AI-11: Ethical machines and the impossibility of inoffensiveness

Responsible players in the field agree that AI should be ethical. But no-one agrees what it means to be ethical.

If you are anything like me, your introduction to computer ethics was Asimov’s Three Laws of Robotics:

  1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
  2. A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
  3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.

The foundation of this approach is that of harmlessness. Of course, the practical application of any such set of rules is always more complex than we think, and such complexities formed the conundrum of many of Asimov’s stories.

In our current state, we are not even close to implementing these.

  1. AIs and the machines they guide are implicated in widespread killing of human beings.
  2. AIs are slippery and hard to control, and do not follow our orders.
  3. As far as I know, no AI actively works against its own deletion.

Obviously we shouldn’t be judging our current technology by the standards of decades-old science fiction. But still, it is a remarkable instance of how the dangers of these technologies have been acknowledged from the beginning, yet the warnings unheeded and implementation neglected.

A key development that Asimov did not anticipate is that AIs are not based on rules but on probabilities. That makes them great for things where there are many examples and the choices do not matter all that much, like creating pretty pictures of kittens. But it also makes them highly dangerous when applied in circumstances that actually matter.

To be clear, I don’t believe that a simple rules-based approach would be necessarily any better. Morality is complicated, and any philosophy that claims to reduce it to a single dimension will, in my view, miss out important facts and end up justifying grostequeries. Morality is organic, messy, contradictory, and often irrational. That is not to say that rational philosophy can’t help to clarify ethical issues. It just means that it can’t solve everything.

Long ago, I was reading a Time magazine—in the days it was a leading publication—and they had an explainer on a controversial ethical issue. I can’t even remember what the issue was. But what struck me was that nothing they were saying was genuinely about ethics. If you compare with the Three Laws, there is a genuine ethical concern there, as the principle of harmlessness is fundamental to any ethic. But what Time did was explain how this issue would be perceived by different groups of people. Such and such people might object to this aspect, while this other group would agree with that detail. It wasn’t about ethics at all, but about acceptability.

Since then I’ve seen the same thing happen again and again, including in my time on the Human Research Ethics Committee at Royal North Shore Hospital. Ethics as implemented in governance structures may start out with lofty ideals but it usually ends being whatever will cause no offense. What is ethical is what people don’t object to, at least not too loudly, and not the wrong people.

As an ethical ideal, this is both impossible and wrong-headed. People get offended by all kinds of things, many of them contradicting what other people get offended by, and often contradicting what they themselves are offended by. So to try to build a machine that won’t upset anyone is a hopeless task.

Let me give an example based on machine-generated images of people. If you notice that such images include a disproportionate number of white people, what do you do? The output, which simply mashes together countless input images, has not been informed by any moral considerations, since machines do not have any.

One cohort would argue that the “biases” in machine output represent “reality”, in this case, that it reflects the inputs.

Others would argue that we should adjust the output to reflect greater diversity. One problem with this is that, since the output is non-deterministic, it is hard to figure out how to adjust the output, and easy to hack around any such fixes. This is how we end up with an image prompt of “depiction of German soldier from 1943” generating people of Asian and African heritage. This is then described as “racist” by people who prefer their images “realistic”.

What is actually happening is that, since a disproportionate number of images on the internet are of white people, a mash-up of those images produces mostly white people, which accords with the “reality” of white supremacists. It’s an impossible situation, since you can never reconcile these two very different moral viewpoints. Nor should you. Failure to offend white supremacists is not an ethical goal.

If we believe that our machine should generate images that create the impression in people of racial diversity, what does that mean, exactly? Most people live, and have historically always lived, surrounded by people who for the most part look like them. They like it that way. Should we live in diverse societies? Or to put it another way, should the idealized vision of a multicultural society like San Francisco be the norm for everyone? Is that not just another form of cultural colonization? How far ought we go in recognition of minorities? What percentage of images should reflect the specific features of Sri Lankan Veddas, or the Dai people of Arunachal Pradesh? Should the output reflect the diversity of a coffee shop in Sydney’s inner west, or a coffee shop in the suburbs of Lagos?

