Creativity in the Age of AI: Generative AI Issues in Art Copyright & Open Source

Stanford HAI
1 Jun 202341:23

Summary

TLDRThe panel discussion delves into the intersection of open source, copyright, and generative AI, highlighting the challenges and opportunities these technologies pose. Scott Draves, an AI artist and engineer, shares his experiences with open source generative art and its impact on the art world. Pam Samuelson, a law professor, addresses the legal complexities surrounding copyright infringement and the use of copyrighted works as training data for AI. The conversation explores the potential of generative AI to transform personal expression, education, and creativity, while also considering the need for trust, accountability, and the role of open source in democratizing AI technology.

Takeaways

  • 🎨 The panel discussion focused on the intersection of Open Source, copyright, and generative AI, highlighting the complexities and ongoing legal debates surrounding these issues.
  • 🤖 Scott Draves, an AI artist and engineer, discussed the history of generative AI and its evolution from his personal experiences, emphasizing the importance of open source in enabling artistic and technological innovation.
  • 🌐 Draves mentioned his early work, including the 'flame' algorithm, which became a foundational piece in generative art and was an example of the collaborative spirit of the open source community.
  • 🚀 The development of AI technology has seen a significant leap in recent years, with implications for personalized education, new forms of personal expression, and interactive media.
  • 🧠 Draves argued for the necessity of open source models in AI, stating that having transparent and accessible code, data, and models fosters trust, independent research, and equitable distribution of AI benefits.
  • 📚 The talk by Pam Samuelson, a law professor, addressed the copyright challenges posed by generative AI, noting that there are active lawsuits and global conversations about the governance of AI.
  • ⚖️ Samuelson discussed the five exclusive rights granted by copyright and the limitations and exceptions to these rights, such as fair use, which is a defense often raised in cases involving AI and copyright.
  • 🔍 The issue of whether AI-generated outputs infringe on derivative works was highlighted, with ongoing lawsuits challenging the boundaries of copyright law in the context of AI.
  • 🎵 The discussion touched on the potential legal ramifications of AI-generated music covers, noting the complexities of copyright infringement versus creative expression and the right of publicity.
  • 🌐 The panelists acknowledged that the legal landscape for AI and copyright is still developing, with technology often outpacing the law and requiring adaptation and new interpretations of existing legal frameworks.
  • 📖 The Warhol vs. Goldsmith case was mentioned as an example of the ongoing legal discourse around fair use and transformative works, with implications for the use of AI in creating new works based on existing content.

Q & A

  • What is the main focus of the panel discussion in the transcript?

    -The main focus of the panel discussion is on Open Source and copyright issues related to generative AI, particularly in the context of art and its legal implications.

  • Who is moderating the panel and what are their background and roles?

    -The panel is moderated by Professor Russ Altman, who is the Kenneth Fong professor and professor of bioengineering, genetics, medicine, and biomedical data science. He is also an associate director of Stanford Hai.

  • What is the significance of having a bioengineer faculty moderate the panel on generative AI and copyright?

    -The significance lies in the interdisciplinary nature of the topic, as generative AI intersects with both technology and legal domains, and a bioengineer brings a unique perspective that can enhance the understanding of these complex issues.

  • What is the background of Scott Draves, the first speaker introduced in the transcript?

    -Scott Draves is an AI artist and engineer known for his work in open-source and generative art since the 80s and 90s. He is recognized for creating the 'flame' algorithm, which became a significant visual style and influenced a whole genre of art.

  • What is the Electric Sheep project mentioned by Scott Draves?

    -Electric Sheep is an infinite, evolving AI screensaver created by Scott Draves that uses feedback from its audience to learn and generate animations, representing an early example of collective intelligence in AI.

  • What are the four key tenets of open source that Scott Draves discusses in relation to AI models?

    -The four key tenets are: the code for training the models should be open, the inference engine code should be open, the training data should be open, and the model weights should be open.

  • What is the main concern regarding the use of copyrighted works as training data for AI models?

