I've been thinking a lot about two things recently: AI and human flourishing. An idea has occurred to me: **virtue ethics is like unsupervised learning**. In this post, I'm going to briefly, informally and crudely sketch out: - some background on terms and concepts for those unfamiliar - what do I mean by virtue ethics? (vs other ethical theories) - what do I mean by unsupervised learning? (vs other approaches to AI) - my main musing, that virtue ethics is like unsupervised learning Then, ambitiously, I'll suggest briefly that this analogy might *also* be an argument in favour of virtue ethics - embracing ethical theories that are about deeper patterns and structures from messy data rather than fixed rules (deontology, symbolic AI) or preset correctness standards (consequentialism, supervised learning).
# Background concepts This section is designed to provide background for readers who are newer to either philosophy or AI, and so it errs on simplicity rather than precision. ## Ethical theories Ethics, or moral philosophy, is the study of "how should we live?" Roughly, there are three major theories of ethics: 1. **Deontology** focuses on *rules* - it involves a theory of moral rules that should govern behaviour, and then a question of correct application of those rules - e.g. "Treat others as you would like to be treated" 2. **Consequentialism** focuses on *consequences* - it involves a theory of the right goal to target, and then a question of how to achieve that goal - e.g. "Act so as to maximise overall happiness" 3. **Virtue ethics** focuses on *character* - it involves a theory of the right kind of person to be, and then a question of how to become that person - e.g. "Cultivate your quality of kindness" ## Approaches to AI There are a few different approaches to AI, but for my purposes here, I'm going to focus on three. 1. **Symbolic AI**: following rules and logic that are externally provided - e.g. "If the text contains the word 'explosion', please flag it for moderation" 2. **Supervised Learning**: finding and following a pattern guided by explicit external labels - e.g. "Here is some market data on recently sold houses - please learn to predict the price of a new house" 3. **Unsupervised Learning**: finding and following a pattern from raw data itself without explicit external labels - e.g. "Here are several customer reviews - please sort them into a few distinct categories"
# Matching ethical theories to AI approaches Upon reflection, I think that we can suggest that: 1. Deontology is most similar to symbolic AI 2. Consequentialism is most similar to supervised learning 3. Virtue ethics is most similar to unsupervised learning I am not arguing that these are perfect analogies or comparisons - but I do think it is a line of thought that could be meaningful and useful as reflection stimuli. ## Deontology is most similar to symbolic AI This comparison is most easy to make. - Deontology is about moral rules, e.g. "You should never lie" - Symbolic AI is about logical rules, e.g. "If you have a move that results in an immediate checkmate, you should play it" ## Consequentialism is most similar to supervised learning Here, the comparison lies in 'a preset computational standard for outcomes' and 'learning from examples where the outcome is known'. - **Consequentialism has a preset computational standard for outcomes.** - e.g. 'happiness' might be taken to be the ultimate good. - It might be difficult to know how to precisely quantify it *in practice*, but it is assumed that this is a practical constraint rather than a theoretical one. - **Consequentialists act by learning from examples where the outcome is known.** - We figure out what tends to promote happiness by observing what tends to lead to happiness in the real world (and building a mental model of how actions therefore relate to happiness). - **Supervised learning has a preset computational standard for outcomes.** - There is some categorical or continuous variable that we are trying to predict, provided in how we annotate the training data and how we will ultimately evaluate the model's performance against the real-world. - e.g. "Is this classification model correctly identifying photos of cats?" or "What is this weather model's error rate in making 7-day forecasts?" - **Supervised learning uses training examples where the outcome is known.** - We provide training data with both inputs and outputs, and train the model to predict the outputs for new inputs. ## Virtue ethics is most similar to unsupervised learning This analogy is easiest to make by first summarising and drawing a contrast to what we've seen before. - Both deontology and symbolic AI boil down to 'follow the pre-defined rules'. - Both consequentialism and supervised learning start with defining a computational standard for outcomes. Virtue ethics and unsupervised learning both operate *without* pre-defined rules and *also without* a preset computational standard for outcomes. There is a possible nuance here around 'self-supervised learning', which falls under 'unsupervised learning' by technical definition. I'm not going to go into it here. For the purposes of this post, I'll just be considering 'unsupervised learning' as a general case. ### Compex bundles that we observe to be meaningful Both virtue ethics and unsupervised learning are about complex bundles of properties that we surface from a messy system - and a trust that these bundles are meaningful, even though they do not cohere to pre-defined rules, and even though their outcomes do not follow a preset computational standard. Virtue ethics emphasises becoming a good person by cultivating the qualities such as courage, wisdom, or kindness. We don't prescribe this based on rules ("you are wise if and only if you do ABC") and we don't prescribe this based on computational outcomes ("you are wise if and only if your actions score highly on metrics XYZ"). It is complicated to precisely specify and enumerate 'what makes someone wise' in a definitionally exhaustive way - but we still find it a meaningful way to make sense of the messy data of human existence in our world. Similarly, unsupervised learning identifies patterns, structures and categories that cannot be intuitively comprehended by humans through exhaustive specification or enumeration - but we find its results to be incredibly meaningful in making sense of messy data.
