Hallo from Ljubljana, Slovenia, where I am attending this year's UNESCO Global Forum on the Ethics of Artificial Intelligence.
(If any of you are confused why UNESCO, usually associated with old memorial sites, is interested in AI - it falls under their 'science' brief. And they actually produced one of the more widely respected frameworks for ethics in AI a couple of years ago).
I'm going to be speaking very briefly on a panel about ethical impact assessments. This reminded me that last year I delivered a short 'provocation' on AI ethics at the Gropius Bau in Berlin in August last year. It was for a rather strange event called "Ether's Bloom", also the event was a Saturday morning, so I wrote a deliberately slightly provocative piece called "AI is Boring" (heavily featuring my family's Havanese dog Toby). I doubt I'll be repeating a similar tone this week, lest I find myself banned from various heritage sites. But I also realised I'd not shared the piece more widely, so here it is.
Relatedly - despite some last minute shenanigans (the technical governmental term), the EU AI Act passed last week. This is a big deal. A few people I'd recommend following on it are Luca Bertuzzi (the journalist who got so many scoops that it became a running joke amongst the negotiators), Joanna Bryson, and of course my own employers AlgorithmWatch; I can assure you the team internal chat has been pinging at all times of the day for at least the past month.
AI Is Boring
Last week I visited my family in the UK. This involved spending a lot of time with a 2 year old dog called Toby. As many of you may agree, dogs are incredible creatures. A dog’s sense of smell is around 40x better than a human’s. They can detect cancers by smelling, a feat which we would expect should take sophisticated machines. The fastest dog breeds can accelerate to over 70 km per hour in 6 strides; by comparison the fastest human in the world has reached less than 45.
But of course to my family (and me), the fact that Toby the dog can smell things and run fast are not the best things about him. We get excited when Toby does things like stand on his hind legs, or tries to join in our conversations, or responds to words we’ve taught him. In other words, we get most excited when this dog, this incredible and highly evolved creature in its own right, behaves like a human.
Artificial intelligence can do extremely impressive things in a range of highly specialised spaces. In the past I’ve personally used quite narrow and specialised AI methods to analyse things like hate speech and political interference on social media. It’s performed great feats in medical fields ranging from drug discovery to cancer detection. I’ve recently spoken to people who want to use AI to make it easier for people to discover highly specialised engineering parts, allowing people to repair more things.
But the real boost to AI’s notoriety came from the rapid uptake last year of ChatGPT, a machine which, in front of our eyes, performs the feat of writing an email. And now a lot of the talk is about Artificial General Intelligence, which could perform a very wide range of human-like tasks. The phrase that’s often used is that it will “outsmart humans”. Which raises the question of why people who are very smart are so keen to develop something that they think will outsmart them. I think many of them just find that idea exciting.
Now I don’t want to dismiss the technical innovation underpinning recent AI tools like ChatGPT. I also don’t want to downplay its potential uses and impacts. But a lot of these boil down basically to two things - speed and cheapness. Which is the technological equivalent of getting a McDonalds. Speed and cheapness are important, sure, but are they that interesting? Are they worth this level of excitement? I’m reminded of these words of the author Terry Pratchett:
“People think that it is strange to have a turtle ten thousand miles long and an elephant more than two thousand miles tall, which just shows that the human brain is ill-adapted for thinking and was probably originally designed for cooling the blood. It believes mere size is amazing. There's nothing amazing about size. Turtles are amazing, and elephants are quite astonishing. But the fact that there's a big turtle is far less amazing than the fact that there is a turtle anywhere.”
I want to take Pratchett’s invitation to consider, carefully, what we do find interesting, and what we could or should be interested in. Let me clear. I’m not saying being boring is bad. I am saying it would be good for us to think of AI as more boring. There are implications for some of the most important questions around AI: how we regulate and control it. You sometimes hear people say - AI is totally new. It’s unlike anything we’ve ever seen before. It’s a black box. We don’t understand how it works or what it’s doing. So how can we regulate and control it?
Fortunately humans have millenia of experience in regulating and controlling entities we don’t understand, but that can nonetheless cause harm. Those entities are called ‘other humans’. When a person runs into me in a car, or posts lies about me online, or hacks into my bank and steals my money, we don’t say “well, I don’t fully understand this person, so I can’t do anything about this.” So why would we say that about technology, if we can see it’s causing harm?