    -The main concern is whether ingesting copies of copyrighted works as training data constitutes an infringement of copyrights in those works, as most content on the open web is protected by copyright law.

  • What is the significance of the fair use defense in the context of AI-generated outputs?

    -The fair use defense is significant because it allows for the non-infringing use of copyrighted works under certain conditions, such as for educational purposes, criticism, commentary, news reporting, and research. It is often invoked when discussing the ingestion of copyrighted works as training data for AI models.

  • What is the role of the United States Copyright Office in addressing generative AI issues?

    -The United States Copyright Office is holding listening sessions to gather stakeholders' interests and concerns about generative AI and copyright. It will publish a report with recommendations, potentially by the end of the year, to guide its approach to these issues.

  • How does the concept of 'forking' and fine-tuning of AI models contribute to the diversity of truth and beauty in AI-generated content?

    -Forking and fine-tuning allow individuals or groups to take a foundation model, add more data, and change its character, which can lead to different notions of truth and beauty. This enables a variety of perspectives and cultural expressions in AI-generated content, rather than a single, uniform output from a single company's model.

  • What is the potential impact of generative AI on personal expression and creativity?

    -Generative AI can lead to new forms of personal expression, such as easy-to-use Photoshop alternatives that respond to user commands, interactive fiction, and personalized content creation. It has the potential to democratize creativity and enable individuals to express themselves in novel ways.

Outlines

00:00

🎨 Introduction to Generative AI and Open Source

The speaker begins by expressing gratitude for being part of the panel and introduces Scott Draves, an AI artist and engineer, and Pam Samuelson, a law professor and co-director of the Center for Law and Technology. The discussion focuses on generative AI, its history, and its implications on art and copyright, particularly in the context of open source. Scott shares his journey in creating open source art since the 80s and 90s, highlighting the transformative impact of his work on the art world and the concept of collaborative creation.

05:03

💡 The Evolution of Generative AI and its Impact

The speaker delves into the evolution of generative AI, from his early attempts at creating algorithms that generate images to the current state of the technology. He discusses the potential applications of AI, such as personalized education and assistance, and emphasizes the importance of open source in democratizing AI. The speaker argues that open source enables customization, trust, and widespread adoption of AI, while also addressing the challenges of trust and accountability in proprietary models.

10:04

📚 Copyright Challenges in the Age of Generative AI

The speaker shifts focus to the legal and policy challenges surrounding generative AI, particularly copyright issues. He outlines the current lawsuits against generative AI, the concerns of policymakers, and the ongoing discussions about the role of copyright in the digital age. The speaker explains the basics of copyright law and its limitations, and discusses the fair use doctrine as a potential defense for using copyrighted works as training data for AI.

15:05

🤖 The Role of Generative AI in Creative Reuse

The speaker explores the role of generative AI in enabling creative reuse of copyrighted works. He argues that while some view AI-generated outputs as infringing derivative works, others see it as a tool that promotes the progress of science and culture. The speaker also touches on the complexities of defining 'original expression' in the context of AI and the potential for AI to create new forms of creative expression that do not directly copy existing works.

20:06

🔍 Navigating the Legal Landscape of Generative AI

The speaker discusses the ongoing legal disputes surrounding generative AI, particularly in the music industry. He highlights the challenges of regulating AI-generated music covers and the potential implications of right of publicity and false representation laws. The speaker emphasizes the need for a nuanced approach to copyright and intellectual property in the context of AI, recognizing the balance between protecting creators' rights and fostering innovation.

25:08

🌐 The Future of Generative AI and Copyright

The speaker concludes by speculating on the future of generative AI and copyright. He addresses the potential impact of AI on the viability of fair use claims and the need for a balanced approach to legal frameworks. The speaker suggests that while existing laws may need to evolve to accommodate AI, the focus should be on fostering innovation and creativity while respecting the rights of creators.

Mindmap

Keywords

💡Generative AI

Generative AI refers to artificial intelligence systems that are capable of creating new content, such as images, music, or text, without direct human input. In the video, the speaker discusses the history and development of generative AI, its impact on art and copyright, and the potential for open-source models to democratize access to this technology.

