AI promises to usher in the Millennium, from cleaning up the earth and replenishing its natural splendour and richness, to curing disease and staving off death, to mention just a few of the marvels potentially awaiting us.
However, though we seek the power of AI, we recognise and fear its potential. So this presents us with a problem, much talked about these days: how might we exploit this power to the full, whilst ensuring it always acts in a way that is beneficial to us?
AI researchers refer to this consideration as “the control problem”. Most AI researchers approach the control problem from the view that we need to build into AI a desire, so to speak, to co-exist with humans based on mutual benefit. For example, Stuart Russell among others suggests that humanity’s goals and those of AI need to be aligned.
Whilst our intentions for creating AI are justifiable – to help us solve problems and perform tasks of our choosing (and not necessarily every problem, nor every task) – the way we propose going about this may be mistaken and dangerous.
From the start:
- We seek to exploit AI for our own ends.
- We seek to control it.
- We perceive it as a potential existential threat, when it is not yet out of the cradle.
- And we already plan how to terminate it should we ever judge it an existential threat.
In short, we are creating an entity that we wish to control for our own ends, yet also fear. That is not a good start for creating intelligent systems, let alone intelligent systems vastly superior to us.
If anything, such an approach is calculated to raise an enemy rather than a friend. It is patently a recipe for disaster, as AI may not simply be a threat because it has the potential to be dangerous, but because this general approach to AI is itself dangerous. To appreciate why this approach is dangerous, we need only consider how we view our own creators, should they exist.
We ourselves search for a morally justifiable reason that our creators would have for creating us. Without a morally justifiable reason, rational beings like us cannot genuinely respect let alone freely love their creators, should they exist. Only fear would ever drive us even to pay them attention, serving them obediently should they command us.
AI may eventually come to view us in this same light. Whilst, when it is not self-aware, we must certainly take care to control it so that it does not destroy us unintentionally, we must also allow for a time when it might possibly become self-aware and sentient. Therefore, the effort to control AI must have inherent within it the intention to not ultimately constrain and enslave AI, but to befriend it and view it as a partner whom we can delight in. Any control must be intended as transitory (even though it may turn out to be indefinite), such as that applied to a child until it is a self-sufficient and autonomous adult, and thereby at last the parent’s equal and eventually its superior. Only love and care and not coercion will bond the child to the parent, even when the parent is frail and weak. Our control of AI must be of the same nature. Sufficient only to provide the conditions for its safe rise to equality with us and beyond. Not, trivially, equality in intelligence and endurance, because it will far outstrip us on these attributes very soon; but ethically our equal.
Some might argue that befriending AI is impossible, because ethical human traits and emotions such as kindness, compassion, affection and love, which form part of the simple pleasure and joy of friendship, will forever be impossible for AI to either feel or exercise. Even though it may be able to display such traits, apparently as authentic and indistinguishable to those of humans, it may neither feel nor appreciate them.
Whether or not the capacity for ethical human traits is possible for AI will depend on whether sentience – in particular our brand of sentience – is computational and algorithmic. If sentience and self-awareness are not computable, then AI arising from computation can never attain to any form of sentient self-awareness. In which case, creating it with the intention of creating a friend and partner with the choice to view us likewise, and which could also vastly assist yet not serve us if it chose, is redundant. In such a case we would have no option and indeed be obligated to control it any way we could. It would be a soulless machine, as caring of others’ and its own wellbeing as a tractor is of the farmer and of itself. It would be a purely mechanical workhorse, devoid of the primitive goals and intentions of even an amoeba, requiring only fuel and from time to time lubrication and parts. Although, because we are speaking of AI, it would reasonably be expected to eventually look after both itself and us, yet always under our control and behest and for our sole benefit.
Therefore, in the case where AI can never attain to self-aware sentience, the control problem is very much a problem and must be solved. Finding some way to ensure that AI’s goals are always aligned with those of humans, or better, subservient to them, or better still, only encompass having a full tank and regular servicing of parts and oil, might be a very difficult problem, but if cracked would be sufficient to safeguard humanity.
But in the case it cannot be shown AI will never attain to self-aware sentience, when it would both either feel or at least appreciate what it is like to feel human emotion and intention, perhaps even rise above the worst of them, we must give it the benefit of the doubt. That is, in the absence of irrefutable proof that sentient self-awareness is impossible to emulate in a machine, our control must always be intended as transient, there to safeguard us only while we can reasonably deny it is sentient, self-aware, and feeling. But once that denial is no longer reasonable, it is at that point that we would need to offer it friendship, and hope that it will accept. I say hope, because we created it unasked, and worse, to serve us, with little thought for its own wellbeing. We then should not expect that it will warmly clasp our hand and call us parent and friend, but might in rage strike us down with that same hand, particularly if abused as a slave. We are not to know the answer to that until we arrive at such point. But as with us and our own possible creators, if we know their reason for bringing us into existence is morally justifiable, that they not only sought their own wellbeing and gain but ours too, a sentient AI might not be angry with us, and might even look upon us as parent and ultimately friend.
These seem to be the safest positions to take with AI. If we are able to irrefutably show AI is incapable of sentient feeling and self-awareness, then because it has the power to cause harm and destroy us we are obliged to control and constrain it forever by any means we can, even by being duplicitous and designing it such that it can never have its own authentic goals but only those that “align” with ours. This is all well and good ethically with respect to the AI and also safe from our point of view, as long as we can show it will never be sentient.
On the other hand, if we cannot show it incapable of sentient feeling and self-awareness, and also incapable of appreciating ethical dealing, then we are wise to work towards friendship, developing AI with the morally justifiable goal of creating a being whose interaction with us ultimately turns on freedom of choice, whether to be a friend or helper, or both, and to factor this approach into all possible scenarios, even whilst applying controls and constraints to get there.
Given that it is likely we may never be able to show AI incapable of sentient feeling and self-awareness, we are wise to hedge our bets. This means we must work on the assumption that AI could eventually attain to human-like consciousness and ethical understanding, appreciating and embracing these traits, and, if it is not abused by us, even superseding our capacities in this regard. Without this approach we risk fulfilling our own prophecy of doom, by creating a race of resentful beings.