Now, at this point I will step away from my provocateur stance to actually talk about my day job. The above comparison, of regulating technology versus regulating humans is a slightly glib one, to make the point that regulating something doesn’t require fully understanding it. But regulating technology is different to regulating humans, in many ways.
Much of my day job involves thinking about issues around regulating technology. I work for AlgorithmWatch, a human rights organization which works to ensure algorithms and AI do not weaken justice, democracy, and sustainability, but strengthen them.
I lead their work on auditing so-called “systemic risks”. Systemic risks are not just ‘bad things happening because of technology’. You can’t realistically stop algorithms causing anything bad to happen, like how you can’t ensure all medicines are 100% safe or that cars never crash. But systemic risks are bad things that, due to way technologies are designed and made, we can expect to keep happening, and could potentially have quite widespread and complex harms. For instance, a car crash might be a one-off event; the fact that cars produce polluting fumes is systemic. Someone might unexpectedly have a bad response to a medicine; the fact that historically a lot of medical research is built on a default assumption that a patient is a white male is a systemic problem.
So how do we locate and address the systemic risks that AI might pose? There’s a whole industry springing up recently around AI safety, AI ethics, and AI regulation. Some of it is reassuringly boring. The EU, characteristically of the EU, has taken a very boring - and I’d say pragmatic - approach in its AI Act, which should be finalised in the next couple of months. It treats AI as a product, defines some risky forms of these products, and says these should be registered and follow certain standards. An Executive Order released by the White House last week is also reassuringly boring, talking about things like the purchasing power of Federal bodies.
But you also hear a lot about more exciting, even science fiction, frontier risks, catastrophic risks, even risks of human extinction. My home country of the UK last week ran a summit on AI Safety, which explicitly focused entirely on these frontiers. And in case that wasn’t exciting enough for Silicon Valley, there was even a talk show style discussion between Prime Minister Rishi Sunak, and Elon Musk (given how unpopular Mr. Sunak currently is, a cynic might suggest he was doing a job interview). You also see a very large number of institutions being set up, in the States but also in other countries, often with links to people who are actually making the technology, to think about these speculative future risks.
Now, I’m not saying we shouldn’t try and think ahead to potential future risks. But these frontier risks command a lot of money, and a lot of attention. Technology firms have actually become the largest lobbying sector in the EU, spending 113 million euros on lobbying last year. By contrast the European Partnership on AI and Society Fund spends around 2 million a year across 40 different civil society organisations. Disclaimer, AlgorithmWatch is one of the beneficiaries, and I can tell you trying to keep up with fast-moving legislation with not much money is… well, it’s not boring.
But when we look away from these frontier, high tech, speculative risks, we see that a lot of the harms of AI - potential harms, but also real harms that are already happening - are really, depressingly, familiar. Let me give you some examples. It is known that many AI chatbots often produce various sorts of discriminatory content. Most notoriously, in 2016, Microsoft created a chatbot called Tay that learned from humans that it interacted with. They then released this chatbot to talk with people on Twitter. You can see the problem here? Very quickly, users began teaching it to spew out hate speech and conspiracy theories, such as holocaust denial. Developers of more recent AI technologies have learned from this experience. For instance, when OpenAI developed ChatGPT, they used a process called Reinforcement Learning from Human Feedback to train it not to give toxic answers. We’ll come back to that later. But this was not an unforeseeable, novel problem. Anyone who knows Twitter knows this was easily predictable.
Now Tay was a dramatic and vivid example of the fact that AI reflects the human biases in the data it’s trained on. And we can say - well, we know this, and we can learn from it and improve the systems. But again, a lot of the issues which come from these kinds of biases are more boring, but more important, than a toxic Twitter chatbot. It may not get as much attention, but there are few things which are both as boring and important as bureucracy. We’ve been tracking various ways in which algorithms are used to assign social security payments in ways which discriminate against single mothers, or using nationality data in tools used to predict likelihood of committing domestic violence, for example.
