will is, it is unlikely that we will be able to model it, and if we would, it would not be possible to tell whether we have succeeded or not. However, this does not mean that it is a good idea to postpone development of trying to technologically model freedom until philosophers solve this issue. It seems more likely to me that it is the other way around, that trying to model free will is what will tell us more about what free will is than what philosophers can tell us, a discussion that has troubled philosophy for more than two and a half millennia. And it is possible that what is discovered will change the concept of free will, and that it is the old philosophical conceptions of free will that is the cause of these “tremendous difficulties”.
Comments:
Upon rereading this, I wonder how many subscribe to these views about consciousness and moral agency that Himma describes. I sympathize with the most of these ideas when it comes to humans, but do we really need to say that for artifacts to be agents that they need to exhibit a human type consciousness? Maybe artifacts can have free will and consciousness, but just of another type and level?
Part 12 here!
Footnotes:
[1] Only one of her five conditions (the last one) deals with morals. More on this in 5.2.
[2] Since Himma does not address Floridi and Sanders’ accountable/responsible distinction I cannot say what he thinks of it, which would have been interesting to know. It could be that he is not aware of it since Floridi and Sander’s “On the morality of artificial agents” is not in the references, but to the best of my knowledge he seems to be quoting it on p. 25, which suggest a reference mistake. However, he uses the terms accountable and responsible interchangeably and only addresses the responsible part of the distinction, which suggest that he does not find the distinction useful/interesting or well argued for.
[3] ‘Mereological simples’ are objects that have no proper parts, as atoms were once was thought to be.
[4] One way is a system which although it is directly programmed, it also interacts with a database that is filled with inputs from the environment, e.g. colors of objects. If the artifact is deployed in different environments, it can have very dissimilar content in the database. Or rather the artifact, which includes both the program and the database as a whole, is an artifact which is indirectly programmed.