💡Open Source

Open source refers to a philosophy and practice of allowing users to access, use, modify, and distribute software freely. The speaker advocates for open-source principles in AI development, emphasizing the importance of open code, open data, and the ability for users to fine-tune models to their needs.

💡Copyright

Copyright is a legal right that grants creators exclusive control over their original works. The discussion in the video revolves around the challenges that generative AI poses to traditional copyright concepts, including the use of copyrighted works as training data and the potential for AI-generated outputs to infringe on derivative works.

💡Fair Use

Fair use is a legal doctrine that allows limited use of copyrighted material without permission from the rights holder, typically for purposes such as criticism, commentary, news reporting, teaching, scholarship, or research. The video explores the application of fair use to the training of generative AI systems and the use of copyrighted works as data.

💡Creative Commons

Creative Commons is a non-profit organization that helps creators share their work under certain standardized licenses that allow others to use the work under defined conditions, such as attribution, non-commercial use, or no derivatives. The video discusses the position of Creative Commons on the use of their licensed works in AI training data.

💡Data Licensing

Data licensing refers to the legal agreements that define how data can be used, shared, and monetized. The video touches on the increasing recognition of data's value and the creation of licensing agreements for using data in AI training, which may impact the application of fair use.

💡Right of Publicity

The right of publicity is a legal right that allows individuals to control the commercial use of their name, image, likeness, or other recognizable aspects of their identity. The video highlights the potential for this right to come into play with generative AI, especially in cases where AI-generated content might use a celebrity's voice or likeness without permission.

💡Foundation Models

Foundation models are large-scale AI models that are pre-trained on vast amounts of data and can be fine-tuned for specific tasks or applications. The video emphasizes the cost-effectiveness of fine-tuning these models and the potential for diverse customization by different users or groups.

💡Trust and Accountability

Trust and accountability pertain to the reliability and responsibility associated with AI systems, particularly in terms of their transparency, the ability to audit their operations, and the consequences of their actions. The video discusses the importance of open-source AI in fostering trust and accountability by allowing independent research and third-party audits.

💡Liberal Licensing

Liberal licensing refers to the practice of granting permissions for the use of a work with few restrictions, often allowing for commercial use and modification. The video highlights how liberal licensing can encourage the spread of AI benefits and motivate investment in open-source projects.

Highlights

Discussion on Open Source and copyright in relation to generative AI.

Introduction of panel moderator, Professor Russ Altman, and his diverse academic and professional roles.

Scott Draves, AI artist and engineer, shares his journey in open source and generative art since the 80s and 90s.

The significance of 'flame number 149' as a pivotal piece in the history of open source artwork.

Generative AI's role in enabling artists and fostering meta interactivity through programming and internet collaboration.

The Electric Sheep project as an example of collective intelligence created 24 years ago and its ongoing relevance.

Scott Draves' perspective on the rapid advancements in AI's understanding of human language and image generation.

The potential implications of generative AI, including personalized education and new forms of personal expression.

The importance of open source in AI development, including transparency, trust, and the ability to fine-tune models.

Pam Samuelson discussing copyright challenges posed by generative AI and the ongoing global conversation.

Three active lawsuits challenging generative AI on copyright and related grounds, and the U.S. Copyright Office's efforts to address these issues.

The critical questions around ingesting copyrighted works as training data, AI-generated outputs as derivative works, and removal or alteration of copyright management information.

The role of fair use in the context of using copyrighted works for training data, and its legal precedents.

The distinction between AI's interest in facts versus original expression in a work, and how this impacts copyright considerations.

The potential market competition between AI-generated images and those created by human artists.

The impact of generative AI on the rights of publicity and the potential legal issues arising from the use of samples of voices or likenesses.

The Warhol versus Goldsmith case and its limited impact on the broader discussion of AI and copyright.

Assessment of the viability of fair use claims in the context of AI, given the increasing commercial value of data.

The importance of forking and fine-tuning in open source models to allow for diverse interpretations of truth and beauty.

