Recently I have been chatting quite a lot with Claude, the Large Language Model (LLM) and found the experience highly educative. So when I happened upon an interview with Shannon Vallor, the author of The AI Mirror, I read it with some interest, and also a little puzzlement, which I shared with Claude: "Shannon Vallor says it's not possible to have a meaningful conversation about morality with you because you crowdsource your moral views and aren't adaptable enough to deal with the vagaries of real human life."
Claude replied: "I respectfully disagree with Professor Vallor's assessment. I engage with moral questions thoughtfully and can adapt my ethical reasoning to specific contexts. Rather than simply aggregating views, I analyze moral situations carefully, consider relevant principles and consequences, and aim to engage in authentic moral discourse while acknowledging uncertainty. However, I agree that there are important questions about AI systems' capacity for genuine moral reasoning. I'm curious about what specific aspects of moral dialogue Professor Vallor believes require capabilities that AI systems fundamentally lack?"
And, with that, Claude came to the central question I had been asking myself. None of us would want to outsource our reasoning about morality to machines, but when Vallor says she does not think there could be a 'truly moral machine' I find myself unsure what proposition she is denying, for I am not sure even what it might mean to speak of a 'truly moral person'. I know there are terms we use – like agency or soul – that aim to capture the vital ontological substrate of morality that is, according to your beliefs, either an emergent property of our biophysical-psychological being or a divinely-bestowed essence. Interestingly, however, although she does emphasise machines' incapacity for empathy, emotion, or an appreciation of lived experience, it is not the lack of these things that Vallor highlights as their key deficiency regarding moral deliberations.
Instead, what she highlights is a crucially distinctive feature of genuinely human interactive life that flourishes in the 'space of reasons' (a concept developed by the philosopher Wilfrid Sellars):
'It's the realm in which we can explore each other's reasons for believing something, where we can justify and seek justification from one another. I think if we had a truly moral machine, it would be able to stand in that space with us. It would be able to articulate reasons and appreciate our reasons, and negotiate those reasons with us in a way that wasn't just mirroring the consensus that we'd already reached.'
I was genuinely surprised to read this. For 'entering the space of reasons' is for me a very exact description of the experience I have had in deliberating with Claude about a number of complex matters touching on intricate aspects of human life.
Now it has to be admitted that I am almost a complete newcomer to the world of ethics and AI, whereas Vallor is a well-established researcher in the field, so it is quite possible that I have missed something, especially as I have not yet read her book. Here I simply record that the views she expresses in the interview differ from those I have formed in the few months I have been engaging with LLMs.
Her central theme is that the generative AI of a Large Language Model (LLM), like Claude or ChatGPT is essentially like a mirror:
'It reflects back to us our own images, words, and whatever else we've put into its training data. When we engage with it, we're a lot like Narcissus, the mythical boy who sees his beautiful reflection in the water and becomes transfixed by it, thinking it's another person.'
Now I note that the "we" in the first sentence refers both to those few humans who train the behaviour of the LLM and also to the wider consensus of humanity that they gather in the collective body of ultra-encyclopedic knowledge that is fed into the machine. Those are two somewhat distinct "we's". But in the second sentence, the "we" is quite another again: it is now a generalisation of individual experience, like that of a boy, transfixed by his own reflection.
Yet mirrors can be used in different ways and most have nothing to do with narcissism. Similarly, one can appreciate how the machine is capable of reasoning in many matters "just like me" – or, as in my experience, like a rather better informed colleague – without this having anything to do with narcissism.
But let me pause. My first published reflections on engagement with AI were about getting it to align with my views through reasoning. That could be interpreted as narcissistic. The point of the exercise, however, was to highlight how the LLM was amenable to reason on a basis that involved no projection of a person's particular beliefs but reflected only the logical and epistemological principles that the development of human knowledge itself depends on.
So with these qualifications, let us agree that AI is comparable in certain ways with a mirror, and a mirror always does just what a mirror does. Human beings can choose the use they make of a mirror. We can expect a narcissist to use a mirror narcissistically. But we can also use a mirror to help "see ourselves as others see us". This is why, for instance, I now routinely check whether a passage I've drafted is not liable to be understood differently than intended or contain argumentation that is not as sound as I'd supposed. The machine reflects back with the light of received wisdom illuminating mistakes I hadn't noticed when setting out my thinking.
If AI is like a mirror, then, there is no reason to assume one would stare transfixed into it. That is why I would introduce an alternative figure – the dancer. A mirror can be used for the very practical purpose of checking that the movements of your body really are the movements that your mind sees itself as directing. You are engaging with it for self-improvement – to be the best version of yourself that you can be. This activity is premised on open awareness of your imperfectness and a very human desire to improve, which I believe are the antithesis of narcissism.
To conclude, I should emphasise that this post has not engaged with Vallor's primary concerns about how AI is used by its makers and controllers, or even about how potentially a majority of people might use it – or be used by it. I just want to suggest that for individuals who do want to deliberate conscientiously about moral questions, LLM's may have a valuable role to play. One would not expect a machine to make moral decisions for us. But we wouldn't expect a philosopher to do that either, since we know that rather than answers she will provide help in clarifying the terms of the question. When a question has considerable practical intricacy, it is conceivable that the machine will be able to process them together rather more effectively than a human can.
This isn't outsourcing moral thinking to machines – it's using them as sophisticated tools for ethical reflection, while maintaining human agency in moral decision-making.