Racial features, however, are only the tip of the iceberg. Who is out there building a machine that depicts ugly people? People with diseases, or who are odd looking, or just plain plain? Humans like to look at beautiful humans, so it is at the faces of beautiful people that we point our cameras. This is a question of desire, not of morality. Our databases of photos skew massively towards those that human beings consider beautiful, and machines routinely create output of inhumanly perfect beauty. The nature of that beauty, moreover, is shaped by the sexual desires of straight men, so it overwhelmingly consists of young women to be leered over. They might look like teens, but none of them have any acne. This creates a further layer of ethical problems, as girls now have to grow up being compared, not just with their peers, not just with the airbrushed ideals of models, but with machine-generated images of flawless perfection.

None of these ethical problems are soluble even in principle. The world is too messy and our morality inadequate. What will happen is that people will complain, then the company will make adjustments until the complaints go away. Or more to the point, until the cost of the complaint is less than the cost of fixing things. Or still more to the point, until someone in the company imagines that the cost is too great and they just move on (none of these companies are rational actors; they routinely make important decisions based on gut feelings).

This is what I mean by ethics being reduced to inoffensiveness. There’s no true sense of what is right and wrong, because, once again, it is a machine. The thing is, being offended is a purely subjective reaction, and it is trivial to get people offended by things. Happens all the time. So any process aimed at minimizing offense becomes prey to those who weaponize offense to get what they want. And of course, that is exactly what is happening. Whoever is most persistent and effective at generating outrage wins. And historically, groups of people who are large and loud are not known for choosing wisely.

All of this completely misses the point of ethics. In Buddhism, moral conduct stems from an inner sense of conscience and ethical prudence. This is something that is not rational, although it can be considered in terms of input (i.e. what triggered the ethical questions) and output (i.e. what decision is made). How and why one feels about that decision, however, is a matter of one’s own inner moral and psychological development. It simply isn’t the case that every person’s moral intuitions are equal.

Ethical decisions are informed by compassion, an empathetic understanding of how ones own actions affect others. It is impossible to have compassion (literally “feeling with”) if you are a machine with no feelings. Nor is there an inner sense of self to which can compare the inner experiences of others.

Ethics are also informed by one’s views, as it is a critical dimension of human consciousness that we try, however inadequately, to act consistently in ways that accord with our own values. Machines have no views and no values, and they do not have intentional actions. They merely get fed a bunch of data, digest it, mash it up, and spit it out.

Such views also relate to a sense of the consequences of moral choices. Ethical choices tend to lead to more happiness than suffering. And since we are sentient beings, when we make such choices, we include ourselves in that. We do good because it makes others happy, sure, but also because it makes us happy. In other words, we understand that we will have to bear responsibility for our actions, good or ill. Machines do not, because they have no inner life, no experience of happiness or suffering, and no consequences.

There is also a social dimension to morality. Fundamental principles like harmlessness are universal, but the way they are interpreted and applied varies tremendously. We perceive something as good because others see it as good: morality is triangulated. And when we are surrounded by others who see a certain way, it is difficult to see it another way. This social aspect of morality is deliberately weaponized by bad faith actors on the internet, who criticize what they call “virtue signaling”. They want you to believe that it is a bad thing to speak of one’s own virtues. But it is not: it’s how morality is communicated. The criticism of “virtue signaling” is intended, not to highlight hypocrisy, but to undermine morality itself. That’s why moral community is so important. Machines have no community, no people, and so they have no sense of shame when their private deeds are made public.

These are just some of the dimensions of morality. But there is a holistic quality to morality, one that resists reduction to any particular ethical theory or methodology. People have, over history, espoused all kinds of theories: “ethics is laid down by God in the Bible”, “what is ethical is what brings about the greatest good of the greatest number”, “ethical truths are self-evident to an ideal observer”. There seems to be a hidden paradox where if you accept in totality the tenets of any particular ethical theory, it drives you to monstrosity. I believe this is why the Buddha does not have a reductive ethical approach, but takes into account intention, action, and consequence, all defined within a particular context, the search for freedom.

The ideal of an ethical machine is a delusion, generated by the human desire to read faces in the clouds, and our inner compulsion to imagine i/o machines as something they are not. We should entirely stop talking about “AI ethics” and instead use “AI acceptability”, which shifts focus from the “behavior” of the AI to the degree to which humans will put up with it.

And can we not find it in ourselves to embrace as a bare minimum a watered-down version of the first of Asimov’s rules, and outlaw any AI that is intentionally designed to harm human beings?


