The Impact of Generative AI on Creative Industries and the Need for Protection

I think these industries you know if you allow AI companies without payment or permission to build a highly scalable

competitor that's much cheaper and quicker to use it's clear that a lot of these industries are going to be totally

gutted we're here to talk about generative AI how they train uh their models and how that's impacting creative

communities and both of you have been working on this issue and are tackling it from um similar yet very

complimentary angle Luke you're a filmmaker tell us a little bit about how you got into this issue and what overlay

is doing yeah so um I got my start on YouTube I love kind of that uh Creator economy uh people can just upload from

anywhere and get followers and get connected to Their audience and uh I started to see this wave of um

disruption coming from Ai and and the way it's been set up in the way they've been training and so I kind of banded

together with our co-founders Paul nckin and Christina midm who are you know top National Geographic photographers and we

started to just think about like how this was going to play out and then what tools we could build now to kind of get

ahead of it and and protect our ourselves really we started there and then we started to go like how could we

expand this to give it to more creators and more creators so overlay is really just uh an app that assigns all the

necessary metadata and new metadata as it comes along in this kind of evolving world uh to give creators as much

protection as as we can possibly provide some of who may be familiar with how generative AI trains their models some

may not Ed can you kind of walk us through as a former AI executive yourself can you walk us through how

training works yeah I mean so at a high level um these generative AI models only work because of the content that they're

fed um so we call it in the industry it's kind of termed learning it's actually like pretty different to human

learning but you know it's a helpful metaphor I mean essentially you what a lot of AI companies do is they gather as

much content as they can um generally from scraping the internet um they download all of it and they and they

train a model to be able to create you know new work that is in the style of that work essentially and of course A

lot of the time um a lot of the time you know there's sort of regurgitations of what went in and that sort of thing but

ultimately the the intention I think is is to try to create um you know work in that style and so this is true across

text generation uh you know things like chat GPT it's true across art generation things like mid Journey it's true across

um music generation uh and it's really true across the creative Industries the same kinds of models are being used

across the board and and unfortunately you know the same kind of approach is being used by a lot of AI companies

pretty suddenly just in the last couple of years or so you know which is basically an approach that says we're

just going to gather as much information as content as we can you know and we're going to and we're going to train our

models on that and and yeah that that's roughly how they're doing it right and traditionally you'd have to license that

work like if you were using you know a photograph from nichan you'd have to license that work I mean in basically

every instance in the past yeah you this is kind of like the first time I've witnessed it at this scale towards just

like no we're not we're not going to license anything so and it's funny I mean yeah I I totally agree and it's

funny like for years I've been in generative AI for since long before we called it generative AI I started in

2010 and for really a long time up until about 2022 No One released commercial models

trained on copyrighted work basically it was understood that you would not do that because you'd get sued um and then

in 2022 you know these people have been working on Research models behind the

scenes and they started working really well and and a few companies I guess decided to release them and they're

companies you know we've all heard of um and then you know those companies I think did so well and raised so much

money that there's been this total rapid domino effect where just the entire AI industry has kind of followed suit and

just thought well if they can get away with it maybe we can and and that's cuz like lawsuits take time to put together

right I mean the first ones came in pretty quickly but you know and so what you have now is an AI industry that that

very suddenly has made scraping copyrighted work and using it to build you know what's essentially a

replacement for much of that work you know as you know has adopted that as basically the norm and as Luke says like

you know generally in the past when commercial entities have needed to use copyrighted work to you know to build

something they'd license it right they'd pay for it so yeah I mean this is this seems to me to be a very different

approach from from prev Technologies and the reason they can get these these you know the reason we've

seen these hyper realistic images is because they're based on photography they're based on on photographers work

talk us through what we're seeing right now and the exper the kind of the tests that you ran so I think there's like a

couple things to to dive into here but but one of them is we try to keep it really simple just to show different

photographers what was going on so we kept it to just like the photographer's name and then the word photo so in this

case it was Paul nckin photo was the only prompt I didn't say anything about Wildlife I didn't see anything about you

know the Arctic or anything like that and that's what it gave back so to us that was a pretty clear indicator that

his content's all in there in some form when we get into their work and and the work that they do like they're

conservationists they're always going and trying to capture you know the current state of of different climates

uh species different things like that the first thing Paul noticed was like that's a spirit bear on the right the

fake one the bottom right but with the polar bear's feet uh simply because like in the training data they didn't have

you know enough training data to properly put together the spirit bear it just kind of pulled the spirit bear and

then threw some polar bear feet on there I'm sure Ed could dive into that a bit more but we just found that super

interesting like that could lead to all sort of terrible things terrible outcomes right

yeah that's fascinating because that first look like I I you know not having any knowledge of the specificities of

the polar bear family I would have not known that had you not told me and tell us a little bit you know I'm curious

about this is a unique style the the photographs on the left Paul nichan he's known for a very unique style of

