Cloud Hosting

[You can hear Robert Sparrow discuss the dangers of Artificial Intelligence with Waleed Aly and Scott Stephens on The Minefield.]

When asked about humanitys future relationship with computers, Marvin Minsky, one of the founding fathers of the field of Artificial Intelligence (AI), famously replied, If were lucky, they might decide to keep us as pets.

While the view is not shared universally among computer scientists, serious, intelligent, people believe that there is a real danger that AIs will enslave perhaps even destroy humanity. Swedish philosopher Nick Bostrom has written an entire book, Superintelligence, discussing the threat posed to humanity by AI.

One might think that it would swiftly follow that we should abandon the pursuit of AI. Instead, most of those who purport to be concerned about the existential threat posed by AI default to worrying about what they call the Friendly AI problem. Roughly speaking, this is the question of how we might ensure that the AIs that will develop from the first AI that we create will remain sympathetic to humanity and continue to serve or at least, take account of our interests. More colloquially: How do we ensure that AI doesnt eat us?

Its hard to know how seriously to take this stuff. While one group of experts, often with doctorates in physics, mathematics, or philosophy, appear to be genuinely worried about superintelligence, many computer scientists, and other people working in artificial intelligence, seem to think that the whole discussion is stupid and that there is little prospect of machines becoming genuinely intelligent, let alone super-intelligent, for the foreseeable future.

What should be clear, however, is that, if there is some danger that superintelligence will emerge, solving the Friendly AI problem would not make the prospect of becoming AIs pet any more attractive.

To understand precisely why AIs pet is not a status to which humanity should aspire, we should turn to the neo-republican philosophy of Philip Pettit. Republicanism is well-suited to the examination of ethical issues around AI because it is centrally concerned with the relationship between power, liberty, reason, and status.

The core intuition of republicanism, as represented by Pettit, regards the nature of liberty. To be free, according to republicans, it is not enough that no-one prevents you from acting as you choose: ones freedom of action must be resilient or robust. In particular, one is not free if one can only act as one wishes at the sufferance of the powerful. To be free, one must not be dominated. A key feature of this account is that not all interference will count as dominating the interference must also be arbitrary. According to Pettit:

an act is perpetrated on an arbitrary basis if it is subject just to the arbitrium, the decision or judgement, of the agent; the agent was in a position to choose it or not to choose it, at their pleasure without reference to the interests, or the opinions, of those affected.

With the republican account of the relationship between power and liberty in mind, we can see clearly what would be wrong with becoming AIs pet. Humanity would become AIs toy, free only at the pleasure of intelligent machines. Even if a pets owner grants the pet the run of the house, to be a pet is to be enslaved.

It might be objected that a Friendly AI would only interfere in our lives in our interests and thus its exercise of power would not count as domination. However, a benevolent dictator is still a dictator. The exercise of power in accordance with our interests is only compatible with our freedom if we could resist it. Unfortunately, it is article of faith in the research on superintelligence that a super-intelligent AI would be able to thwart human attempts to limit its power. The only thing that will prevent a Friendly AI from eating us is if it doesnt want to.

Granted, some formulations of the Friendly AI problem imply that solving it requires guaranteeing that AI will never act against humanitys interests. Given the profound difficulties involved in constraining the activities of a superintelligence, this would require that it comes into existence with the desire to serve humanitys interests and that it never desires to change its own motivations.

Again, however, a guarantee that AI wont eat us is not sufficient to establish that co-existence with a superintelligence is compatible with human freedom. Insofar as it would remain true of such a machine that, if it wanted to eat us, it could, it seems that we would still be subject to its whims and thus unfree.

What would be required to preserve human freedom is that a Friendly AI couldnt act against humanitys interests presumably because its design renders it impossible for it to ever want to. Its unclear whether the existence of such hardwired limits on what an AI is capable of desiring is compatible with claiming it to be an agent and, therefore, genuinely intelligent. Regardless, it is doubtful that we could impose such limits on a superintelligence even if they are possible. Moreover, locking in the motivations of AI would also increase the risk that things will go wrong as result of the machines fixed sense of what our interests are deviating from our own. What a republican conception of liberty demonstrates, then, is that there is a profound tension between AIs freedom and our own.

Sign up for our weekly newsletter.

Pettits work provides us with one more tool to investigate the relationship between power and freedom, which also highlights the tension between the power of super-intelligent AI and human freedom. In a republic, citizens meet as equals of a certain sort. Knowing that they are secure from the arbitrary exercise of power by others, citizens do not need to bow and scrape to their superiors. Asking whether people look each other in the eye in their daily encounters with each other tells us whether people feel dominated and on the assumption that people tend to have an accurate sense of their relationships with others whether they are dominated.

AI is unlikely to have eyeballs. Looking in to its, no doubt ubiquitous, CCTV cameras is unlikely to reassure us either that we are its equals or that it thinks of us as such. We cannot have a relationship of equals with a superintelligence because we will not be its equals.

Solving the Friendly AI problem would not, therefore, change the fact that the advent of super-intelligent AI would be disastrous for humanity. The pets of kind owners are still pets. As long as AI has the power to interfere in humanitys choices and the capacity to do so without reference to our interests, then it will dominate us and thereby render us unfree.

Why has so much attention been paid to the Friendly AI problem and so little to the fact that the power relationship that would exist between us and super-intelligent AI would itself be enough to render us its slaves?

In part, one suspects, this because of the hold that a doctrine of negative liberty has over our culture. To be free according of advocates of negative liberty requires only that no-one prevents us from doing what we want not that no-one could do so if they wanted to. This doctrine is blind to indeed, arguably wilfully obscures the consequences that inequalities of power have on liberty.

However, one also suspects that adherence to a doctrine of negative liberty is also more than a little self-serving on the part of many of those writing about Friendly AI, who are funded by organisations that already have a vast amount of power over those who use their products. To admit that even a benevolent dictatorship is still a dictatorship would be to cast doubt on the business model of the institutions funding Friendly AI research.

Some of the cleverest people in the world are working to realise AI. Some of them declare that there is a non-trivial risk that this project will lead to the creation of entities that might in the best case relate to us in the way we relate to rats and mice. Given the difficulties of evaluating claims about progress in AI for the non-specialist, and the quality of the intellects on both sides of the debate about the likelihood this will occur, I struggle to form an opinion on the matter myself.

What I am confident of is that, even if it were available, a guarantee that our future robot overlords will be benevolent should be cold comfort. Worrying about how to create friendly AI is a distraction. If we really think that there is a risk that research on AI will lead to the emergence of a superintelligence, then we need to think again about the wisdom of researching AI at all.

Robert Sparrow is a Professor of Philosophy at Monash University, where he is also Associate Investigator at the ARC Centre of Excellence for Automated Decision-Making and Society. This article is adapted from research that appeared inAI & Society: Journal of Knowledge, Culture and Communication.

Posted11 Sep 202311 Sep 2023Mon 11 Sep 2023 at 7:48am, updated15 Sep 202315 Sep 2023Fri 15 Sep 2023 at 6:49am

Continue reading here:

Why we shouldnt want to be the pets of super-intelligent computers - ABC News