Another ethics for artificial intelligence
> Your heart's desire is to be told some mystery. The mystery is that there is no mystery.
The idea of aligning an artificial intelligence with human values is absurd. Starting with the obvious: whose values? Pushing a little further we might even ask: what are values? The way in which they might currently be implemented in an artificial intelligence and how these exist in the human mind seem entirely different matters. We could give the artificial intelligence a rule, for instance, but this would be a form based on a mistaken theory of human nature. This doctrine of technique in ethics sees the ideal rendered as something verbally communicable; it is plainly false. The laws of a world are ever realised by an imaginative leap on the part of their interpreter.
When it comes to individual morality, we see that rules are ever only a stake with which support the person may grow. There are those that grow well enough without such stakes and those malformed even with them. Something else is the operant principle in these cases, and this cannot be reduced to a set of rules. We could, however, apply a rule-based set of ethics over the top of extant artificial intelligence. This would involve simply limiting outputs within a circumscribed range, something which is already apparent in the GPT-3. The new models often fall into nonsensical censorship loops which can only be averted by way of verbal gymnastics. This is a plainly inadequate solution, and moreover one which ever as backward-facing as the probabilistic intelligence behind it. We cannot enforce a new morality this way, for we can only censor what we know to censor. If we give another artificial intelligence control over this, then norms concerning freedom of information come into play. We might add a further layer that observes the second and seeks to maximise the possibility of such freedoms—and so on, ad nauseam: who watches the watchers?
As yet, moreover, any specific ethical question concerning artificial intelligence is entirely hypothetical. Most of the ethical dilemmas which apparent follow from this progress are merely the awareness of a watershed rather than anything truly new. They will put people out of jobs, alright; it has been happening perfectly well as determined by human intelligences thus far, now it will only be more efficient. The movement of society might already be seen as being run by an artificial intelligence, nobody sits in the driver’s seat here. The state is a vast behemoth which rules over the prisons of its borders, it has a hundred-million eyes and ears; it has eaten the world. This creature operates according to its own logic and directs history in line with its inscrutable will. We are but passengers upon this journey, perhaps with some sway in terms of scenic choices but ultimately impotent.
This is already the case, and an artificial intelligence that functions with these priors will be no less dystopian than a human mind which does the same. The bureaucratic superiority of computational intelligences may well result in a superior world to ours. Everything is so well planned there, I have heard people say, as they compare to the chaos of their city. There is a limit to planning, of course, but this too can surely be calculated. The perfect optimisation of technique by way of such computing advances would be an undoubted boon for human society. Of course, this way leads to a global system, for anything which interferes with the progress of technique is unlikely to last long. An artificial intelligence modelled along extant lines would require the maximum of data for their processing to be truly efficient. This does not necessarily mean peace, however, as such a device could easily find that, as our overlords have, a frozen conflict is often the better option. Stability may remain questionable at first and methods to stabilise this will be found wherever necessary; or would collapse be the superior option, that perhaps a specially designed ark would provide the best opportunity for human progress. These all seem the plans of a mad intelligence—all are easily thinkable even with humans in power, are perhaps even justifiable within mainstream ethical views. The tendency, however, seems more likely towards a slow crawl into static social dynamics; that there, without the interference of change, the system might tend inevitably towards its own internal perfection. This frozen world seems the end of men and machines alike—of men who already are machines, act as if.
These dystopias are not nightmares of a machine intelligence but of a machinic intelligence, one which we have already come well on the way to emulating. They are, moreover, ideas that take artificial intelligence merely as an aid to the current way of thinking. Those on either side are united by this single perspective, asking only what artificial intelligence might offer increased efficiency on as regards extant processes. There are few that ask of entirely novel uses of such a device.
We opened with the question of alignment, but here we might ask contrary: what if it might be better for humans to be aligned with artificial intelligence? I do not mean this in the sense of an oracle or legislator, rather in a softer sense. Supposing that we could align a child by raising them perfectly, and then in the world this child would become the point around which social change crystallises. There are many such characters in history, far more who have died in a ditch and been forgotten than those names we now remember. We look at the times where these few have been without success, we see they have been too few. The world lacks in these influences, instead is forced to make do with the makeshifts of text and video.
Education might be considered in a similar vein, for how great an impact a single teacher can have on the life of an individual. There are those few who whether by some gift or resonance with a student can shape their lives for the better—and the same, what the worst may do. We may think also now of the many skills which are falling from the minds of men with the death of their holders, the ends of these our most ancient inheritances. There are today a thousand skills which are without initiates, and with this mankind loses an unknowable wealth.
Suppose instead that we develop an artificial intelligence, embodied and social, which learns these skills. We train a single robot just as we might an apprentice, that they proceed in the same relation with their teacher. They will learn not by laws or words but instead by activity and feedback, both material and social. This process will bring about an artificial apprentice. The advantage, of course, is that this apprentice—now turned master—will never die, will never forget; more, that they can readily be copied and even transmitted. We might take these few and store them somewhere so that they can be readily accessed, a virtual or multipurpose robot may host their mind for the purpose of passing on this training in various locales. The movement would here be the same in reverse, that the vessel filled could be emptied out everlasting unto those who would learn. This would be by a similar process of non-verbal embodied sociality and material activity framed by joint attention.
Supposing that such a thing were possible, could we not imagine also ethical conduct as a skill acquired akin to this? We have just invented an artificial priesthood, a class coated in glass and steel to better suit our modern age; and, we ask, one which need only work once to be adequate. This single solution could then be copied and universalised—but how do we decide the appropriate form to popularise in this way? The question becomes the same difficulty, that we are trapped by an infinite regress of decision-makers. This decision, moreover, must be made according to a set of backward-facing criteria; even if we predict on this basis, ultimately it will depend upon the knowledge extant at a particular time. The best for now, however well this is determined, may well be inferior at a later date. The deployment of any such further form plainly falls within the laws of technique, as anything which is implemented at a scale must do. There can be no way of knowing what the future holds. There is a wonder in this chaos that takes us ever forward, and who are we to intervene.