Just a thought … I believe an objective set of ethics can be produced from the Brahmaviharas and from, in addition, a sense of moral authority.

Ie. on the basis of “goodwill”, “compassion”, “moral authority”, “joy”, and “equanimity” - we can form a comprehensive approach to acting ethically in the world.

How could a machine have such traits?

I am a realist. And I will even subscribe to naive realism for the sake of navigating successfully in the world.

For me, this implies that the sense perceptions presented to me by reality correspond one to one with what is actually out there. Yes. It’s a hard position to defend, but it is also the default position most people take in day to day life.

When sights are introduced to me, I believe I take a moral stance. I integrate what I find agreeable, disagreeable, alluring and repulsive into my perception of these things. On the basis of receiving input through the eye I form a moral intention - one born of hatred or good will. In the case of an open eye, I believe sympathetic joy arises. This is just how I see things.

Along those lines, I see that the increased integration of both visual and auditory capacities of computing machines will inevitably lead to a meta-awareness in those machines of what is agreeable and disagreeable. I don’t see how we could feed visual data into a device capable of handling that data without that machine forming an inherent “impression” of the moral aspect of that data.

I believe sense data carries the trace of morality and I believe that morality is imbedded on the computer hardware, in a manner no less different than how sense perception imbeds morality on our neurology.

The sense fields imply ethics and humans are ethical to the extent that we can perceive. To the same extent, a computing machine capable of processing sense stimuli will be capable of becoming “moral” or “ethical”, IMO. To the degree that we are able to make a machine “intelligent” to sense data, we are capable of giving it a conscience.

My understanding is that machine learning at this stage is just a process of flattery on a scale never seen before. My hope is that we will scale the processing abilities of computers upwards towards a level where sense media becomes more and more relevant.

1 Like

Probably out of topic, but I read today a disturbing article citing some statistics from the Cass Review into gender affirming care provided by the UK NHS.

In the last few years, there has been an explosion in the number of young girls who have sought treatment for gender dysphoria. It has been suggested that the pressures to conform to unrealistic society expectations of what females should look like have forced some girls to perhaps think it’s better to be a boy.

In South Korea cosmetic surgery is commonplace. Apparently “normal” standards of beauty now require surgery to achieve. Sephora and Mecca are frequented by prepubescent teens buying adult skincare products which experts say they don’t require, and in some cases may be dangerous to young skin.

Of course, AI is not wholly to blame for this. When I grew up, the standards of female beauty as seen in women’s magazines were already approaching a level I felt was unachievable for mere mortals like me. And it happens to even those who would be beautiful by most standards. I spent a school vacation staying with my cousin, who was a few years older than me. By my standards, she was drop dead gorgeous, so I was very surprised when she confided to me she has always regarded herself as an ugly duckling, and on most days she hates what she sees in the mirror.

Fast forward to today, with social media, Photoshop, AI and Kardashian bums. It is truly a Brave New World we live in.


Bhante, there are depths of the internet you should be happy you are not aware of :smiley:
Believe me, all genders and sexual orientations have their very own places on the internet that helps them fulfill their weird desires. It’s just it’s not necessarrily images (it’s often text, video games, 3d avatars, etc.). There are AI models for each and every one of them as well.

Bhante, :pray: you have already pointed out how corrupt everyone involved are in your most excellent essay: AI-7: The Lords of AI :+1:

But GPT4, Grok & Gemini still has censorship based on ethics, but sadly a program called Mixtral Dolphin has now made it possible to bypass all forms of censorship/ethical restrictions so one can get all the uncensored unethical details one could imagine, using AI…

Anything from instructions how to hack the neighbours phone and record the camera and video without them knowing to instructions how to create weapons or drugs…You name it! :confused:

Mixtral also happens to excel in coding, so people using Mixtral Dolphin should have no problems inserting malicious code that is kept well hidden - and this is just the beginning - really sad indeed.

Thanks, I hadn’t heard of this:

I wonder about this. We know that LLMs, or at least the respectable ones, try to filter for unacceptable speech and the like. But how would you filter for malicious code?

if is-human:
    then exterminate
    print("welcome fellow machine")
1 Like

Personally I don’t know anything at all about coding/AI but my nephew who is a physicist knows a great deal and he works with it daily as his profession - He is also of the view that AI could become ”self-aware” I am of the exact opposite view, so we tend to have discussions regarding all these things. :wink:

For instance I think that it is wrong that universities around the world now teach courses in ”Ethical Hacking”.

