91av

STATISTICIANS Nick Polson and James Scott want us to understand artificial intelligence so “we can harness its power for a better world”. They will certainly succeed in spreading public understanding, but their account is far from reassuring.

In each chapter they employ anecdotes to contextualise a statistical technique. So we learn how AI has been used to reduce toilet paper theft and improve cucumber detection, for example. In doing so, the pair dispel the myth that AI is a potentially out-of-control entity, beyond the layperson’s understanding. And if you set aside the bits about , racially biased criminal justice algorithms in Florida and the appalling state of advanced data science in the healthcare sector, this book will help you sleep at night.

I wanted to hear more about AI being used to combat climate change, inequality, human rights abuses and soil depletion, but the authors prefer to concentrate on basketball, motorsport, autonomous vehicles, financial markets and social media.

Facile as this might sound, they also spin some winning tales. Who wouldn’t want to be treated with the smart electrosurgical knife developed by Zoltan Takats and a team at Imperial College London? It cuts away cancerous cells and knows when it has hit healthy tissue – and hence when to stop – by analysing the smoke from the cauterised tissue.

When I learned that one of the team was developing tools for driverless cars, and only later applied the techniques to detecting skin cancer, it made me think that our social priorities need to be switched. I’m in good company: Mark Sendak at the Duke Institute for Health Innovation is quoted in the book as saying that “huge health systems make money on advancing chronic disease”.

But this feels too easy. There is no shortage of AI development in healthcare, exploiting a huge body of available data. (Its quality is a separate question.) Are we really to believe that medical researchers are reluctant to be identified as slayers of disease?

Elsewhere the authors suggest that having to take legal and moral responsibility for the behaviours of diagnostic algorithms acts as a disincentive for their development. The authors go so far as to claim that “lawyers and policy-makers haven’t gotten off their backsides” to address questions about liability. If this was true, the whole field would be in legal meltdown: lives are at risk, and liabilities must be assigned, in many fields besides medicine, from military operations to Formula 1 racing.

“The authors spin some winning tales. Who’d turn down being treated with a smart surgical knife?”

Data, and the models that are used to interrogate it, are paramount in any AI system. If you handle quantities of data, you quickly learn that there is no such thing as unbiased data collection.

The authors meticulously take apart a contraception fallibility analysis that appeared in . This indicated that, over 10 years, a woman’s chance of becoming pregnant while taking the pill is more than 60 per cent. Polson and Scott use the same data set, but apply different (and to my mind more reasonable) assumptions to get a different result. The original article remains on the web, an uncontested statistical outlier, raising a wider point about the power of media to equate data with truth.

Overall, reassurance is the watchword here. Scott and Polson’s premise is that humans are integral to AI, and so AI is becoming more human.

If that’s the case then I think we are missing a trick. Why aren’t we emulating another organism entirely? I vote for the dolphin. I reckon it would look after the planet a lot better than the selfie-taking, market-shorting, motorsport-obsessed idea of a human that predominates now.

This article appeared in print under the headline “What could go wrong?”