Can an AI learn to suffer?

Question

I had first this question in mind "Can an AI suffer?". Suffering is important for human beings. Imagine that you are damaging your heel. Without pain, you will continue to harm it. Same for an AI. But then I told myself "Wait a second. It already exists. It is the errors and warnings that shows up". We can say it has the similar purpose as suffering. However, I felt something missing. We feel pain. The errors and bugs are just data. Let's say a robot can use machine learning and genetic programming to evolve.

Can an AI learn to suffer? And not just know it as mere information.

Answer 1

At a very high level, regarding evolutionary game theory and genetic algorithms, it is absolutely possible that AI could develop a state that is analogous with suffering, although, as you astutely point out, it would involve conditions that a computer cares about. (For instance, it might develop a feeling analogous to "being aggrieved" over non-optimality in the algorithmic sense, or "frustration" at equations don't add up, or "dissatisfaction" over goals that have not been achieved.)

The robot tormented by small children at the mall can certainly be said to be "suffering" in that the children block the performance of the robot's function, but the robot is not conscious and suffering might be said to require awareness. However, even without consciousness, this very simple robot can learn new behaviors through which it mitigates or avoids the "suffering" brought on by not being able to fulfill its function.

You definitely want to look into the concept of suffering in a philosophical context and Epicurus would be a very useful place to start.

Epicurus is directly relevant in an algorithmic sense because he uses the term "ataraxia" meaning calm, and is derived from the verb "tarasso" which means to agitate or disturb.

Ataraxia can be mathematically expressed as an equilibrium. Tarasso can be mathematically expressed as disequilibrium.

This relates directly to Game Theory in that disequilibrium can be said to be the primary requirement of games, and to AI in that Game Theory can be said to be the root of all AI.

Ataraxia is also understood in the sense of "freedom from fear", which is temporal in that fear is a function of uncertainty as it relates to the future in a predictive sense, and involves current condition vs. possible, less optimal future conditions.

Thus fear, which is a form of suffering, is rooted in computational intractability, even where the "computer" is is a human brain.

Early philosophers such as Democritus are especially useful because they were exploring critical, fundamental concepts, many of which can now be expressed with modern mathematics.

To wit: you can't arrive at suffering until you first define "the Good" and "the Bad", which is a binary relationship in which neither term can be said to have meaning without the opposite. (Mathematically, it can be expressed in its simplest form as a finite, one-dimensional graph.) This understanding is quite ancient.

---

It is worth noting that the continuing value of the early philosophers is partly a factor of wisdom not being dependent on the volume of knowledge, demonstrated by Socrates in the idea that wisdom may be as simple as knowing you don't know something.

The ancient sages didn't have the benefit of powerful measurement tools, advanced mathematics, or scientific method, but they were very smart, and even more importantly, wise.

Answer 2

This is a subject Dr. Joanna Bryson has been writing about from a couple of decades. The website has a list of published papers and drafts, and one that immediately leaps to mind is "How Do We Hold AI Itself Accountable? We Can’t."

• A core argument of Bryson is that we can't offload responsibility to AI because we can't meaningfully punish AI—current algorithms cannot be said to experience suffering

This is partly explicated in the linked paper:

What matters is that none of the costs that courts can impose on persons will matter to an AI system in the way the matter to a human. While we can easily write a program that says “Don’t put me in jail!” the fully systemic aversion to the loss of social status and years of one’s short life that a human has cannot easily be programmed into a digital artefact.

This leads to a deeper argument about the nature of computing applications in general:

generally speaking, well-designed systems are modular, and systemic stress and aversion is therefore not something that they can experience. We could add a module to a robot that consists of a timer and a bomb, and the timer is initiated whenever the robot is alone, and the bomb goes off if the timer has been running for five minutes. This would be far more destructive to the robot than ten minutes of loneliness is to a human, but it would not necessarily be any kind of motivation for that robot. For example again of a smart phone, if you added that module to your smart phone, what other component of that phone would know or care? The GPS navigator? The alarm clock? The address book? This just isn’t the way we build artefacts to work.

Genetic and other learning algorithms can certainly be designed to maximize rewards and minimize penalties, but, if they are not conscious and sentient, they can't be said to suffer—there is no coherent self to experience suffering.

Answer 3

Perhaps this strikes at the heart of the provincial way humans interpret their world. Is suffering pain the optimal way of fending off entropy? If asked, would a dog marvel at how we get through our days without smelling everything? I think the true value of this alien lifeform we are creating called AI (perhaps more accurately MI, machine intelligence, intelligence just is) is that it will not be like us. Of course this is it's big risk also.