New Scientist - USA (2019-12-21)

96 | New Scientist | 21/28 December 2019

The back pages Q&A

Janelle Shane is on a mission to

reveal the weirdness of AIs. She

explains why they write hilarious

recipes and see giraffes everywhere

As a child, what did you want to

do when you grew up?

I wrote my first science book when I was 5.

I  decided the world needed to know about horses

and that, as I had once seen a horse in real life,

I was uniquely qualified to write about them.

Your book is about why AI is making the

world weirder. What do you mean?

AIs sometimes do things that reveal how very

different they are from human-level intelligences,

like when they delete a list of numbers instead

of sorting it because their job was technically to

eliminate sorting errors. Or when they mistake a

photo of a house for a photo of a giraffe. Because a

lot of science fiction AIs are humanlike, we tend to

forget that what we have today is a lot simpler.

It doesn’t have the understanding that

we have, so it’ll do weird things.

You are known for giving AIs unusual things

to learn about, with amusing results.

Can you give us an example?

One of my favourite experiments is to have AIs

try to imitate recipes. They’ll get the overall title-

ingredients-directions format correct, but they

tend to lose track of what the ingredients are, or

they’ll ask for things like chopped flour and

peeled rosemary and shredded bourbon.

It becomes pretty clear that they don’t

understand that the recipe refers to

things that you do to real ingredients.

Has your field of study changed in the time

you have been working in it?

One of the tricky things about writing a book on

the state of today’s AI is how quickly things are

changing. When I started writing the book in 2017,

state-of-the-art algorithms could barely generate

complete sentences. Now, they’re generating

readable articles, even if the articles don’t make

much sense. I had to stick to what’s been true

about AI for a long time.

“ When AIs try to

imitate recipes,

they will ask

for things like

chopped flour

and peeled

rosemary”

LIFE ON WHITE / ALAMY STOCK PHOTO; JESS KORNACKI

What achievement are you most proud of?

I’m always proud when people read my work and

say they’re no longer afraid of an AI takeover. I’m

also tickled when people say they laughed so

hard they made people stare.

What scientific development do you hope

to see in your lifetime?

I’m really looking forward to the first major film

to use AI-generated imagery. AIs tend to generate

pictures that have lushly realistic textures but

completely weird geometry, so you’ll get melting

clocks with multiple hands and illegible writing,

or cars with lots of mismatched wheels. As AI

tools get more powerful and easier to use, I’m

excited to see what artists will do with them.

Do you have an unexpected hobby, and

if so, please will you tell us about it?

I play the Irish flute, and it was interesting to look

at some of the Irish tunes generated with AI. Some

are weird but others could have been written by

humans. Yet nobody plays the AI music, and I think

that’s because the point is the history of the tunes

and their social context. It’s a glimpse of the role

AI-generated music will have in the future.

It might be useful for background music,

but it won’t replace the music we listen to.

You say AIs tend to spot a lot of giraffes

in pictures. What’s that about?

People take way more pictures of giraffes than

they do of boring rocks or bushes. As a result, AIs

seem to have learned that giraffes are everywhere.

If they’re not sure what’s in a picture – and they do

get confused a lot – they’ll often guess “giraffe”.

What’s the best thing you’ve read or seen

in the past 12 months?

I really enjoyed Exhalation, Ted Chiang’s latest

collection of science fiction stories. There’s a story

in there, “The Lifecycle of Software Objects”, that

explores the moral dilemmas around creating –

and therefore being responsible for – sentient AI.

How useful will your skills be after

the apocalypse?

I don’t have many build-it-from-sticks-and-stones

kinds of skills. But at least people will always need

music and stories.

OK, one last thing: tell us something that will

blow our minds...

Someone once trained a neural net to place bets on

horse races. Its winning strategy? To place zero bets.

Janelle Shane’s book You Look Like a Thing

and I Love You: How AI works and why it’s making

the world a weirder place is out now (Voracious)