Photography and it takes a lot of money yeah even say with his style it's more about the time and the effort that goes

into capturing it which he he gets into it a bit but it could be as much as six months right like trekking to this

location uh spending days on end not getting anything useful and then being there in that in that perfect moment to

capture that um that's obviously very valuable stuff and the business model that he relies on to go out and capture

that data is fragile it's not like bulletproof um so if we start messing with these new business models that are

you know fairly new in the social media age um algorithms start to change the internet's flooded with AI content

they're going to be replaced pretty quickly just from an exposure standpoint an algorithm share standpoint so I think

AI companies really need to think long and hard about like how long they're going to need uh organic data and if

they want to be messing with these fragile business models of the people out there capturing this stuff you

strongly believe that they were trained on on work that would have been licensed and in this case the this photographer

Paul nckin and the others we have other examples as well but the others did they receive any compensation for this no and

I mean another good example is anel Adams you can type anel Adams photo um and and get his stuff back they are very

strict with where they license his work and it's never been cleared for that use so

yeah there's no payment and then the companies that do this are are kind of receiving the bulk of that payment yeah

these are just some of the other tests that you ran and you know Ed I think this is a good segue into the big push

that you've made uh with your statement and now 50,000 creatives signing

on I mean this is exactly what the statement is talking about yeah I think so I mean you know

it's important to remember that while these examples are obviously totally egregious and you see them across

domains you see them in text you see them in music you know you see pop music coming out of some of these systems

sometimes um while they're egregious the the there's a broader issue here which is that even when these things aren't

creating something in the specific style of a certain artist you know what they're still doing is they're still

taking you know the basically the world's creator life's work and they're ingesting that and they're using that to

build a crazily scalable competitor to those creators and to that work right and there's a bunch of examples already

where you know people are losing work and their work is less valuable because AI companies have taken it and have this

replacement machine so you know so so even outside of style regurgitation like you know like this which is obviously

awful there's a broader issue here and yes so you know this statement that we put together is meant to just be a very

short simple statement um you know because I think short simple statements can be powerful um and just as meant to

show the extreme strength and breadth of feeling you know around this question at the moment we hear a lot from AI

companies who have a lot of money to Lobby governments uh right and get the word out it's generally harder for

individual creators to have their voices heard recently in the UK the UK government has actually suggested

upending copyright law and allowing AI companies access to all this content at the moment it's illegal in the UK you

know I had a bunch of creators come to me and say what can we do about this right people you know these are

professional creators maybe not many people have heard of them but they're saying look this is they they're

stealing our work and they're going to put us out of a job by doing this what can we do and honestly it's hard to know

what to do but I think one thing we can do is we can shine a light on this View and we can say well look every time an

AI company says hey what we're doing is good for creativity is good for creators you know we can say well is it really

like are you paying them are you asking permission like to build this competitor to them using their work obviously or

not is it too late now since the damage has been to a certain extent done what can be done moving forward I I mean I

can't speak for what people are hoping for apart from the fact they've signed this statement um and I'm sure there are

a lot of different views among the 15,000 and Counting signatories here there'll be people who just don't like

AI at all there'll be people who like Ai and want to use it but want to use you know licensed models there'll be a whole

range of views um you know personally what I'd like to see is I'd like well for starters I'd like governments not to

change cop change hundreds of years of established copyright law to introduce exceptions to copyright law that allow

AI companies to take creators life's work and build competitors to them I mean that that feels like table Stakes

here and I'd like to see you know governments not introduce that um you know I'd like I'd like that to be the

broader way the discussion goes and I ultimately I'd like to see more AI companies starting to pay creators and

like I'd like to get a fairer deal for creators in the age of generative AI is it too late I don't think so I mean

again like it can feel like it's too late because so many people use this product or that product like open AI is

worth $157 billion dollar it must be too late but in simple time scale terms we're talking about two years like

things that like and it always takes regulators and you know and and frankly the courts time to respond to things so

I I actually don't think it's too late and I think if if if you know if we make it as clear as we can that you know the

creative Community does not accept this and does not accept it as fair in any way you know maybe maybe we can affect I

mean who know who knows you know but but all you can do is try and Luke you know as we as these creatives await for

regulation or legislation and and and the law to to kind of come um pick up with the times

overlay has a solution right now for photographers who are posting online can you walk us through how it works I yeah