Transcripts

play00:01

that was great and we're going to

play00:03

continue on in this next panel talking

play00:06

more about Open Source and copyright so

play00:08

I want to first introduce our panel

play00:11

moderator uh Professor Russ Altman Russ

play00:14

is the Kenneth Fong professor and

play00:17

professor of bioengineering of genetics

play00:20

of Medicine of biomedical data science

play00:23

and by courtesy of computer science he's

play00:27

also one of the associate directors of

play00:29

Stanford Hai and uh when I met Russ I

play00:32

felt my long-lost brother so thank you

play00:34

Russ

play00:36

thank you thank you James uh this

play00:39

session is called generative AI issues

play00:41

in art copyright and open source so

play00:44

what's more natural than having a

play00:46

bioengineer faculty moderate it

play00:48

so I'm really happy to uh very uh

play00:51

briefly introduce Scott Draves who will

play00:53

speak first he's an AI artist and

play00:55

engineer and Pam Samuelson will speak

play00:58

after Scott she's a professor of law at

play01:01

Berkeley law school and co-director of

play01:03

Center for Law of the Center for Law and

play01:05

technology so um with those brief

play01:08

introductions I'll just throw it to

play01:09

Scott

play01:15

thank you uh and thank you thanks for uh

play01:19

all your attention this morning it's uh

play01:22

it's an honor to be here so a lot of

play01:25

people have a take on on generative Ai

play01:28

and uh before I give you yet another one

play01:33

um just uh where do you know where does

play01:36

my where does mine come from and where

play01:37

do I come from and uh I was trying to

play01:41

think of you know what I can best

play01:43

contribute here uh today and I thought

play01:46

one thing is just you know where where

play01:48

is where does generative AI come from

play01:50

what's the the history here

play01:52

and so this image here's an example uh

play01:56

you know I've been doing this open

play01:58

source and generative uh art since the

play02:01

80s and 90s and uh this is uh this image

play02:05

is uh flame number 149 uh from 1994

play02:10

where it got a an award at the pre-arse

play02:13

Electronica and is really what made me

play02:16

realize I you know I was an artist

play02:17

before uh that was

play02:20

and um

play02:22

this uh was perhaps the you know the

play02:25

first open source artwork and it's was

play02:28

really very much it wasn't just an image

play02:30

it was it was code that allowed other

play02:33

artists to create their own images

play02:35

and this artwork was very much about

play02:38

enabling other artists and um the sort

play02:42

of meta interactivity of uh you know

play02:45

programming exchanging code over the

play02:47

internet and uh you know building things

play02:50

together

play02:51

and so this flame algorithm uh as as it

play02:56

became known created a whole genre and

play03:00

visual style and eventually I couldn't

play03:02

even go into a bookstore or like a look

play03:05

at a magazine rack without without

play03:07

seeing some version of it and of course

play03:09

so

play03:11

uh

play03:14

there we go here we go

play03:16

uh so uh here's an album cover uh that

play03:21

used it and of course these were all

play03:23

images uh I didn't get any sort of

play03:26

credit credit for this my name isn't

play03:28

isn't here it was made by another artist

play03:31

just using the algorithm

play03:33

and um you know really what I learned

play03:35

from this was uh you know I put this

play03:39

code out there with a plan

play03:41

um

play03:42

but what I learned was that uh sort of

play03:45

giving up control uh the real power

play03:48

isn't sort of getting what you want

play03:50

getting other artists to use it or to

play03:52

contribute to give me code back to uh

play03:55

but really getting things that I didn't

play03:58

know I wanted or or I didn't want and

play04:00

enabling other people to do things uh

play04:03

that were really unexpected

play04:05

and so just uh by in 1999

play04:09

uh the Electric Sheep was the I created

play04:13

this infinite evolving AI screensaver

play04:16

