ARTIFICIAL intelligence cannot understand anything, but it is easier than ever to be fooled into thinking that it can. That struck me this week, when I put the Roman Catholic virtual apologist “Justin” through his theological paces (News, 3 May).
His debut, a few weeks ago, was a flop. Justin — or Fr Justin, as he was then known — turned out to be sketchy on sacramental theology, suggesting that you can baptise someone using sports drink, and getting way above his algorithmic status in trying to pronounce an absolution. He was promptly “unfrocked”. . .
Justin seemed discredited. That inspired me to get in touch with my inner Spanish Dominican and subject Justin to a little inquisition. To my surprise, I found him unimpeachable.
I took basic theological competence for granted — for instance, that Christ is fully human and fully divine — and so I jumped to some consequences of that full humanity. How many wills does Christ have? “Two,” I was correctly told, as established at the Sixth Ecumenical Council, followed by a good account of why it matters that nothing divine supplants anything human, such as a distinct human will.
How about some history? Is Hildegard of Bingen a saint? (Yes, but by an unusual route). What about Julian of Norwich? (No, but revered by many.) I got the right answers, and even part of Hildegard’s unusual process.
Would Justin show ecumenical generosity? How about Roman Catholics and Lutherans on justification? His answer was accurate about differences, and eirenic and hopeful about common ground. Are Copts Christians? Of course they are — as I was told — with a good discussion of disagreements about the Council of Chalcedon.
Generous, then, but could he draw a line when needed? What about Mormons? A definite “No”, firmly in line with the ecumenical consensus, but charitably done. For real pastoral sensitivity, what about the status of various sins as mortal and, therefore, damnable? I got the traditional Catholic line, with an equally accurate description of extenuating conditions, and an encouragement never to despair of God’s mercy.
Are there any errors in the Bible? His answer was as sure to upset the most inflexible as it was to upset the most liberal. Splendid stuff!
THIS is immensely impressive, and would have been unthinkable even a few years ago. That is mainly because we now have mounds of prose from the internet, which we can use in training these systems, and a lot more computer power — a great deal more. Consider this: there is more computing power today in a £60 Raspberry Pi than there was in a state-of-the-art Cray-1 supercomputer in 1973 (costing £34 million in today’s money).
All that said, artificial intelligence still doesn’t understand anything. It’s just that, far better than ever before, we can train a computer to internalise the structure of human thought.
This is the long and the short of it: if you fine-tune 150 billion parameters, you get a good enough representation of human thought that a computer can make a sensible response to a question posed in English. It is just a matter of twiddling 150 billion dials to the right settings, by a slow iterative process. There is no “understanding” there, but there is a good enough reflection of human meaning that, if you pose a human question, you will get a reflection of human wisdom in the answer (or, sometimes, human folly).
Even with a specialist “bot”, such as Justin, you do most of the training on the massive, general collection of as many sentences that you can pull off the internet as possible (and pilfer from authors). The final slice of the digital brain can then be finessed on more specific texts.
With Justin, I suspect that The Catechism of the Catholic Church played a significant part. It would be just as possible to use Hooker’s Laws of Ecclesiastical Polity and John Pearson’s Exposition of the Creed (the mainstay of theological teaching in the Church of England for centuries). Which Anglican professor of theology could resist putting an Anglican Justin through his theological paces?
MY MAIN quibble with the RC Justin was an overly rigorist line on whether lying is always a sin. “Yes,” he said, although there can be circumstances when it is legitimate to withhold information, without actually speaking a falsehood. That doesn’t cut the ice if, say, a misogynist psychopath with a machete turns up at your door, asking whether there are any women at home (and your sister is staying).
The correct answer is “No.” Why? Because the psychopath’s real question is actually “Can I kill your sister?” And to that the right and proper answer is emphatically “No.”
Justin needs to appreciate the legitimate place of informed creativity in ethics. In other words, if, for this little inquisition, I got in touch with my inner Dominican, Justin could do with finding his inner Jesuit.
Canon Andrew Davison is Starbridge Professor of Theology and Natural Sciences in the University of Cambridge, and Fellow of Theology and Dean of Chapel at Corpus Christi College. He is currently a Visiting Fellow of the Center of Theological Inquiry, Princeton, in the United States.