I I think we do I mean we're trying to right so uh what we're essentially doing is taking all of the available metadata

that is out there that could help photographers or creators and that comes in the form of c2p which we're starting

to see a lot more adoption of like uh Google has said they will recognize c2p metadata in Google search and YouTube

and we're starting to see that roll out uh and then we Al but we're not committed to just you know one bit of

metadata we want to assign copyright we want to do IPC um and we want to Stack all of these

statements up and put them in one place and take the best available statements for photographers and put them in one

place but like the underlying issue has been that metadata is consistently stripped uh as soon as you do anything

with your cont you uploaded to a website uh any statements that you might have attached to that piece of data are gone

um and so these standards are important because they are starting to establish uh you know ways to keep that metadata

attached to the file and if it does get stripped different ways of like calling it back and going hey I found this

picture uh I don't know who owns the picture who took it and here's a way I can find that out um it it seems like

honestly it seems like uh work that should have been done on the internet a long time ago but it's you know the

generative AI push has kind of Taken everybody uh collectively to come together and go we should probably

figure this out and everybody has their own reasons for doing that but what I like to see is that it's it's happening

and and I think creators can benefit from that work and the content credential c2p is about is really about

transparency and traceability giving users online the ability to understand where a piece of content comes from in

terms of overlay how are you hoping it might help uh creatives especially photographers you know Paul nckin

is is joining you to to help launch the app or as part of as part of the the effort how would practically how would a

photographer be able to protect his work using overlay like we tried to make it as simple as possible it's it's an

iPhone app you simply just take your photos from your library on your phone uh add the protections and then it's

ready to share online obviously uh different companies like Google uh they look at the metadata directly and then

we also have an invisible Watermark that's fairly robust so that if it does get stripped on one of these social

media apps like Twitter Twitter's fairly well known for just like discarding all that stuff we can attach it back to the

file via our Watermark so it's pretty simple um we tried to make it uh easy and for everybody to use but it's not

bulletproof right like I think a lot of these things I don't know if there will ever be a bulletproof text solution so

it's kind of like with what Ed's doing it's on us to kind of keep up on it and apply the pressure on multiple fronts if

we actually want to do anything here and you have a do not train component to it right yes so there is a do not train

component um and that's using the c2p uh there's also on on the IPC metadata side I believe they might have a do not train

component as well but as you can see there on the b-roll you know we we stack that metadata and we stack those Asser

ions um instead of just you know putting all our eggs in one basket uh one standard we want to adopt M multiple and

we think more will come you know it's looking like potentially uh Sony and Japan might have their own so that would

be something we would add yeah it seems like we're moving towards a world where we're going to see more of these digital

standards and more of this uh transparency and the reason that I you know the reasoning for these companies

adopting it might not be this you know they're not doing it for ethical reasons they're doing it for house housekeeping

but at the end of the day like I said I think creators can take advantage of that and then Ed you have a solution uh

when it comes to to data sets um these a company if they wanted to use a data set that was based on licensed work you

launched a nonprofit called fairly trained tell us more about how that works yeah so fairly trained is really a

trust Mark it's a certification badge um we give it to companies AI generative AI companies that don't train on

copyrighted work without a license um we've certified 18 companies today and really it's you know you know it's it's

to try to elevate these companies and show that there is an alternative you don't have to go and scrape copyrighted

work and use it without permission so that's what we're doing and I think that um you know and I think as Luke says you

need all sorts of things like you need you need protections uh you know like overlay and others you need these things

as well you need certification badges I mean one of the things that you know I think a lot about is you do need you

know you need the ability to opt out of training right now unfortunately because you know everyone is just training on

your work you know this in itself is totally unfair the burden should not be on the Creator to use metadata to opt

out of AI training you know opt out I mean opt outs for better or worse you know tend to get missed most people

never realize they can opt out of this training I've run opt out schemes for AI companies and even the most well-run

schemes get missed by most people because you know frankly you never hear about it you never see the email if

that's how it works like maybe there are a few different protocols you can use you don't quite know how they work you

don't know which one to use you know I think if a government really thought that AI training was kind of a good

thing for creators you know it wouldn't create an uptown scheme you create an uptown scheme if you you know you

basically know that no one's going to go for this and you want it to happen anyway you know if you think it's a good

thing for creators you create an optin scheme because they'll all be jumping at it and the fact that um you know AI

companies don't tend to be advocating for that I think tells you all you need to know about whether they think

creators want this or not my last question to both of you is if we don't get to this right if we don't get this

right what happens to the creative community and this is a global issue it's not just the US UK Europe right I

kind of always approach this question of like I think you need to look beyond that and go so we have to get it right

to set precedent for how AI will interact with every profession right and so I I think it's even a little bit

bigger than that but I I do feel like um I'm optimistic I I do feel like creators will bounce back a lot of this will

bounce back but it won't happen if nobody does anything right it does require uh people putting at work people

organizing um but I'm I'm fairly optimistic that things will bounce back I think I do think if we don't get this

right then you know a lot of the creative Industries are in a lot of trouble I think you're basically going

to see a lot of the current creative Industries you know especially the parts you know the probably the vast majority

that come from people who aren't necessarily household names right you know be it background music composing be

it much photography be it lots of illustration be it lots of writing and journalism you know I think these

industries you know if you allow AI companies without payment or permission to build a highly scalable competitor

that's much cheaper and quicker to use it's clear that a lot of these industries are going

to be totally gutted you know so I I I think we have to get it right for that reason