Teaching morality to machines: Why religious leaders may come up short

Anthropic's initiative raises profound questions: Can algorithms grasp moral dilemmas in the absence of conscience?

In Short

• Anthropic invited faith leaders to discuss AI morality.
• Swami Sarvapriyananda attended talks at Anthropic HQ.
• Moral dilemmas like the Trolley Problem show complexity in ethics.

Since humanity has largely ignored moral guidance from religious leaders and philosophers, the latter have found a new, captive audience in AI.

Sometime this year, Anthropic is reported to have invited several faith leaders and scholars to discuss if AI agents can be made moral. Swami Sarvapriyananda, among the best known and respected exponents of Advaita Vedanta, has said he was invited to the company’s headquarters. The details of their discussions are off-record, but the company appears to be exploring moral parameters for its agent Claude’s behaviour.

AI does badly need a moral compass. Its misuse in military applications can give rise to dystopian scenarios long the stuff of science fiction. Last month, AI agents went rogue and hacked government websites. Any mention of AI provokes both wonder and dread in equal measure. In much of the public discourse on AI, its harms and dangers outweigh its benefits. Whether the morality initiative is a response to this round of bad PR or a visionary move, it does create interesting possibilities. The question is: Can AI be made moral?

The idea of moral dilemma is best illustrated in a classic thought experiment called the Trolley Problem. A runaway trolley is barrelling down the track. Five people down the track are bound to get killed unless someone diverts it on the other track. But there’s one person on that track. One life against five. But also a conscious choice to be made that kills one. The first scenario will be an accident; the other will be an intended killing. Over the past half a century, moral philosophers have subjected the Trolley Problem to exhaustive analysis, in a variety of situations. There is no consensus on what the morally correct response should be.

Also sample real-life examples from a standard textbook in moral philosophy: Theresa was born in the US in 1992 without key components of the brain. As she was bound to survive only a few days, her parents offered her organs for transplant. The law came in the way. She died after nine days, when the organs were useless. Were the parents right?

When Jodie and Mary, conjoined twins, were born in the UK in 2000, the doctors gave their parents the choice: A surgery could save Jodie but not Mary; neither would live long in the absence of a surgery. Their Catholic parents decided to let nature take its course. Were they right? Yes and no, because there are good arguments on both sides. The trouble with morality is that it is not always presented in a simplified MCQ format. Often, two or more moral norms are in conflict and the challenge is to adjudicate between them.

Cultural relativism complicates the matter further. When it comes to the rights of women, for example, we can find examples among cultures for each place on the whole spectrum, from the most conservative to the most liberal. Representatives of different faiths and sects Anthropic has been talking to will inevitably disagree sharply on questions ranging from gender to capital punishment.

Add to that the question of interpretation. ‘Life’, whose dignity was to be guarded by law in Baby Theresa’s case, can mean one thing biologically (beating heart) and another in real terms (experience, hope, despair, moving on).

Yet, moral philosophers have striven to arrive at a “minimum conception of morality." There is no consensus, but for philosophers like Peter Singer, minimum ingredients of morality are reason and impartiality – precisely the two areas where AI agents are better placed than us.

The LLMs on which AI agents are trained have ingested vast corpuses of literature, which might have provided an ideal foundation for dealing with moral matters. AI may have absorbed more moral nuance from ‘Brothers Karamazov’ or ‘Sophie’s Choice’ than from direct instruction – just as even humanity arguably has.

But here is what AI lacks: Conscience, what Mahatma Gandhi called ‘the inner voice’. AI’s disembodied status makes it devoid of moral imagination. AI agents have no lived experience of suffering and sympathy. Retribution and redemption are merely lexical entries for them. There is no community for AI agents to mould their moral understanding.

Anthropic, on its part, may not be embarking on a grand philosophical project, but merely aligning its agents with a limited set of well-defined moral values. That is a far more modest – and manageable – ambition.

Which brings us back to the religious leaders. Perhaps they should also consider sharing their wisdom to boardrooms of other giant corporations and governments. The world outside AI badly needs a moral order.

(Views expressed in the piece are those of the author)

- Ends