There’s No Such Thing As ‘Ethical A.I.’?
onezero.medium.com
onezero.medium.com
What we don't have, or need (yet) is "ethics built into AI algorithms" or "ethics derived from AI" - we just not don't have an AI system that's advanced enough to require it. Self-driving cars might be the first to raise these kinds of questions, but I would still be reluctant to call that "ethics".
I think it should, but in its current state, it doesn’t really.
In "Life 3.0" and his recent lectures, Max Tegmark repeatedly stresses the pursuit of safe AI requires that we articulate the world we want to live in.
In his concluding crie du coeur, Tom Chatfield writes,
> give me the capacity to contest passionately the applications and priorities of superhuman systems, and the masters they serve
If Tom or anyone else wants a seat at this table -- to participate in developing and defining ethical AI -- by all means describe the world you want to live in.
Until then, it's just tantrum-throwing.
But there’s really no way to debate with “hey, it’s impossible, you can’t do it!” Someone still has to decide what AI systems are going to do, and as they get more complex that’s going to implicate ethics more and more.
The only "safe AI" that I can think of is one that has a healthy dose of intellectual humility built into it, and can resist the urge to map-reduce conflicting expectations into a single solution.
What parallel universe is this? From the NYT to Medium the entire opinion writing industry seems to run on pieces that scrutinise the ethics of tech.
Whereas an AI is a person of some sort (when speaking of generic AI, but also at the limited level of decision making of a less capable AI).
So one could legitimately ask whether there's an ethical AI.
I think what the article is saying (as compared to the title) is that a universally agreed upon as ethical AI cannot exist. But this seems obvious given that a universally agreed upon set of ethics does not exist.
>Depending upon your priorities, your ethical views will inevitably be incompatible with those of some other people in a manner no amount of reasoning will resolve.
If we say there is no such thing as an ethical A.I. because we cannot all agree upon a standard, then it seems we should also say there is no such thing as an ethical human.
Commonly expressed opinion, and for a reason of course. But I have come to think that rather there is pretty much no such thing as 'morally neutral' technology. Everything lends itself more or less well to different purposes, and can, and often will, interact with an immensely complex society rife with morally loaded choices and actions. There was an eloquently presented thought experiment I heard many years ago, for which I have unfortunately forgotten the exact context, but I suspect it may be from former US Secretary of Defense, Robert McNamara. (If anyone recognizes this one, I'd be grateful if you refresh my memory)
Play with the thought that someone comes to you and presents you a ground-breaking invention, namely a procedure to manufacture dirt cheap, compact atomic bombs. Nuclear weaponry needs no longer be beyond the means of even the humblest citizen. Well, that would be a stunning scientific and technological achievement. We should probably give this person the at the very least the Nobel Prize and an annual feast in their honor. Is that what would happen? No, more like it, we'd make sure we have all copies of the schematics and then lock him up in isolation and throw away the key.
It's an extreme case, naturally, but I think it's a neat illustration that there are powers whose properties have a profile which give them a completely predictably huge and horrifying potential for abuse and great difficulty to use for good. And conversely, arguably those which have very readily apparent beneficial uses and would be much harder to use for nefarious purposes, though that direction may be tempered somewhat by the fact that there are many more ways for something to be broken than to be whole.
It's really important to separate the non-intelligent and over-hyped "AI" tools of today and the near future - which it seems reasonable to expect an ethically guided path and constraints - from the human-sentience AGI equivalents that can only be programmed to be ethical by hypothetical coded laws and "inserting emotion chips" in naive fantasy.
As of now, where A.I. (specifically ML/DL) is mostly used to solve isolated problems the question of ethics is more about the use of the technology. Technology itself is ethically neutral and the ethical framework is provided by the people that use it (and arguably implement it).
I also feel like ethics if A.I. posts/books etc. generally focus too much the bad use case and not enough on the good use case. One can find a lot about the potential issues of autonomous vehicles killing someone in a crash or A.I. surgeons cutting an artery but the other side of the coin is usually under-investigated. How ethical is it to limit autonomous vehicles if they are statistically more safe than human drivers. How ethical is it to put regulations on the use of automated ML in medicine to scan for cancer and require a human to sign off if the human is statistically more likely to make a mistake.
Not trying to argue either side but I think the "how ethical is it to not use A.I." case is under-argued.
Those are two absolutely gargantuan “if”s.
I’d like to see more discussion about whether it’s appropriate to project the idiosyncrasies of our imperfect, competitively-evolved biological brains— like resentment and anger— onto hypothetical thinking machines.
A lack of public efforts seems to be the equilibrium we're at, because any public effort is definitely going to attract a bunch of essays like this. Any partial attempt that looks like "we're trying to be ethical" is going to get scrutinized hard.
