Meredith Broussard usefully identifies the ways in which AI systems incorporate bias - sometimes directly in the systems, at other times in the unjustified ways that they are used. We see powerful examples, for example, of the hugely problematic crime prediction systems where it's entirely clear that these AI systems simply should not be used. A useful pointer is what a 'white collar' crime prediction system would do (and why it doesn't really exist). We get similar examples from education, ability issues, gender rights and medical applications.
What I'd hoped would make the difference from earlier books were solutions, when Broussard brings in the concept of 'public interest technology' and outlines a 'potential reboot'. Again, there is some interesting material, though it can seem to be in conflict with other parts of the book. Earlier Broussard argues powerfully that it's not enough to fix bias in AI systems, because the systems have no understanding of the circumstances - this will always need human input. But under public interest technology, we are told 'algorithmic auditing shows great promise for decreasing bias and fixing or preventing algorithmic harms'. Algorithmic auditing is doing exactly what was said earlier wasn't really possible - 'examining an algorithm for bias or unfairness, then evaluating and revising it to make it better.' In the end, this is the kind of problem where the devil is in the detail - and there is little evidence here of solutions that are aware of this, just as proved the case with the way that GDPR in the EU adds layers of bureaucracy without doing the job.
The book sometimes make statements as fact that don't seem backed up. In part this is because it is so intensely US-focused. There's no attempt to look at different cultural settings. So, for example, early on Broussard makes the statement 'People also consistently overestimate how much of the world is made up of people like themselves.' Yet 'the world' is not the US. Data from the UK, for instance, shows consistently that white people significantly underestimate how much of the UK population is white.
The content sometimes sets things against each other that either don't ring true or don't really go together. For example, talking about a shoplifting incident involving the theft of watches worth $3,800, Broussard states 'It's not a good idea to prosecute shoplifting that is this low in dollar value.' Really? To a small business, losses like that can be catastrophic. We are then told that retailers are partly to blame for shoplifting by introducing self-checkouts to reduce costs. Really? But shops that sell $3,800 worth of watches rarely use self-checkouts - and for many customers, self-checkouts are very useful. Why should they be disadvantaged because it makes it a bit easier for criminals?
Without doubt an interesting book, but it doesn't add much that is useful to the discussion that hasn't already been said.
Hardback:   |
Review by Brian Clegg - See all Brian's online articles or subscribe to a weekly email free here
Comments
Post a Comment