WEBVTT

(upbeat music)

- Governments are mostly focused

on labeling AI-generated content.

Creative professionals like you and me

are more focused oftentimes on the copyrighted data

being used to train the LLMs.

So today we're going to look at the very differing concerns,

what countries have made, what legislative measures,

and what all of this means for creatives like us.

Now there's a pretty big disconnect

in the global conversation about artificial intelligence.

So governments in India, the EU, and the US

are primarily focused on regulating the output of AI.

And these regions are acting on potential societal

or business and political harms of things like deep fakes,

misinformation, and other forms of deceptive content.

But creative professionals, we are often focused

on what's going into the LLMs,

like what's being scraped basically from the internet.

We are pretty concerned with the unauthorized

and very uncompensated scraping of copyrighted work

to train AI models in the first place

for these corporations to profit from.

And in this episode, we're going to look at this input

versus output gap in the conversation.

I'll cover some of the latest legislation aimed

at regulating AI, which is usually not addressing

the primary economic and ethical concerns of creators,

but at least doing something to help identify

what is synthetic versus what was created by a human

is good.

And just this week, a couple of days ago,

we got a major update from Australia that shows

that some governments are finally starting to listen

to the people impacted by all of our IP

being scraped into LLMs.

Hi, I'm Jen, and this is the Human Internet Theory.

And in this show and podcast, I talk about changes

to human creative content on the internet

and how that is affecting the creative professional industry.

What we need to know and when possible,

what we can do in response.

Before we get to the core issue for creators,

we will first look at what governments around the world

are actually doing right now.

So let's start with India.

The government there has just established rules

for what it calls Synthetically Generated Content, or SGI.

And this is defined as information that is artificially

or algorithmically created, generated, modified,

or altered using a computer resource in a way

that appears reasonably authentic or true.

This framework puts a burden on both the user

and the platforms that are displaying this content.

So users of those platforms have to declare

if their upload is SGI.

That's kind of one issue right there.

Then the platforms must use technical tools

to verify that declaration.

So if the content is SGI, it has to be clearly marked.

And that clear mark is covering 10% of the surface area

for visual content, or the first 10%

of the duration for the audio.

So this is applying to all content

from creative work to those more malicious deep fakes.

Now, China has also implemented new requirements.

AI generated content must be labeled with a watermark

that covers at least 5% of the shortest side

of the content, along with metadata

that tags that content.

So the tech platforms there are legally required

to enforce this measure.

The EU has its AI Act now,

which applies to any provider in the EU market.

So most generated content is falling

into what they call a limited risk category.

And according to IBM,

this category imposes transparency requirements,

which means disclosing that content that's AI generated.

And I'll put that article in the description.

So this obligation is on the provider of the AI system,

while the disclosure of the,

say deep fakes of the generated content

is on the person who is deploying that content.

So the USA and UK have very little in place right now.

Now in Canada, a new minister of AI has been appointed,

but new legislation is still just in the planning stage.

It seems like old legislation was kind of scrapped

when this change occurred.

So this is now restarting with the new Canadian government.

This kind of inaction or slow action

from many of the Western countries leads to many court cases.

That's what happens when there isn't really

legislation in place.

So that's what makes the recent news from Australia

pretty important and interesting.

On October 27th, just a couple of days ago

from the time of recording,

the Australian government confirmed

that it is ruling out a fair use style of exemption

for AI training, for that ingesting of all of IP

into LLMs.

And that exemption of fair use

has been what corporations have been pushing for.

They want that exemption so they can use all the stuff.

So this is a really positive sign

for creative professionals and all creatives anywhere.

So the government's position is that AI models

using copyrighted works for training

already require a license

under current existing Australian law.

So it doesn't sound like Australia is really interested

in creating a new loophole for the tech companies to use.

And Australia is moving in a direction

that supports and protects artists and creative professionals.

So this is a good thing,

which is why it's also so important to voice concerns

and advocate for creative protections of IP

remaining in place for the future.

So the global regulatory focus is mostly,

not entirely, but mostly on labeling AI content,

on transparency and declaration,

if anything is in place at all.

But that focus really kind of misses

the main issue for creators.

So a good example of what's happening right now

is the musician, the YouTuber, Davey50.

He discovered that an AI service,

which was kind of implied in those videos,

I'll put the links in the description,

you can check them out,

allowed users to upload his copyrighted recordings

into the service.

And then they used those stolen songs

to generate new derivative songs.

And this example is just one of many

of why creative groups are lobbying governments.

So in Canada, the Coalition for Diversity

of Cultural Expressions, or the CDCE,

is pushing for three measures.

And one of these measures is a requirement

for AI developers to disclose the data used

to train their systems.

And I'll put that source in the description.

So we see this sort of thing happening in the US as well.

So the WGA strike that was fairly recent

led to contract protections for creatives.

SAG-AFTRA and the RIAA have been lobbying hard as well.

SAG-AFTRA for protections against deep fakes

and the RIAA for copyright protections.

And the WGA strike, for example,

won those protections that AI cannot be used

as a source material or to rewrite scripts.

And this protects writers' credits

and compensation as well.

So speaking of writing, join my free newsletter

at humaninternettheory.com

and I'll send you some real human writing.

