How To Think About AI by Richard Susskind review: 'measured and insightful'

This wide-ranging exploration of our AI-assisted future strikes the right note of seriousness without being depressing, writes Stuart Kelly

One of the many topics Richard Susskind deals with in this measured and insightful book about AI is “future-proofing” – essentially, taking steps to ensure the continuation of a business – which is ironic, in that I have rarely read a book so haunted by its own imminent obsolescence. Susskind, educated at Glasgow University and Balliol College, Oxford, did his doctorate on AI and law in the 80s, going on to develop actual systems for legal use, and becoming Technology Adviser to the Lord Chief Justice of England and Wales.

He recounts underestimating AI at almost every turn, thinking it at best a specialised tool at first and quoting New Scientist in 1986 on the “barefaced public deception” about the potential of AI. So he writes with both humility and expertise, both theoretical and practical. As for the current endeavours on future-proofing, he wryly notes that much was written in the 1890s about the “Great Manure Crisis” in New York caused by horse-drawn transport. Nobody thought the solution would be the motor car.

Susskind is ideally placed to step back in order to look forward. If there is one major point to his discussion, it is that policy should be directed towards “what-if AGI thinking?”: in order words, to proceed as if “artificial general intelligence” or “full-scale, human-level performance by machines” is the most likely outcome. One caveat: human-level is a minimum threshold, since already machines can far outperform humans at some tasks. Neither gradualism nor make-do-and-mend will suffice. Nor does Susskind play down the possibility of catastrophe, either by malicious actors or unforeseen consequences; though he does not go quite as far as Batman on Superman – “if we believe there’s even a one percent chance that he is our enemy, we have to take it as an absolute certainty”.

In terms of humanity, he diagnoses the major impediment being “not-us thinking”, a forgivable reaction whereby lawyers think lawyers can’t be replaced – or oncologists think oncologists can’t be replaced. (Dare one add – politicians? writers? priests?) This is just the tip of human-centric thinking. Do self-driving vehicles require manual gearboxes? So much effort and thought went into trying to make robots walk like us – to be C3PO, basically – when, in terms of mobility, insects have always been more flexible and adaptable.

ChatGPT app on a smartphone. Photo: OLIVIER MORIN/AFP via Getty Images | OLIVIER MORIN/AFP via Getty Images

Susskind is enlightening about range, especially since at present ChatGPT is receiving the lion’s share of attention. We do not talk as much about its use for synthesising vaccines (although most of us benefitted from that), determining mineral deposits (I’m sure Ukraine and Greenland has focussed some minds on that), or shipping logistics.

He acknowledges that part of the problem is linguistically rooted, and not just because we infer personhood and agency that might not exist, as when we say “come on, you stupid machine” to a recalcitrant printer, or “hurry up” to a download. Ted Chiang famously referred to artificial intelligence as “a poor choice of words in 1954”, and his preferred “applied statistics” feels less ominous. Susskind’s preference for the prefix “quasi-” (as in quasi-empathy or quasi-judgement) might be etymologically sound – Latin for “as if” – but common usage has more of a sense of not or not quite: quasi-beautiful, quasi-sincere.

This attention to the wording of things is evident in how he approaches the “not-us thinking”, which is inextricably linked to his interest in the difference between process-thinking and outcome-thinking. With slight mischief, he recounts telling an audience that “people don’t want neurosurgeons”, with his follow-up being “they want health”. There are areas where process is important – if we want to look at emergent machine consciousness, we need to “look inside”, rather than just contenting ourself that it does a damn good impersonation of consciousness. But in other areas, outcome is everything. It doesn’t matter to me if there is a silicon chip or a genie in my iPhone. (It does worry me, however, if the genie is selling my data to corporations on the sly). I retain some scepticism about, for example, his analogy of ATMs and bank tellers. He poses interesting points – how did we get out cash at 11pm in the past and why aren’t all-night human cashiers a solution? – but doesn’t answer the question: what are the bank staff doing for a crust now?

Normally, books about AI tend to leave me mildly depressed, but Susskind’s strikes the right note of seriousness. One of the practical suggestions seems an eminently good idea – a national discussion similar to the one Mary Warnock fronted on questions like in vitro fertilisation. I certainly cheered when I read him state “we need to be guided by Plato, Aristotle, and Kant rather than – with respect – Sam Altman, Elon Musk and Mark Zuckerberg”. Although if I had been his editor I would have suggested “with respect” unnecessary. Technologists do not make good decisions about technology, let alone humanity.

At the outset, Susskind relates a public debate between Henry Kissinger and Noam Chomsky, where two titanic figures took diametrically opposed positions. This book is part of the “Guide for the Perplexed” series, and there is no shame in being confused against such a polarised question, especially when the stakes are high. The people Susskind cites – Jaron Lanier, David Chalmers, Nick Bostrom, Martin Rees, Max Tegmark and others – would make a pretty decent bibliography of non-hysterical, sensible futurologists. He says this books wasn’t written with the help of AI. I believe him – the AI would have curbed references to books by his sons.

How To Think About AI, by Richard Susskind, Oxford University Press, £10.99