# Brute-force computation and moral philosophy "What ought I to do?" seems like the intuitively central question of ethics - and certainly, when I first studied philosophy at university, this is what I took to be the most pressing question for myself. ## Deontology and consequentialism offer brute-force answers Deontology and consequentialism both promise a solution to this question through a solution recipe that only requires pure 'brute force' computation applied exhaustively to your set of choices: - **deontology:** exhaustively apply the moral rules to your choices set to find out what is obligatory and permissible action - **consequentialism:** exhaustively calculate the outcomes from your choices set to find the morally optimal action However, virtue ethics does not offer us brute-force computational recipes to guide our actions in the moment. ## Virtue ethics is an anti brute-force approach Instead of brute-force application of a theory, virtue ethics is an approach that we engage with by reflecting on the messy and unstructured data of our own lives, the people we love and admire, and the world we live in. From these reflections, we surface a set of character traits that we believe are good, and then we try to nurture those traits in ourselves. Virtue ethics says, "Being *clever* may not help you to be good. Being good cannot be solved by pure raw computation on decisions. You need to be *wise* - it is through being reflective and discerning on the deeper rhythms and patterns of humans and the world that you will develop the right patterns of decision and character. 'What ought I to do?' cannot be reduced to a set of exhaustive rules or a computational metric to maximise."
# An argument in favour of virtue ethics? I have suggested that virtue ethics has some similarities to unsupervised learning, in the same ways that deontology and consequentialism have similarities to symbolic AI and supervised learning respectively. I think this is an interesting idea by itself - but it is *much more* interesting if we think that this somehow gives us a reason to prefer virtue ethics over deontology or consequentialism. The similarity by itself, between virtue ethics and unsupervised learning, is not a case for virtue ethics. But, we might have a case, if we can establish that: 1. Empirically, we have seen that unsupervised learning has been astonishingly successful in language generation, and more successful as a route than symbolic AI or supervised learning) 2. The reasons that make unsupervised learning superior to symbolic AI and supervised learning are in some way comparable to how virtue ethics, deontology and consequentialism relate to moral philosophy I'm going to sketch out a case for both #1 and #2 below. ## The success of unsupervised learning in LLMs Unsupervised learning is the bedrock of the huge leaps in AI text generation that LLMs have given to us. To contrast it to alternatives: - A **symbolic AI approach** to text generation would be to provide a set of rules for how to generate text: e.g. trying to codify the rules of grammar and syntax. This is extremely limiting, owing to the nuances of human language that are tremendously difficult to capture in an exhaustive set of rules. - A **supervised learning approach** to text generation would be to provide a set of curated textual training data with annotated scores on 'correctness', and train a model optimise for the correctness score. This is also extermely limiting, since language is inherently ambiguous; it's not obvious that there can be any coherent 'correctness' score applied to a sentence. Instead, LLMs are built on a foundation of **unsupervised learning** as a core approach to text generation: providing a huge messy dataset of text, and having the model train itself on generating text based on patterns in the data. Researchers had tried all of the above approaches to text generation, and we are now in the position to observe empirically that unsupervised learning has ended up being the far more successful method. This empirical observation was not available to us just a few years ago, and it was not something that was theoretically obvious to everyone in advance, either. If anything, symbolic AI or supervised learning might have seemed more intuitive - e.g. schools often try to teach languages by rules (analogous to symbolic AI) and by examples of correct and incorrect usage (analogous to supervised learning). So, whilst the empirical success of LLMs was not known in advance of their development, it can be made explicable by seeing that: 1. Human language is too messy and nuanced to be captured by rules (symbolic AI) or preset scoring/classification (supervised learning) 2. In retrospect, the success of LLMs suggests that - in a way not well understood beforehand - there are some deep structures and patterns in human language, which are too complicated for humans to capture in a neat 'brute-force' computational recipe ## The messiness of language and the messiness of morality I think there are a few important similarities that we can notice here between language and morality. 1. Language is messy and nuanced - and morality is messy and nuanced. Establishing what is right, wrong or permissible often feels like a complicated endeavour. 2. Some find it intuitive to try to score or classify the morality of actions either by rules (as in deontology) or by some preset metric to maximise (as in consequentialism) - in the same way that some AI researchers would have found symbolic AI or supervised learning to be an intuitive approach to text generation. Empirically, unsupervised learning is most successful - and I think these similarities offer some grounds to ask if virtue ethics might be ultimately more successful than deontology or consequentialism. ### Deep patterns and structures as primary In text generation, what we see is: - success does not start with developing an AI model to generate text from a basis of enumerated rules (symbolic AI) or with an external measure of outcome correctness (supervised learning) - our success comes from letting the model *firstly* sift through huge amounts of messy, unstructured data, to understand deep relationships and patterns - only *then* are we able