Transcript of YouTube Video: How to Govern AI — Even If It’s Hard to Predict | Helen Toner | TED

When I talk to people about artificial intelligence,

something I hear a lot from non-experts is “I don’t understand AI.”

But when I talk to experts, a funny thing happens.

They say, “I don’t understand AI, and neither does anyone else.”

This is a pretty strange state of affairs.

Normally, the people building a new technology

understand how it works inside and out.

But for AI, a technology that's radically reshaping the world around us,

that's not so.

Experts do know plenty about how to build and run AI systems, of course.

But when it comes to how they work on the inside,

there are serious limits to how much we know.

And this matters because without deeply understanding AI,

it's really difficult for us to know what it will be able to do next,

or even what it can do now.

And the fact that we have such a hard time understanding

what's going on with the technology and predicting where it will go next,

is one of the biggest hurdles we face in figuring out how to govern AI.

But AI is already all around us,

so we can't just sit around and wait for things to become clearer.

We have to forge some kind of path forward anyway.

I've been working on these AI policy and governance issues

for about eight years,

First in San Francisco, now in Washington, DC.

Along the way, I've gotten an inside look

at how governments are working to manage this technology.

And inside the industry, I've seen a thing or two as well.

So I'm going to share a couple of ideas

for what our path to governing AI could look like.

But first, let's talk about what actually makes AI so hard to understand

and predict.

One huge challenge in building artificial "intelligence"

is that no one can agree on what it actually means

to be intelligent.

This is a strange place to be in when building a new tech.

When the Wright brothers started experimenting with planes,

they didn't know how to build one,

but everyone knew what it meant to fly.

With AI on the other hand,

different experts have completely different intuitions

about what lies at the heart of intelligence.

Is it problem solving?

Is it learning and adaptation,

are emotions,

or having a physical body somehow involved?

We genuinely don't know.

But different answers lead to radically different expectations

about where the technology is going and how fast it'll get there.

An example of how we're confused is how we used to talk

about narrow versus general AI.

For a long time, we talked in terms of two buckets.

A lot of people thought we should just be dividing between narrow AI,

trained for one specific task,

like recommending the next YouTube video,

versus artificial general intelligence, or AGI,

that could do everything a human could do.

We thought of this distinction, narrow versus general,

as a core divide between what we could build in practice

and what would actually be intelligent.

But then a year or two ago, along came ChatGPT.

If you think about it,

you know, is it narrow AI, trained for one specific task?

Or is it AGI and can do everything a human can do?

Clearly the answer is neither.

It's certainly general purpose.

It can code, write poetry,

analyze business problems, help you fix your car.

But it's a far cry from being able to do everything

as well as you or I could do it.

So it turns out this idea of generality

doesn't actually seem to be the right dividing line

between intelligent and not.

And this kind of thing

is a huge challenge for the whole field of AI right now.

We don't have any agreement on what we're trying to build

or on what the road map looks like from here.

We don't even clearly understand the AI systems that we have today.

Why is that?

Researchers sometimes describe deep neural networks,

the main kind of AI being built today,

as a black box.

But what they mean by that is not that it's inherently mysterious

and we have no way of looking inside the box.

The problem is that when we do look inside,

what we find are millions,

billions or even trillions of numbers

that get added and multiplied together in a particular way.

What makes it hard for experts to know what's going on

is basically just, there are too many numbers,

and we don't yet have good ways of teasing apart what they're all doing.

There's a little bit more to it than that, but not a lot.

So how do we govern this technology

that we struggle to understand and predict?

I'm going to share two ideas.

One for all of us and one for policymakers.

First, don't be intimidated.

Either by the technology itself

or by the people and companies building it.

On the technology,

AI can be confusing, but it's not magical.

There are some parts of AI systems we do already understand well,

and even the parts we don't understand won't be opaque forever.

An area of research known as “AI interpretability”

has made quite a lot of progress in the last few years

in making sense of what all those billions of numbers are doing.

One team of researchers, for example,

found a way to identify different parts of a neural network

that they could dial up or dial down

to make the AI's answers happier or angrier,

more honest,

more Machiavellian, and so on.

If we can push forward this kind of research further,

then five or 10 years from now,

we might have a much clearer understanding of what's going on

inside the so-called black box.

And when it comes to those building the technology,

technologists sometimes act as though

if you're not elbows deep in the technical details,

then you're not entitled to an opinion on what we should do with it.

Expertise has its place, of course,

but history shows us how important it is

that the people affected by a new technology

get to play a role in shaping how we use it.

Like the factory workers in the 20th century who fought for factory safety,

or the disability advocates

who made sure the world wide web was accessible.

You don't have to be a scientist or engineer to have a voice.

(Applause)

Second, we need to focus on adaptability, not certainty.

A lot of conversations about how to make policy for AI

get bogged down in fights between, on the one side,

people saying, "We have to regulate AI really hard right now

because it's so risky."

And on the other side, people saying,

“But regulation will kill innovation, and those risks are made up anyway.”

But the way I see it,

it’s not just a choice between slamming on the brakes

or hitting the gas.

If you're driving down a road with unexpected twists and turns,

then two things that will help you a lot

are having a clear view out the windshield

and an excellent steering system.

In AI, this means having a clear picture of where the technology is

and where it's going,

and having plans in place for what to do in different scenarios.

Concretely, this means things like investing in our ability to measure

what AI systems can do.

This sounds nerdy, but it really matters.

Right now, if we want to figure out

whether an AI can do something concerning,

like hack critical infrastructure

or persuade someone to change their political beliefs,

our methods of measuring that are rudimentary.

We need better.

We should also be requiring AI companies,

especially the companies building the most advanced AI systems,

to share information about what they're building,

what their systems can do

and how they're managing risks.

And they should have to let in external AI auditors to scrutinize their work

so that the companies aren't just grading their own homework.

(Applause)

A final example of what this can look like

is setting up incident reporting mechanisms,

so that when things do go wrong in the real world,

we have a way to collect data on what happened

and how we can fix it next time.

Just like the data we collect on plane crashes and cyber attacks.

None of these ideas are mine,

and some of them are already starting to be implemented in places like Brussels,

London, even Washington.

But the reason I'm highlighting these ideas,

measurement, disclosure, incident reporting,

is that they help us navigate progress in AI

by giving us a clearer view out the windshield.

If AI is progressing fast in dangerous directions,

these policies will help us see that.

And if everything is going smoothly, they'll show us that too,

and we can respond accordingly.

What I want to leave you with

is that it's both true that there's a ton of uncertainty

and disagreement in the field of AI.

And that companies are already building and deploying AI

all over the place anyway in ways that affect all of us.

Left to their own devices,

it looks like AI companies might go in a similar direction

to social media companies,

spending most of their resources on building web apps

and for users' attention.

And by default, it looks like the enormous power of more advanced AI systems

might stay concentrated in the hands of a small number of companies,

or even a small number of individuals.

But AI's potential goes so far beyond that.

AI already lets us leap over language barriers

and predict protein structures.

More advanced systems could unlock clean, limitless fusion energy

or revolutionize how we grow food

or 1,000 other things.

And we each have a voice in what happens.

We're not just data sources,

we are users,

we're workers,

we're citizens.

So as tempting as it might be,

we can't wait for clarity or expert consensus

to figure out what we want to happen with AI.

AI is already happening to us.

What we can do is put policies in place

to give us as clear a picture as we can get

of how the technology is changing,

and then we can get in the arena and push for futures we actually want.

Thank you.

(Applause)