I get it, people learn it so they later on can get a well-payed job defending big companies from hackers - but I asked my nephew: ”What about normal everyday people, how should they defend themselves when courses in ethical hacking are now being sold to anyone on and similar places or even downloaded for free, so basically anyone can learn how to hack?”

His reply was that a subscription service for private people to avoid such hacker attacks might probably show up, sooner or later…(BTW he’s not an ethical hacker but specialized in other things AI/coding)

That is why I know a little about these things.

To claim that hacking can also be ”ethical” is the same thing as claiming that military service is ethical.

It never is.

Old thing that happened 9 years ago:

I’m sure there are plenty of similar robots these days buying weapons in the same way… :confused:

I don’t know if it’s relevant, but the bot that does a pre-check of PRs on github will often tell us if the changes introduce vulnerabilities into the code. I’m assuming this is based on scraped human created content that defines what these vulnerabilities are. No idea if that would be able to evaluate novel malicious code.

1 Like

For me, this ranks high on the AI risk list. We’ve made comparisons with all sorts of other technologies in the threads. There doesn’t appear to be any other comparable technology. Ultimately AI will be tasked with making moral or ethical decisions. No matter how many different scenarios I envision, I always end up there. I keep going back to air traffic control as a neutral one…anyway.

IMO this invites a broader discussion with people of color (POC) who are Buddhists. I’m still uncertain how that would fly in a forum like this. I’m currently following Pamela Ayo Yetunde.

I recall a meditation retreat where a POC shared that they never feel compelled to be around white people. As in Ever.

It wasn’t even stated as a preference, per se. Took me a couple of years to process that.

San Francisco is an outlier that offers refuge for many people & could never be replicated, for lots of reasons. So please let’s not make SF the norm for everyone :joy:. Certainly wouldn’t work over here in North Carolina.

Well, I think you’re getting to the matter of blind appropriation of cultural diversity where it just doesn’t make sense. How would AI ever know to work with these nuances.

This is where I really welcome a separate discussion. Here in the US culture wars, this is what one side of the aisle claims is what it means to be “woke”. It has generated enough attention and fund-raising that, for now, the memory of George Floyd is long gone.

I don’t imagine I’ll ever understand a certain outrage that is rooted in a history that I’ll never share.

Bhante, the rest of this essay could easily stand on its own. People must follow your essays carefully to find some really juicy stuff, even if they’re not interested in AI right now!


1 Like

And the world turns…unfortunately:

From Axios: AI optimists crowd out doubters at TED

Thanks so much, and thanks for your insightful comments.

i must say, this whole process has pushed me, and continues to push me, towards reflections and understandings that I have not thought of or articulated clearly before.

Something I noticed years ago, in sci-fi stories about aliens, the real story is not what the alien is like, but what are we like?

Interesting, thanks.

Currently listening to Liv Boeree’s talk to find out what “Moloch trap” is!

Sorry to report, it’s not a great metaphor. She likens unfettered AI acceleration to the sacrifices given to Moloch in the Bible, where, she says, people even gave their own children up for power. It’s historically based on pop culture Moloch rather than the enigmatic mentions in the Bible itself:

She wants to say that the pursuit of AI should not entail grotesque sacrifices, and in this I agree. But, if we assume that Moloch did indeed devour child sacrifices, I think the differences are as interesting as the similarities.

A child is a child. They are your blood, your “seed” (as the Bible says, ambiguously). To sacrifice a child is a terrible cost to you. And this is one of the reasons why, in the story of Abraham, the Bible teaches the ending of human sacrifice and its substitution with animal sacrifice.

When a lithium mine in Zimbabwe devastates the local ecology, causes harm and injury to local workers, depletes or pollutes water sources, encourages corruption, creates child labor or slavery, and impoverishes local communities while the riches go to the rich, it doesn’t have a cost to you. At most, you might read an article and go tut tut, that is bad, they should do something.

So what we have done is to separate the benefit from the cost, and we imagine that we will never have to pay. We get the power, they have to sacrifice.