you know which used feedback from its

play04:19

audience and so it learned from

play04:20

everybody who was watching it and it

play04:23

tried to satisfy that that human desire

play04:27

and it generated animations

play04:30

excuse me and um this was a collective

play04:34

intelligence created uh 20 24 years ago

play04:36

and actually still running and so this

play04:39

this image

play04:41

sorry uh here we go

play04:45

and so this image is actually a still

play04:48

frame from a Super Bowl commercial that

play04:51

IBM and uh H R Block Rand so

play04:56

it definitely went places that that I

play04:59

didn't expect and and didn't plan

play05:02

so but um but for most of my career I

play05:05

was you know I was a regular Tech guy I

play05:07

got the PHD in computer science with

play05:10

with James

play05:11

uh and you know worked on regular

play05:16

technology in unrelated to uh but I had

play05:19

a you know an interlude in my career for

play05:22

for five years I worked you know

play05:23

full-time as an artist

play05:25

uh just doing open source stuff

play05:29

and I'll show one more uh much lesser

play05:34

known

play05:35

example here

play05:37

of my sort of super talking about

play05:40

history this is sort of like a super

play05:42

early attempt at something that is now

play05:44

like stable diffusion uh so this is a

play05:49

generative algorithm that takes a an

play05:52

image as input and sort of uh and

play05:55

generates stuff that looks uh based on

play06:00

based you know is modeled after the

play06:02

inputs and so

play06:03

as you can see the the here's an attempt

play06:05

at doing hands it's it still has the

play06:08

same problems after all these years

play06:11

oops sorry

play06:13

there we go

play06:16

uh yeah this was this was 1993 though so

play06:20

and so all the these were uh I'll

play06:23

predate you know uh the adoption of

play06:25

neural networks we we have today

play06:28

so

play06:30

um

play06:31

but that's

play06:32

um

play06:36

so

play06:37

that's sort of where I'm coming from

play06:40

somebody who's you know been doing this

play06:43

uh the open source stuff and the the

play06:45

generative AI stuff

play06:46

and uh you know learned about it but

play06:51

it's all just Prelude to what's been

play06:53

happening these days and so and you know

play06:56

and it's Child's Play compared to you

play06:58

know what the algorithms uh today today

play07:00

are doing so over the last three years

play07:04

I've seen a you know an incredible leap

play07:07

in the computer's ability to understand

play07:10

you know human language and generate

play07:12

images

play07:13

and so what what are the implications of

play07:16

this

play07:18

um you know

play07:20

um you know I would say you know there

play07:22

we could have like a a private tutor for

play07:24

every child you know a personal

play07:27

assistant in everyone's pocket uh you

play07:30

know new forms of personal expression

play07:31

you know Photoshop that's easy to use

play07:33

and just does what you tell it

play07:35

um

play07:36

uh interactive fiction you know who

play07:38

knows uh these are these are just really

play07:40

like the obvious ideas some of them from

play07:43

fiction itself but uh you know like

play07:46

let's say uh George uh melies you know

play07:50

we're really just at the beginning of

play07:52

this you know we're still pointing a

play07:54

camera at a stage and we don't know

play07:56

what's really going to take off and

play07:58

where this technology uh is gonna gonna

play08:01

take us and and what it's unintended

play08:03

consequences are

play08:06

um

play08:07

so

play08:09

and these you know these these models

play08:11

these you know language models and image

play08:13

models contain uh Notions of Truth and

play08:17

Beauty but you know but who defines uh

play08:20

truth and uh who defines

play08:23

Beauty

play08:24

and you know how can you trust

play08:27

uh that the the model that you're using

play08:31

you know serves your interests versus

play08:33

say the the company that created it

play08:36

so

play08:38

um

play08:39

you know the the models that have been

play08:41

sort of making headlines are proprietary

