Yesterday I read Cory Doctorow’s (latest) book on AI. I found it enormously helpful. It’s a stimulating, short read. Two hundred pages, and I finished it in a day, which is more a signal on the quality of the writing than it is on my attention span.

I’m immensely thankful when I read books like this because they are only two hundred pages. The author has resisted the temptation to pad it out with personal anecdotes and spurious guff to meet the usual three-hundred-and-twenty-page requirement of publishers. This book is disciplined, much like his blog.

I’m going to summarise a few of the key points that jumped out at me. I’m aware of the dangers of summarising a book when it is so well written. Often, summarising a book helps clarify the points, here though I feel I am diluting his message. But I need to write more, and if nothing else, writing this post has helped me organise my thoughts on the matter. So, good for me, but perhaps not for you reading this. You should read the book.

Which is to say, if there are any useful insights in my notes below, they are, of course, directly lifted from the book. And if there are any confusing (or incorrect) points, they likely reflect my misunderstanding and therefore subsequent misrepresentation on my part. Again, please read the book.

Firstly, he’s not anti-AI by any stretch. The closing chapters talk passionately about the benefits and undeniable societal improvements we can enable with AI. If I were to describe his view (always fraught when trying to summarise another person’s outlook in a few words), he’s strongly anti-AI harm (and hype) - in particular, the direction he sees focus being pushed, the desperate need to justify increasing investor funding, to push a growth narrative, and the workforce disruption this mandates.

Second, good and bad AI.

To oversimplify his thought: AI can be good if it assists a person (this is the ‘centaur’ in the title).

And bad if it forces the person to be the assistant to AI (the ‘reverse centaur’). Checking the AI results, and now taking responsibility for output from the AI model. The human is no longer a creator. They are a checker, held accountable. Which brings us to the only avenue that can justify the massive investment - massive disruption of the workforce.

In order to provide a return on the AI investment, bosses need to get rid of wages any way possible, whether that’s getting rid of the people altogether, or reducing the work they do and therefore the salaries that they can demand. So, generating return on AI is essentially about getting rid of workers. Which brings us to the problem of: what if the technology can’t adequately replace workers?

What if the AI seller’s promise - reduced workforce - can’t be delivered? In some industries the promise is obviously happening, but in others (despite the breathless claims) it isn’t. He covers automation blindness, with interesting examples. Radiologists losing their effectiveness when relegated to only checking AI diagnoses, TSA agents perfectly catching water bottles in bags, yet missing bombs added by testing teams.

There’s an interesting aside around copyright, which I was not aware of. And to be clear, I’m not a lawyer, so this is simplified, and you should check this yourself (obviously). But my takeaway is that in the case of imagery (as an example), copyright is not a protection, nor a legal claim against AI. Let’s assume someone has copyrighted their image. If a technology:

Analyses the image to work out, statistically, how the pixels are placed next to each other, that then becomes a fact
If that same technology then puts those facts into an algorithm for how an image might be generated, that’s a piece of software. At no point is copyright being broken.
And then, of course, the punch to the gut: if that algorithm then produces another image in a similar way, that’s not a copyright breach. It’s simply using a piece of software to create something.

You and I can look at an AI generated design that to our eyes is a blatant copy, and then be bewildered that copyright hasn’t been compromised.

This important, problematic issue is difficult (perhaps impossible?) to solve. And in the book, he discusses it at length. There’s a lot of nuance. It’s a complex area. But the summary is: Midjourney taking your image, analysing it, ‘algorithming’ it (ie making model), and coming up with a similar output when the algorithm is used, is not breaching your copyright.

Then there’s discussion of environmental impact. How it is played down (with best case environmental accounting), and the net benefits to society over-elevated. The audience for all this PR is not you or me, not even our bosses (who want to replace us), it is primarily the investors, the protected fools (with their circular financing no less) needing to be sure of the future payback. The AI companies don’t care whether your boss actually buys it, they just need the investors to believe your boss will.

For me, the main takeaway is the topic of replacement of workers, and the shifting of monetary reward from the workers who are left displaced to the few controlling companies. This is the problem, the real impact.

So what should we do? How do we respond? That’s probably the main question. And really it comes down to pushing back on a few things:

Push back on when an AI tool is used to replace a person without benefit to that person or their employer (while embracing an AI tool that assists and increases the value of the user)
Push back on public resources being used to provide data centres and infrastructure that only benefit a few big companies
And finally, being prepared to use these leftover resources once the bubble does burst, many of these AI companies disappear. But there is useful infrastructure left behind for good, local practices.

He finishes with an uplifting example of how human rights organisations can use AI to stand up and advocate for the downtrodden. A benefit that, poetically, will increase significantly once the AI bubble pops.

TL;DR: it’s a really helpful book, worth reading.

Streamlined

Reverse Centaur