Who owns the content? The issue with generative AI and copyright

Imagine you're chatting with your boss.

She asked you a question about an

approach to social media marketing.

You were just reading a book about this last

night, so the topic is fresh in your mind.

You're ready to answer, but what you say

is, "Sorry, my knowledge on this topic is

derived from a copyrighted work.

Here is a link to Amazon where you

can buy a copy of it to read yourself."

That would be weird, right? Your boss

probably wouldn't be too pleased, and

that's not the way the accumulation of

knowledge works for humans. But in the

debate about AI, a tool that as our

co-pilot is being thought about as an

assistant or a colleague in our work,

that is one potential road we may go

down because of our copyright laws and

because of the assertions
of copyright holders.

In this video, I want to dig into the

issue of training AI, and how should

the ownership of the material that's

used for training be thought about,

managed, and compensated. Since the

introduction of ChatGPT last year, the

issue of ownership of training data for

AI models has come up again and again.

We have seen various efforts through which

content owners, whether they be

well-known authors or celebrities,

artists, even Reddit and the New York

Times have tried to put up roadblocks—

legal and otherwise—to having their

content consumed by AI companies in the

training of their models.

There are good reasons for us to

be concerned about what data our AI

companions are trained on. A training

dataset consisting of
only the New York Times

versus only Reddit would clearly

lead to some very different

understandings around topics of

importance to our world. Unless the data

demonstrates a broad range of the

experiences we have and the concepts

we've come up with as humans, AI will be

limited in its ability to help us in a

relevant way. But the issue I'm talking

about today isn't really connected to

that quality issue, apart from my concern

that if rights holders to high-quality

work aim to have it extracted from our

AI models, what would be left? While there

have been various arguments against

using certain data for AI training, most

of them come back to one thing:

compensation. Those who assert ownership

rights over certain content that could

be useful to AI training datasets want

their slice of the pie that these

companies will generate with the AI

models they have trained. And given that

the AI economy could be worth up to

13 trillion dollars by 2030,
according to McKenzie—

which, if it were a country in its own

right, would be the world's third-largest

economy based on 2022 data—
who can blame

them, right? Well, it makes perfect sense,

apart from the fact that in most most

cases, we don't license knowledge in this

way. A rights holder is completely

entitled to compensation when their work

is consumed, but they don't normally get

residuals based on derivative output,

except in situations where those

derivatives are substantially similar to

the original. Under copyright law, we have

concepts such as fair use, but we also

have thousands of years of human history

where we can see a line where our

knowledge today simply derivates from

those issues that have been the subject

of work in years gone by. Let's do a

little quiz. I'm going to describe

something, and you go ahead and think

about what I'm describing. Maybe write it

down, maybe call it out.
See if you can guess.

So, number one: I'm green, I'm kind

of slimy, and I can jump really well.

What do you think that is?
Well, if you thought

frog, then you'd be right. Number two: I'm

covered in black and white stripes,

I look kind of like a horse, and I live on

the African Plains.
What do you think it is?

Did you think zebra? Number three: I'm a

building on Pennsylvania Avenue in

Washington D.C., where the

President of the United States lives.

Did you think White House?

But assuming you can get some or all of

those correct, can you tell me how you

know these things? Did you learn them

from a copyrighted work? Did you see them

in a copyrighted picture? Has the rights

holder given you permission or licensed

that information in such a way to allow

you to use it for your
benefit in daily life?

Think about everything else you know.

Are you sure that in your last

meeting at work, you didn't inadvertently

regurgitate something you read in a blog

for which you're not the rights holder?

If you excluded every piece of

information you've ever consumed that

was delivered by a book, or an article, or

a photo, or a video that may be owned by

someone else, how much
would you know now?

It's tricky, right? This is a knot that's

very hard to untie. While I can put

together a list of specific sources for

a lot of my content, there's a lot of the

flavor of what I talk about that's

simply built on years of knowledge.

I have no idea why I know many of the

things I know, and there certainly isn't

a demarcation in my brain for

copyrighted and non-copyrighted material.

But you're probably thinking, why is this

relevant to the issue of AI?

Well, while AI doesn't learn in the

same way humans do, its accumulation of

knowledge is in some ways remarkably

similar to ours. When we come across

someone who seems to know a lot and can

reference a lot of information, we will

sometimes describe them as being

well-read. In some ways,
GPT4 is just the most

well-read intelligence on the planet—

it just happens to be an artificial one.