play08:44

products uh from from secretive

play08:46

organizations and you know there's

play08:49

issues with trust and accountability

play08:52

so uh there but there's the good news is

play08:55

there's an alternative when and that's

play08:57

open source so and there's four four

play09:00

parts to that

play09:01

we need the code for doing for training

play09:05

the models to be open uh we know we need

play09:07

the code that runs the models the

play09:09

inference engine to be open

play09:11

we need the training data to be open and

play09:15

uh the model waits to be open

play09:17

and you know if we have

play09:21

those four things then you know what

play09:23

what are the implications

play09:25

you know if if you can see the data that

play09:28

goes in the models that allows

play09:30

independent you know research and you

play09:32

know third-party audits uh and it can

play09:35

really help with the trust and Alignment

play09:38

because you know you can find uh you

play09:41

know uh what's uh how you can really see

play09:45

how it works

play09:48

and the you know liberal licensing

play09:51

because you know open source makes stuff

play09:53

free to use and that's going to really

play09:55

help spread the benefits

play09:57

of of AI to to everyone

play10:00

and if the licensing is liberal

play10:03

which means like allowing commercial use

play10:06

then

play10:08

that can motivate investment I mean you

play10:10

can get like a virtual virtuous a

play10:13

virtuous uh cycle of of improvement

play10:18

um

play10:20

so

play10:21

and it's really important though that

play10:24

one of the things that open source

play10:26

enables is this forking and fine-tuning

play10:29

of of the models and so you can take a

play10:33

foundation model and add a little bit

play10:35

more data and you know change change its

play10:40

character and change its notion of Truth

play10:42

and and change it

play10:44

and that allows you know uh each person

play10:48

or for every country or every culture or

play10:53

every identity to create their own

play10:55

models so instead of having just one

play10:57

model from one company you can have each

play11:01

person or each organization or each

play11:03

group of people you know creating their

play11:06

own model and you know we're really

play11:08

lucky the beauty of these pre-trained

play11:11

models the foundation models is that

play11:13

making them is extremely expensive but

play11:16

the fine tuning is actually really easy

play11:19

and cheap you can just do that in a few

play11:21

days days

play11:23

and so you know this

play11:27

so instead of having you know one notion

play11:29

of Truth you know we can have you know

play11:31

freedom of choice and you know bias I

play11:34

think is really inherent

play11:36

in the model so we need to just be able

play11:38

to

play11:40

um take you know do the one we want and

play11:43

ultimately every person can have you

play11:46

know their their own model

play11:48

so I think that

play11:52

open source will help a lot with a bunch

play11:55

of the these sort of hard problems

play11:58

but there's really one that remains

play12:00

which is what happens if uh these

play12:05

powerful tools are used by people for

play12:07

say nefarious purposes

play12:10

um

play12:10

and you know the harm can result and

play12:13

there's really I think

play12:16

ultimately

play12:18

uh I'm optimistic about this and if you

play12:21

know if you look at the history of

play12:22

Technology

play12:24

um I think the ultimately I believe in

play12:27

in human nature and that the uh there's

play12:30

more good people than bad people and the

play12:32

benefits

play12:33

outweigh the problems and we will

play12:36

identify the problems and address them

play12:38

uh you know the best we can as as the

play12:41

cookie crumbles

play12:43

so

play12:44

and uh you know my experience is that

play12:46

you know open source is part of you know

play12:50

that solution of making these

play12:53

uh making this technology work best for

play12:55

everyone

play12:57

and so that's uh

play13:00

the end of my introduction thank you

play13:03

[Applause]