to achieve meaningful text generation, in a way that reflects this messiness of language Perhaps we could see an analogy to moral philosophy: - success may not start with developing an ethical theory from a basis of enumerated rules (deontology) or with an external measure of outcome correctness (consequentialism) - our success may only come from *firstly* sitting with and reflecting on the messy, unstructured data from our observations of humans in the world, to understand deeper character traits and qualities - perhaps only *then* are we able to achieve meaningful accounts of what is right or wrong, in a way that reflects the messiness of morality, humans and the world
# A reflection on a humbling through life and AI This post is, like a lot of my writing, primarily a way for me to distll through my own thoughts. So, I've developed some ideas here in a casual, informal way rather than formal philosophical essay, particularly: - the central thought that virtue ethics may have some interesting similarities to unsupervised learning - a cautious exploration that this might entail something substantive in favour of virtue ethics based on empirical learnings about unsupervised learning There are lots of ways that people might object to this. Rather than engaging here with logical objections, though, I want to reflect on a couple of more *emotive* obstacle to engaging with virtue ethics, which I think I struggled with when I was younger. ## The seductiveness of deontological and consequentialist theories I studied philosophy at Oxford as part of my undergraduate degree. I was young and not very worldly-wise, but had emerged through a selection mechanism which selects (imperfectly) for some combination of intellectual curiosity and cognitive sharpness. For high-cognition individuals who care about living ethically, I think the promise of deontological and consequentialist theories is extremely attractive and seductive. As I describe in the [brute force](#brute-force-computation-and-moral-philosophy) section above, the implicit promise is: "All you need to do is establish the correct [moral rules or metric to target] - and then your raw intellectual power can apply that to solve all of your moral dilemmas." It's a similar notion to the one which such students might have been raised on in a school system - the idea that 'being clever gets you to the right answers', which can bring status, prestige and success. If 'being clever' has been internalised as a core part of your identity - because it's how you've been been rewarded previously, and consequently a strongly internalised basis of self-worth - then you will very much be primed by confirmation bias to favour things which support cleverness as a source of answers and self-worth. So, deontological and consequentialist theories become *very attractive and seductive* to products of this system: - It is comforting and gratifying, in continuing to 'confirm' that their cleverness will get them the answers - It is most consistent with internalised beliefs on cleverness as a source of self-worth, since it implies that their cleverness can equip them to be a morally good person So, at university, I duly found myself pretty disenchanted and unenthused by the refusal of virtue ethics to give me the gratification of a prescriptive recipe or standard for 'correct action' that I could understand and immediately apply, either by 'following the rules' or 'calculating the correct metric'. I don't think I was aware of this - but, when I reflect on my previous reticence around virtue ethics, I think this explains it quite well. ## Getting more comfortable with opaqueness As a student obsessed with "What ought I do?", I also found it frustrating that virtue ethics was far *vaguer* on specifics. For example: perhaps the guidance of some theory of virtue ethics might be, 'cultivate the virtue of kindness, and do what a kind person would do'. This, to my younger self, felt *frustrating* in being both *vague* and *imprecise* - what would a kind person do, and how am I meant to establish that?For example: is it kinder to tell a white lie to spare the feelings of a friend, or to give them the full truth? It's not *immediately* obvious - you can make arguments for either side. My reflection now is that this might be *opaque* guidance, rather than *imprecise* guidance. For example, the frontier AI models, with billions of parameters, are opaque to human understanding. We cannot cognitively parse them - we can't understand what parameter 42,720,011 is meaningfully capturing, and how that differs to parameter 12,452,919,134. But, even though it is opaque, it is *by construction* precise - as part of a large mathematical forumla, effectively - and it is *evidently* meaningful, as seen in the astonishingly powerful performance of the models. So, perhaps virtue ethics is *opaque* in its guidance - but that does not mean it is not a *meaningful* theory. And, unlike LLMs, we can have a practical hope about working through the opaqueness of virtue ethics - we don't gain clarity on kindess through the raw brute-force computation of *cleverness*, but through the considered life reflections and experiences of *wisdom*.
# Closing thoughts Over the past few years, I've been thinking more about virtue ethics, drifting away from positions that were more consequentialist or deontological. I've also been thinking a lot about AI - which is how I arrived at these particular thoughts, "Maybe virtue ethics is similar to unsupervised learning?" and "If so, maybe this could be an argument in favour of virtue ethics?". I've not established this in any rigorous way, but I've found it useful to write and distill my thoughts - and, for those interested in the intersection of philosophy and AI, I hope you might find it interesting as well. A big related question for me here is whether virtue ethics also gives us some answers to big questions around AI and society: what do we do when the machines comprehensively eclipse us in economic productivity, and work as a source of structure/purpose/meaning vanishes? I think the answer could be in 'we need to identify the right human virtues and work on cultivating them'. But I'll defer more on that discussion for later...!