Book review: You Look Like a Thing and I Love You

I just finished reading Janelle Shane's You Look Like a Thing and I Love You, which must be one of the best books I've ever read on the realities of artificial intelligence (AI). It is certainly one of the funniest I have read for a while, and many times had me laughing out loud. Repeatedly. Shane blogs at AI Weirdness, where she writes about the myriad ways in which AI gets things wrong. And this book is a collection of those examples, as she explains in the introduction:

...the inner workings of AI algorithms are often so strange and tangled that looking at an AI's output can be one of the only tools we have for discovering what it understood and what it got terribly wrong.

Shane illustrates AI getting things wrong with examples of trained algorithms that do all sorts of things from create recipes, heavy metal band names (Inhuman Sand was particularly good, I thought), ice cream flavours (carrot beer sounds intriguing, if weird), to My Little Pony names (Raspberry Turd is probably one that Hasbro already thought of, but dismissed). The title of the book itself is the output of an AI algorithm, trained to write pickup lines.

The hilarious examples are designed to illustrate a number of things, including the 'five principles of AI weirdness':

1. The danger of AI is not that it's too smart, but that it's not smart enough;
2. AI has the approximate brainpower of a worm;
3. AI does not really understand the problem you want it to solve;
4. But: AI will do exactly what you tell it to. Or at least it will try its best;
5. And AI will take the path of least resistance.

If you want to anticipate an AI that is about to fail, you should pay attention to the 'four signs of AI doom':

1. The problem is too hard;
2. The problem is not what we thought it was;
3. There are sneaky shortcuts; and
4. The AI tried to learn from flawed data.

And Shane notes that AIs can fail because we:

"gave it a problem that was too broad;

didn't give it enough data for it to figure out what's going on;

accidentally gave it data that confused it or wasted its time;

trained it for a task that was much simpler than the one it encountered in the real world, or

trained it in a situation that didn't represent the real world".

As you can probably tell, Shane appears to like lists. Some readers will probably find it a little off-putting, but it is an effective way of signposting the various examples that Shane then uses to illustrate the points that each list summarises. Readers who have paid attention to the last ten years of AI development, even just in the mainstream media, will have heard many of the key stories that Shane uses. However, that's why the highlight of the book to me was the examples that Shane gives where she has trained her own AI algorithms and they do dumb things, like the Buzzfeed article titles that include '18 delicious bacon treats to make clowns amazingly happy', and '43 quotes guaranteed to make you a mermaid immediately'.

This book is more upbeat than Cathy O'Neil's Weapons of Math Destruction (which I reviewed here), although it does cover some of the same ground. Shane isn't so much sounding the alarm, as pointing out the key flaws that should lead us to a more realistic assessment of the current potential for algorithms. Think of this as an antidote to the breathless hype that often gets repeated online and in the media. AI might be able to beat the best humans at chess or Go, it may be able to create passable abstract art, but it can't give your cat a halfway sensible name, and it sees giraffes in way too many landscape photos (so much so, that there is even a term for this, giraffing). In fact, right now the best AI might not be purely algorithmic. As Shane notes in Chapter 10:

...AI can't do much without humans. Left to its own devices, at best it will flail ineffectually, and at worst it will solve the wrong problem entirely... So it's unlikely that AI-powered automation will be the end of human labour as we know it. A far more likely vision for the future, even one with the widespread use of advanced AI technology, is one in which AI and humans collaborate to solve problems and speed up repetitive tasks.

This is a great book, which I highly recommend for anyone interested in AI. Clearly I'm not the only one who thinks this way. Diane Coyle rates it highly as well (and it was her recommendation that led me to it in the first place), and it got a special mention in her best of 2020 books list. And if you can't get the book, you can always chuckle along with Shane's blog.