Mine, and I'm working on improving my segue.

So you're gonna see different ones all the time.

So because we don't have many or any regulations

in place for ingesting all of the data

and creative assets that lead to a whole bunch

of unauthorized derivative work,

creators are often taking matters into their own hands.

And as such, this has led to a whole bunch

of class action lawsuits.

According to the website chat,

GBT is eating the world.

that are filed against AI companies worldwide.

And 56 of those are apparently cases just in the US.

I'll put a link to that in the description, of course.

So clear labeling of synthetic content is important.

That helps the humans consuming content.

That's all of us.

It helps us identify human-made creations

when it's now getting really hard to tell what is what

and what's real, what's not, what's a deep fake,

what was created by hand versus computer.

Now that copying is so accessible and easy

to everyone to do.

The labeling also demonstrates to the internet people

consuming all of this stuff,

just how much content is being generated

without much of any human oversight.

And some estimates are that it's already over half

of the entire internet.

And this is going to allow people, the humans,

to make choices about what content they even engage with.

Like if you can't be bothered to even read

what you've pasted onto the internet,

why do I want to consume it, right?

But a regulatory focus on labeling,

that misses a couple of things.

Like will people even use or declare what they're creating?

For example, watermark removers are very common,

very popular in the app stores right now.

And will the platforms even disclose what is declared?

An investigation by The Washington Post

tested eight major social platforms

by uploading a video containing the standard metadata

to flag synthetic content.

And only one platform correctly added a warning label

to let people know that the content was generated by AI.

So there's still a lot of issues

in the labeling conversation.

So that makes the more pressing and pertinent issue

in all of this the copyright part of it.

Derivative works are being created for material

that was never intended to have derivatives.

And access to something on the internet, as we know,

does not give anyone the right to create derivatives from it.

And even if derivatives were allowed in the licensing,

the current usage is often incorrect

because credit is often required as part of that license.

And then it has to be released under the same licenses again.

So the 82 lawsuits show what happens

when regulation isn't in place or the regulation fails.

And this is exactly why the Australian government's

decision to enforce existing law

is pretty darn refreshing to hear this week.

So the copyright owners are being forced

to take on this fight themselves in the courts

because lawmakers are kind of focusing

on the output side of the thing

and are not working on the input, the copyright issue.

And as a creator, this input issue is what I'm concerned about.

Obviously, by this point, the part where our assets

are being used by these LLM corporations without asking,

and those LLMs are allowing all of these derivatives to happen.

And that is the real fight here.

AI companies are building their models on an opt-out basis

quite often, even if they're allowing the opt-out,

even if that opt-out works, which is why we're seeing

how the Sora2 app came out with the opt-out

and there's a whole bunch of derivatives

still being created from that material.

And I'll put a link in the description

to the video I did regarding that.

So this is where the LLMs are kind of taking all the work

and allowing these derivatives

and the creatives need to fight instead

to have it removed from the dataset after the fact.

That's broken.

And that's not how things generally work.

So I believe we need to be advocating for this opt-in model,

like we're seeing Australia support,

which is the standard for all of our licensing in the past.

One more thing to think about is how addressing the input,

the ingestion, the scraping, the copyright stuff,

that also could help what the lawmakers are trying

to address with this labeling.

Lawmakers want to stop the harms on the output,

like the misinformation and the deep fakes.

That's what they want.

And that could be accomplished to some degree at least

by controlling or checking the training data

that's being used on the way in.

And that work can involve checking for the bias

or any dangerous material that causes the output harms.

Anyways, all of this work can be opted into ethically

and controlled more.

And these kinds of models can then be used

to create highly targeted, useful applications.

And then those tools can be made

that are very specific and useful and benefit humans

and are on a smaller scale and use less energy

and all that kind of thing,

as opposed to the services now

that reduce the number of apps and tools

that are available to us for creativity.

'Cause there's services that just like replace

the tools entirely.

But let's get back to us creative professionals now.

So unions and organizations such as WGA and RIAA

have the money and the resources to fight

for the artists that they represent.

And those are big organizations

with collective bargaining power and money.

So what about independent YouTubers?

And individual writers and visual artists or podcasters?

What about all of us that don't have those resources

or organizations behind us?

We can't afford to launch one of those 82 lawsuits.

So therefore the input, what the LLMs are taking

from us as well is kind of like the small guy

versus big guy situation.

And that's why we need regulation from the top down

like in Australia.

The current system of forcing the fight

onto the court systems through these lawsuits

only helps larger artists and people who can afford it.

And many of us independent creators

are kind of left out of the conversation.

And I mean, assuming that the court cases

actually succeed in the end, right?

So right now the government, some governments,

if anything are focused on the outcome, the symptom,

the deceptive content like the deep fakes,

but us creators are focused on what could be

an actual solution to that,

which is the unauthorized training data

is being controlled more.

And this is a solution that at least one government

is starting to recognize and that's great.

And I'll be back soon with another episode.

Bye for now.

Episodes are written, directed, edited and produced

by Jen deHaan of STereoForest.com

Find out more about this podcast

and join our free newsletter for additional resources

at humaninternettheory.com.

Find additional videos at the YouTube channel

called Human Internet Theory.

Links are also in the show notes.

(upbeat music)

(upbeat music)