Oxford University Press's
Academic Insights for the Thinking World

What are the moral implications of intelligent AGI? [excerpt]

The possibility of human-level Artificial General Intelligence (AGI) remains controversial. While its timeline remain uncertain, the question remains: if engineers are able to develop truly intelligent AGI, how would human-computer interactions change? In the following excerpt from AI: Its Nature and Future, artificial intelligence expert Margaret A. Boden discusses the philosophical consequences behind truly intelligent AGI.

Would we—should we?—accept a human-level AGI as a member of our moral community? If we did, this would have significant practical consequences. For it would affect human–computer interaction in three ways.

First, the AGI would receive our moral concern—as animals do. We would respect its interests, up to a point. If it asked someone to interrupt their rest or crossword puzzle to help it achieve a “high-priority” goal, they’d do so. (Have you never got up out of your armchair to walk the dog, or to let a ladybird out into the garden?) The more we judged that its interests mattered to it, the more we’d feel obliged to respect them. However, that judgement would depend largely on whether we attributed phenomenal consciousness (including felt emotions) to the AGI.

Second, we would regard its actions as morally evaluable. Today’s killer drones aren’t morally responsible. But perhaps a truly intelligent AGI would be? Presumably, its decisions could be affected by our reactions to them: by our praise or blame. If not, there’s no community. It could learn to be “moral” much as an infant (or a dog) can learn to be well behaved, or an older child to be considerate. Even punishment might be justified, on instrumental grounds.

“Some philosophers deny the reality of the self, but AI-influenced thinkers don’t. They see it as a specific type of virtual machine.”

And third, we’d make it the target of argument and persuasion about moral decisions. It might even offer moral advice to people. For us to engage seriously in such conversations, we’d need to be confident that (besides having human-level intelligence) it was amenable to specifically moral considerations. But just what does that mean? Ethicists disagree profoundly not only about the content of morality but also about its philosophical basis.

The more one considers the implications of “moral community,” the more problematic the notion of admitting AGIs seems to be. Indeed, most people have a strong intuition that the very suggestion is absurd.

That intuition arises largely because the concept of moral responsibility is intimately linked to others—conscious agency, freedom, and self—that contribute to our notion of humanity as such.

Conscious deliberation makes our choices more morally accountable (although unconsidered actions can be criticized, too). Moral praise or blame is attributed to the agent, or self, concerned. And actions done under strong constraints are less open to blame than those made freely.

These concepts are hugely controversial even when applied to people. Applying them to machines seems inappropriate—not least due to the implications for human–computer interactions cited in the previous section. Nevertheless, taking the “mind-as-virtual-machine” approach to human minds can help us to understand these phenomena in our own case.

AI-influenced philosophers analyze freedom in terms of certain sorts of cognitive-motivational complexity. They point out that people are clearly “free” in ways that crickets, for instance, aren’t. Female crickets find their mates by a hardwired reflex response. But a woman seeking a mate has many strategies available. She also has many other motives besides mating—not all of which can be satisfied simultaneously. She manages, nevertheless—thanks to computational resources (a.k.a. intelligence) that crickets lack.

These resources, organized by functional consciousness, include perceptual learning; anticipatory planning; default assignment; preference ranking; counterfactual reasoning; and emotionally guided action scheduling. Indeed, Dennett uses such concepts— and a host of telling examples—to explain human freedom. So AI helps us to understand how our own free choice is possible.

Determinism/indeterminism is largely a red herring. There is some element of indeterminism in human action, but this can’t occur at the point of decision because that would undermine moral responsibility. It could, however, affect the considerations that arise during deliberation. The agent may or may not think of x, or be reminded of y—where x and y include both facts and moral values. For instance, someone’s choice of a birthday present may be influenced by their accidentally noticing something that reminds them that the potential recipient likes purple, or supports animals’ rights.

All the computational resources just listed would be available to a human-level AGI. So, unless free choice must also involve phenomenal consciousness (and if one rejects computational analyses of that), it seems that our imaginary AGI would have freedom. If we could make sense of the AGI’s having various motives that mattered to it, then distinctions could even be made between its choosing “freely” or “under constraint.” However, that “if” is a very big one.

As for the self, AI researchers stress the role of recursive computation, in which a process can operate upon itself. Many traditional philosophical puzzles concerning self-knowledge (and self-deception) can be dissolved by this AI-familiar idea.

But what is “self-knowledge” knowledge of? Some philosophers deny the reality of the self, but AI-influenced thinkers don’t. They see it as a specific type of virtual machine.

For them, the self is an enduring computational structure that organizes and rationalizes the agent’s actions—especially their carefully considered voluntary actions. (LIDA’s author, for instance, describes it as “the enduring context of experience that organizes and stabilizes experiences across many different local contexts.”) It isn’t present in the newborn baby, but is a lifelong construction— to some extent amenable to deliberate self-molding. And its multi-dimensionality allows for considerable variation, generating recognizably individual agency, and personal idiosyncrasy.

That’s possible because the agent’s Theory of Mind (which initially interprets the behavior of others) is applied, reflexively, to one’s own thoughts and actions. It makes sense of them in terms of prioritized motives, intentions, and goals. These, in turn, are organized by enduring individual preferences, personal relationships, and moral/political values. This computational architecture allows for the construction of both self-image (representing the sort of person one believes one is) and ideal self-image (the sort of person one would like to be), and for actions and emotions grounded in the differences between the two.

In sum: deciding to credit AGIs with real human-level intelligence— involving morality, freedom, and self—would be a big step, with significant practical implications. Those whose intuition rejects the whole idea as fundamentally mistaken may well be correct. Unfortunately, their intuition can’t be buttressed by non-controversial philosophical arguments. There’s no consensus on these matters, so there are no easy answers.

Featured image credit: “abstract-lines-numbering-system” by geralt. CC0 via Pixabay.