A Year of Reading: Machines That Think reviewed by Katherine Roberts

I love the serendipity of allowing random books to arrive in my life from unexpected sources... this month's offering is a non-fiction title picked up from a stall at our local Apple Pie Fair. I went for a slice of the giant apple pie, and came home with rather more to digest than I'd bargained for!

Machines That Think
(New Scientist Instant Expert)

'Machines That Think' is part of a New Scientist series written by multiple human experts in the field. Subtitled 'Everything you need to know about the coming age of artificial intelligence', this book practically jumped into my hands. On checking the copyright page, I was rather disturbed to see that it was published in 2017, which is ancient history where technology and AIs are concerned. However, if you study the dawn of computing, it's clear these clever little programs have been around a lot longer than you'd think.

The first chatbot (named Eliza, after Eliza Dolittle in the play 'Pygmalion') was developed in 1966 as an artificial psychoanalyst to help patients by asking simple questions based on their responses during the chat. Eliza was obviously much more basic than ChatGPT et al, but AI learns fast. You've probably spoken to a few Eliza-clones every time you phone or message a helpline, since AIs are increasingly used to screen calls before a human picks them up. And your smartspeaker is essentially a sophisticated chatbot with a pretty name. Siri, Alexa... DO WHAT I SAY!

In the early days, developers approached AIs in a pure way from the top down, attempting to program them for specific tasks in a theoretical manner. Games are perfect training grounds for such specialised systems. The first AI to win a game of chess against a human being was Maniac I in 1956, and in 1997 an AI called Deep Blue famously defeated world champion Gary Kasparov. But when an AI called AlphaGo defeated their leading player Lee Sedol at the game Go in 2016, South Koreans were devastated - how could a machine possibly triumph in a game like Go that requires a certain amount of human intuition?

The top-down approach to programming AIs produced accurate results but limited learning. Following Google's 2009 research paper 'The Unreasonable Effectiveness of Data', developers now approach the problem in another way and allow AIs to train themselves in a more human manner, using a bottom-up approach. This requires massive amounts of data, freely available since the dawn of the world wide web. Using this approach, specialist AIs are being developed for facial recognition, driverless cars, mapping, robotics, image generation, etc, with more general AIs on the horizon.

Some AIs even claim to write stories and music. However, these have been trained in the bottom-up manner on original material produced by human authors and musicians, so their creations are at best derivative. Even if you subscribe to the view that there are only seven basic plots (Christopher Booker), then why does an AI need to devour everything already produced in order to create new work? If it were genuinely creative, why not just program it with those seven plots and let it get on with the job? Those of us who create material for a living - in my case, fantasy and historical fiction for young readers - are watching its development very closely, while also arguing our case that using our published (and unpublished) works without permission to train AIs is a breach of copyright and we should, at the very least, be fairly remunerated for their use. The jury is still out on this matter, although class-actions are currently being brought against AI developers in the US that could pave the way for other such copyright cases.

And here we come to a sticky question in the book... what exactly is intelligence? Is an intelligent human someone who can learn a set of rules and apply them to - say - a game of chess? AI has already proved it can play chess and Go better than human masters of the game. Is an intelligent child the one who has perfect recall and can produce the material they have memorised to answer every question correctly in an exam? AI does that quicker than I could every time I ask Google a question, and I usually find the summary it produces more useful than the original material it must have digested in order to supply its answer. Whether this original material is 100% correct and true, of course, is another question - so maybe that proves AIs are not genuinely intelligent as they can't tell the difference between true and false information? But... er, can you???

A computer brain is a neural network similar to a human brain, so perhaps the difference between an AI and a human is not actually a question of intelligence, but of emotion? Or, if you prefer, does AI have a soul? The science fiction movie 'I, Robot' (2004) claims it is possible for ghosts to exist in the machine. It features an emotional robot called Sonny who apparently overrides Isaac Asimov's three laws of robotics to kill a human being. There's also a powerful AI in the movie called Viki who controls the robots and pretty much everything else. With our definition of an AI, Viki could one day exist (we already have killer robots designed for use in war) - but Sonny cannot, as he seems to feel fear and even love.

In summary, this is what Google's AI has to say about 'Machines That Think':

"Machines That Think" is a series of books from New Scientist, an imprint known for providing accessible introductions to complex scientific topics, with one book in the series focusing on everything you need to know about the coming age of artificial intelligence (AI). The book explores the relationship between AI and human intelligence, how machines can create music and stories, and whether AI poses a threat.

Scream all you like, artificial intelligence is out of the box and living among us. If we can manage to keep AIs harnessed for human use, great. If AIs will eventually enslave or destroy the human race instead remains to be seen.

Meanwhile, I am celebrating being human - the apple pie was delicious!

*

Katherine Roberts is a (real human) author of fantasy and historical fiction for young readers. Her debut fantasy 'Song Quest' won the Branford Boase Award in 2000.

Song Quest (Element first edition 1999)
plus Branford Boase Award

Thirteen of her copyrighted titles appear in a pirate library online that has been used to train AIs.