In general, when we describe a person as

well-read, it's as a result of them

showing off their knowledge, and it's a

compliment. I'm not sure whether there's

ever been a case where a human being

described as well-read has resulted in

them being questioned as to whether they

legally obtained the books they've been

reading, or had properly licensed the

knowledge they used.
Whether that well-read

person is a poor student or whether

they're a billionaire C.E.O., we just take

for granted that they have spent time

accumulating knowledge, and many of us

respect that, and we don't place a

knowledge tax on the benefits they've

accured from that knowledge, financial or

otherwise. In my opinion, by focusing on

the value generated with knowledge and

attempting to find a way of sharing in

it, many content rights holders are

inadvertently seeking to upend centuries

of progress in how we educate and better

our society that was initiated with the

development of the printing press. And if

any modern author or artist claims that

anything they have ever produced is

truly original and not developed through

standing on the shoulders of centuries

of accumulated knowledge and art in our

society, then they're
just kidding themselves.

But more than trying to upend

the status quo, it gets worse. Have you

heard of subliminal advertising? This is

the concept that if you insert a message

or idea in a piece of media, maybe a

video or an audio ad, for example, that is

below the threshold of human conscious

perception, that your subconscious will

still pick up on it. The efficacy of this

approach to communicate anything is

debated, but there are limitations on

this sort of messaging in certain

jurisdictions, because ultimately, as

humans, we don't particularly like the

idea of our subconscious brains being

gained by someone to their advantage.

Imagine if this approach were used in

the teaching materials used in schools.

Perhaps a textbook publisher takes

that page that taught you
that the White House

is the White House and secretly

inserts content that tries to convince

you that the structure is actually

Buckingham Palace in London. Maybe it

sends those particular textbooks to the

school districts that pay the least for

their materials, or it sends them to the

school districts that refresh their

materials less frequently than the

publisher would like. This ends up as a

penalty on knowledge based on the books

rights holders perception of some

financial slight or other. You have a

bunch of children out there in the world

whose knowledge has been intentionally

poisoned for commercial reasons, and as a

society, we have no idea what the

long-term implications of that might be.

The good news is that, as far as I know,

that isn't going on in schools. The bad

news is that some innovative individuals

with issues with AI are attempting to do

it for AI model training by similarly

poisoning the digital images they

publish on the internet. This technology

advance called Nightshade was released

as an open-source poison pill by

researchers at the University of Chicago

back in October. Their goal is to give

visual artists and image publishers a

tool to protect their work by corrupting

the AI data set with images that show

one thing, but the AI has learned to show

something else. But imagine if the

University of Chicago sanctioned

research that helped those textbook

publishers protect their intellectual

property by teaching a generation of

school children whose chool boards have

not approved new book purchases that

zebras were actually frogs.

We would be outraged, and we should

be outraged by this research, as we have

absolutely no idea how far the long-term

consequences of such actions will go.

Now, I'm certainly not saying that rights

holders shouldn't be compensated. They

definitely should. And when we talk about

rights holders, we aren't just talking

about giant, wealthy corporations, we're

also talking about individuals, we're

talking about small businesses. I'm not

arguing that anyone should be made

worse-off by AI, and
it's really important that

AI tools still provide us
information like citations

to make it really easy to go off and

purchase or read the work if we need to.

Anyone who currently makes their living,

or even part of their living from

creating copyrighted work should be able

to have a road map for the future with

the inclusion of AI that allows them to

continue being successful. But AI also

shouldn't be some licensing gravy train

that takes authors or artists who are

making meager incomes from their work and

rewards them more handsomely on the

off-chance that it was their work that was

critical to some answer that ChatGPT

just gave. Just because companies like

Microsoft and Google have deep pockets

shouldn't radically change how we reward

and compensate those who create

knowledge in balance with the overall

societal good of sharing it. Now, it's

important to point out that there have

been some elements reported around the

training of AI models that suggests that

outright stolen works have been used.

This should be a line that isn't crossed,

and it is, and should continue to be

illegal. AI companies don't get to opt

out of the knowledge economy that

already exists, but their use should be

able to operate within existing guard

rails without too much change, and

without radically altering anyone's

protections. If you choose to publish

your content on the open internet, you

probably do so because it gives you some

advantage in life or business.
For example,

I publish here on YouTube, as it

helps me participate in the wider tech

community. I sometimes get clients from

people who watch my videos, and I gain

connections that are
relevant to my interests.