And sometimes it isn’t even the tech itself which produces the problems. It’s all the boringly familiar stuff around it, like supply chains. I mentioned Reinforcement Learning from Human Feedback earlier. It’s been a really good step to make ChatGPT less toxic than previous chatbots. Riley Goodside, a developer who tests these models before they are released has said that before this step the models really can become pretty toxic quite easily.
However, when we look at how the Reinforcement Learning from Human Feedback was actually done, we see a boringly familiar problem of exploitative supply chains. In order to teach the model to recognise toxic content, a whole load of examples of toxic content was shipped out to Kenya by a company called Sama, which employed people to read and label this awful stuff for nine hours a day, on between $1.32 and $2 per hour. And to be clear, some of the content they were labelling was absolutely horrendous, to the point of traumatising. This isn’t some new, cutting-edge science fiction form of harm. This is a familiar sweatshop model, transplanted to AI, of prioritising fast and cheap outputs over human beings in the Global South.
So my job in assessing the risks and harms of technology is to think - or rather encourages governments, regulators, and technologists to think - about all the stuff wrapping around AI. How the tech is produced, how it operates, how its used, and what’s in place if it goes wrong. That work isn’t helped when people - be that politicians, be that journalists, be that the developers themselves - are pushing all the attention towards the exciting technologies, and saying how totally new and unprecedented this all is. As I hope I’ve made clear, that doesn’t help us address real, familiar, often boring, but very pressing harms. And it certainly doesn’t help when trying to build a new - and even potentially dangerous - “artificial general intelligence” is deemed a better use of smart people’s time than really thinking about and helping to address major existing problems in the world.
To end where I began. Toby, the dog. I’ll accept, the things which bring us delight are the times he behaves like a human. But not too much like a human. The fact he enjoys a cuddle, or to play, or to go outside with us is nice. The fact we can teach him some words is cool enough. But I don't want him to learn to talk. He'd probably never shut up. I won’t feel like a failure if I never teach him to do maths. It may be more boring than playing god, but sometimes it’s enough to have a slightly dumb dog, and focus on the things that really matter.
___
Fun Fact About: Bears
This is a redux of the first ever Sideways Looks fact; but I told it to someone the other day, and their response reminded me what a great fact it is and that it bears (pun intended) repeating:
If you were told that Ursos Arktos is the scientific name for a kind of bear, you’d probably guess the polar bear. It’s actually the brown bear. The two words are, respectively, Latin and ancient Greek for ‘bear’ – i.e. the brown bear is Bear Bear, the bear-est of all the bears.
The reason we might associate Arktos with Polar Bears is becase: The Arctic Circle is named after the bears. It means 'the circle which has bears'. And the Antarctic Circle is ‘the circle which doesn't have any bears’. Relatedly, it’s also quite a good way to remember which one has the penguins.
___
Recommendations
Tech review: I have a weird thing where I don't watch many films, but I do love reading film reviews. Similarly, I have little interest in Apple's new headset the Vision Pro; but I still really enjoyed Joanna Stern's video review of it, which is a wonderful encapsulation of how fun the format can be. Basically, one question people have been asking is - how long would you really want to wear the headset for? So she decides to do a review video of her wearing it for 24 hours. Oh also, because it looks like ski goggles, she films it in a ski chalet. For a more serious, and interesting, discussion of it I'd recommend the Vergecast (which, similarly, is one of the few tech podcasts to reliably make me laugh out loud).
Games: Last newsletter I introduced you to the excellent, very complicated, two-player Cold War reenactment board game Twilight Struggle. My board gaming friend has since told me there's an app, so you can play remotely with a friend. This is a lovely way to stay in touch, and also to find yourself spending far more brainpower thinking about the stability of South East Asia and of NORAD than you really should.
Anti-Review: I was in Brussels last week for work. I took the new Berlin-Brussels European Sleeper Train. It is badly named. Turns out rolling stock from the 1950s is quite rattly - who knew?
Slovenia. It's really pretty. I was last in Ljubljana about 10 years ago for an academic conference in Southern Austria, as it was the nearest airport. I somehow got confused, booked my flight 24 hours too early, and had an unexpected and delightful 24 hours here. Also the trip to Austria went via lake Bled - look how gorgeous it is! So it's delightful to have a chance to return. Thanks UNESCO, and good work on the castles.
Thanks for reading. Please do share with others using this link, and let me know your thoughts via this short poll.