Continuing to put out more hypothetical theoretical work ("we prove this result about this algorithm, and it achieves a well-defined fairness goal on this common dataset") while cautioning that actual deployments should be put off until we understand everything better ("the fairness goal we define is probably not perfect, we are not saying you should use it") is much safer.
What I'm really interested in is if you can use formal specifications to constrict AI systems - which is what made me start my search. But even widening it considerably, I still didn't find anything tangible. I only looked for an hour or two mind you.
Why would that fundamentally change if we involve tech in that?
Technology first and foremost helps those with the resources to use it. That tends to be the people with power in the first place.
In our list of things that have happened in a jail cell without a camera present so far, we have someone with broken ribs, and someone who was kicked repeatedly in the back of the head against a concrete floor after a failed attempt to break their arm, as well as more minor attacks.
We have committed no crimes; we are legal, peaceful residents who have done nothing except for make a property management company go through a legal process to evict us, which we believe we will win. Nobody residing in our house, nor our visitors, have hurt anybody - we have taken part in no activity which resulted in anybody being hurt, and there's no reason for this violence against us other than a property management company which is upset at likely losing a multi-million euro redevelopment contract by failing to evict us.
I'm honestly really glad that this isn't something within your worldview; I wouldn't wish this on anybody. But it is happening.
Either way, the quote was about different political ideologies and who will get it their way, i.e. big government, small government etc. It seems that it can be figured out in a democratic process and people aren't rioting. From a historical perspective there is stability and peace even in the US. No civil wars or large scale revolutions or chaos.
The point is that people disagreeing with ethics doesn't mean the constitution or the law stops existing, it just means they ignore it. The existence of a constitution or a law usually just means that some people who our schools teach us are important decided to write something.
I would disagree and say that laws and constitutional amendments are an aggregation of decisions passed by various majorities and coalitions of minorities during several decades.
If you ask any current party, even when a majority, if they agree with all the laws that currently exist they would answer a resounding "no".
So that means that there doesn't exist a legal common ground, just a ground we currently stand on that various parties want to change but don't have the votes, time or alternatives for.
The utilitarianism approach still depends on how you define utility and who you define it for. People, animals, the planet, consciousness are different things you can define it for.
Another issue is the scale of time you are measuring things. It's not clear whether our approach is sustainable.
In an analogy to censorship, consider a world where “prior restraint” and “chilling effects” apply to every aspect in our lives, where deviations from the ideal are reliably detected and rejected, where we would not be able to earn credentials, get a job, find a partner, buy a home, evade surveillance, obtain necessary health care, or safely express a political opinion unless we either conform to or cheat the prevailing machine inference gatekeepers. What an insane additional cognitive burden this will be for everyone, everywhere, all the time.
This kind of AI is a mechanized, high-speed race to the bottom. And it will operate to discover undreamed-of bottoms to race to.
In the name of a few extra percent of efficiency we will doom our children to a rigid and lifeless dance with pattern matching algorithms.
So the idea that (e.g.) libertarians have a free choice about what they believe is - ironically - not even close to being true.
The second flaw with this argument is that all of these positions could, in principle, be tested objectively.
The best possible outcome for true general AI - also the least likely in practice, but not impossible in principle - is that it could provide definitive answers to matter-of-opinion ethical questions.
And it could do this not just in terms of incomprehensible statistics, but as comprehensible narratives derived from explicit arguments.
In fact a hyper-AGI would know which arguments to use to be most persuasive and convincing - which isn't particularly reassuring, but it would be an inevitable outcome of hyper-AGI.
What is the objective measure by which you would test them; against what standard? And by what standard is that standard the best standard to use? By what measure in turn is that standard the best measure to use? It's impossible to find a truly objective standard due to https://en.wikipedia.org/wiki/Regress_argument; a standard can only exist if defined circularly or upon some arbitrarily chosen axioms. A superintelligent AI can't somehow wave away the laws of logic, any more than it could prove that 1+1=3.
There are many interesting details and problems in this research area and there are plenty of conferences about it every year, but these problems are ultimately solvable. In any case, the content of those normative systems comes from humans (and human authorities/institutions), just like there are also laws, traffic regulations, and social norms in every country. Just like laws and regulations are not perfect and do not represent a morality carved in stone, A.I. systems will have revisable human-made contents that represent a more (or less) reasonable&lawful consensus.
Do you have any particular ethics in mind? There is a vast range of positions, almost any position that can coherently be defended has been defended by one author or another. Some of them are by definition not culturally biased, for example many forms of utilitarianism, whereas others are strongly emphasizing that ethics requires a great deal of contextual understanding, for example Dancy's particularism.
I just wonder why you address these particular points, because it's more common to hear the opposite criticisms, that ethics is unable to provide any sensible guideline because there is too much persistent disagreement between ethicists (Mackie's error theory), or because contemporary ethics is too relativist.
Instead, people are preparing to role this junk out at a scale that it's impossible to give it human supervision, and will then use that scale as an excuse to not give it human supervision.