It's for me to assess whether

publishing on YouTube, which is open to

everyone, is what I want to do, but once

I've made the decision to put content

here rather than behind a paywall

somewhere else, I don't get to decide in

any substantive terms who gets to see

it and who gets to distill the

information I share. But there are

different paths. With my new book,

"Who's in the Co-Pilot Seat?", I've made a

different business decision. I've

published that book for commercial sale

and if someone wishes to have a copy,

they need to buy it. But again, once they

bought or otherwise legally obtained a

copy, if they use the information that

book contains to their benefit, including

their financial benefit—

as long as they don't infringe my

copyright by regurgitating my content

verbaitm—I don't get a say. And talking

of paywalls, these are odd things.

A while ago, ChatGPT had to remove its web

browsing function because it was

defeating pay walls and presenting

non-public information. However, to my

understanding, it's not like the AI was

seeing the paywall and hacking its way in.

The way that paywalls are

implemented is rather fragile in lots of

cases. Some websites publish the full

text for search engine indexing to get

up the Google rankings. Others allow a

certain number of articles to be read

before the paywall kicks in.
In some cases,

paywalls are really just presentational

elements on a page, rather than something

clever going on in the background of a

website. The point is that, in the cases

I'm familiar with, an AI's ability to

access paywalled content has come down

more to how the publisher has chosen to

publish their content in order to drive

the most traffic to it, rather than

some malicious act on the part of the AI

vendor. Perhaps if I left copies of my

new book on tables at coffee shops

across the city, I would sell more copies

or get more clients. And maybe not, but we

wouldn't consider anyone who picked up

one of those books and read it to be

stealing my content. The fact is that in

the online world, publishers often do one

thing with their content and then lean

on their terms of use to regulate it.

Where if you compared to a largely similar

approach in the physical world, it simply

wouldn't stand up to scrutiny. I have no

business in what any consumer of my

content gets out of it, other than

wanting to make the best content. If you

take what I say, or write and turn it

into a business 10 times the size of

mine, good for you. I have no remedy

further down the road to come knocking

on your door to say, "Actually, when I

charge you $12 for my book, I really meant

$12 million, because you've done so well."

That just isn't the way the

transfer of knowledge works in our

society, and creating a road for it to

become the way the transfer of knowledge

works, in my opinion, is a very dangerous

path. For published works that aren't

accessible on the open internet, there is

already a perfectly suitable

non-copyright infringing model that

could work today for AI training: that's

the public library. Libraries buy books

and other materials and are able to

share them widely in the pursuit of

maximizing society's knowledge. We see

libraries and their content as a public

good, and authors are pleased to see

their work in libraries for the most

part, rather than seeing it as theft.

Libraries are able to operate because of

both legal protections and licensing

agreements, and many are transitioned

into a more digital world where their

content isn't solely books on the shelves.

The concept of a library, both as

a source of knowledge to turn a

non-reader into someone who is well-read,

and a model for how protected content

can be shared without harming copyright

holders is, in my opinion, the most

readily relevant to the issue of both AI

training and AI model use. After all, if

someone who becomes a tech billionaire

shares a life story that involves them

starting with no books at home and

relying on the public library to allow

them to break barriers of social

mobility, they are lorded as an example

to us, not derided as a thief. Companies

like Microsoft, like Google, like open-AI,

have the resources to build the fullest

libraries known to our civilization, and

that is what they should be encouraged

to do to train the tools that will be

the next chapter of our planet. And those

who are trying to poison the well of

knowledge should not be held up as

warriors for the rights of creators, but

as troublemakers against one of the

pillars of the betterment of society: our

ability to build on top of the knowledge

and creativity that has come before us.

Ultimately, this is something from which

we all benefit, and in my opinion, solving

this problem is a simple as opening some

new libraries that are as relevant to

the AI challenges of today as they were

to the literacy challenges
of years gone by.

It might require tweaking some laws

or tweaking how some things are licensed,

but in my opinion, the
model is already there.

What do you think? Have I got this

issue confused, or should we be looking

for a simpler and more universally

beneficial solution to what has become a

very complex question? Let me know down

in the comments. Thanks for watching

through to the end. I hope this was

useful to you. Until the next video.

Bye-bye.

[Music]

bye-bye