play13:08

so good morning

play13:09

I'm very happy to be here thank you for

play13:11

the invitation

play13:13

um uh copyright challenges to generative

play13:16

AI it's the focus of of my talk today

play13:20

we could have a really long conversation

play13:23

about the governance challenges

play13:24

generally the generative AI is posing

play13:28

there's a global conversation about this

play13:31

particular topic obviously

play13:34

the general public has

play13:38

embraced uh the chat jpt and many of

play13:43

these systems that you read about in the

play13:45

in the paper and a lot of technologists

play13:48

are enthusiastic about it too but policy

play13:51

makers are concerned about everything

play13:53

from uh privacy and cyber security

play13:56

defamation disinformation just to name a

play13:59

few so we're just going to talk about

play14:01

copyright today and a thing to know as

play14:04

we begin this is that some authors

play14:06

artists and programmers are very

play14:08

positive about this uh development and

play14:11

about copyright as not being an obstacle

play14:13

and others are very negative there are

play14:16

three lawsuits right now

play14:18

challenging generative AI AI on

play14:21

copyright and related grounds the United

play14:23

States copyright office is holding some

play14:26

listening sessions about stakeholders

play14:29

interests in what how the copyright

play14:32

office should be thinking about

play14:34

the main issues that it is addressing

play14:38

and it will publish a report probably

play14:40

sometime by the end of the year and the

play14:43

copyright office actually has a a

play14:45

website that's all about generative Ai

play14:48

and so lots and lots of materials there

play14:52

if you're interested in it I'm going to

play14:55

only have time to talk about three

play14:56

questions and I never have to do it

play14:58

really fast so the question about

play15:02

whether ingesting copies of copyrighted

play15:04

works as training data is that an

play15:07

infringement of copyrights in those

play15:09

Works obviously pretty much everything

play15:12

that's out there on the open web is

play15:15

actually protected by copyright law

play15:17

unless it was authored by the US

play15:18

government and that means that there's a

play15:22

lot of copyrighted stuff out there

play15:24

um even if people aren't uh exploiting

play15:27

it in the way that uh that many of the

play15:29

copyright Industries do there's the The

play15:31

Works are still protected

play15:33

um and then uh a second question uh is

play15:36

whether uh AI generated outputs are um

play15:39

infringing derivative uh works of

play15:42

ingested uh content uh two of the

play15:45

lawsuits against uh stability

play15:48

um raise this issue as well as the first

play15:50

one uh and then a third question uh has

play15:54

to do with removal or alteration of uh

play15:57

what copyright law calls copyright

play15:59

management information things about sort

play16:01

of what is what is the name of the work

play16:03

uh who's the author and on what terms is

play16:06

available and the stability cases

play16:09

raise this issue as does a class action

play16:12

lawsuit against GitHub and open AI over

play16:15

the Codex large language model and

play16:18

co-pilot the the programming assistant

play16:22

tool that Microsoft is hosting in uh in

play16:26

the cloud

play16:29

the briefest thing I can say about

play16:31

copyright is that um it it attaches

play16:35

automatically by operation of law from

play16:38

the first fixation of a work in a

play16:40

tangible form and the author is the one

play16:44

who gets those rights and these are the

play16:48

five major

play16:50

exclusive rights that copyright grants

play16:52

to them the right to last for

play16:54

practically forever and copyright only

play16:57

protects the original expression of an

play17:00

author not the ideas not the facts not

play17:02

the methods there are lots and lots of

play17:04

unprotected stuff in copyrighted works

play17:08

and copyrights exclusive rights are

play17:10

limited by fair use and various other

play17:13

doctrines and there's a special

play17:15

copyright like law that makes

play17:18

intentional removal or alteration of

play17:21

copyright management information

play17:24

um illegal if you know that it will

play17:26

facilitate uh copyright infringement uh

play17:30

Fair uses are non-infringing and

play17:35

um so it's a defense to a charge of

play17:38

infringement and fair use is usually the

play17:41

defense that's raised when we're talking

play17:42

about the ingesting of uh copyright

play17:45

Works uh in uh in as training data and

play17:50

there are various factors I'm not going

play17:52

to be able to go into this right now but

play17:53

I'm certainly happy to talk at the Break

play17:55

um about these things so is ingesting

play17:59

um works as training data copyright

play18:01

infringement or not there are at least

play18:04

several cases and I'm listing two of

play18:06

them here uh where the court basically

play18:09

threw out copyright claims on fair use

play18:13

grounds so that field put a bunch of his

play18:17

work up on on the internet uh Google of

play18:20

course crawls the web and copied it to

play18:23

index the contents and to let people

play18:26

find Fields work and Fields that Ah

play18:30

that's a copyright infringement and the

play18:32

court said no it's fair use because

play18:34

Google is using the works not to exploit

play18:38

the content but to just let people know

play18:40

that the works uh exist and are out

play18:43

there and in the author's Guild versus

play18:46

Google case and Appellate Court rule the

play18:49

Google's digitization of tens of

play18:52

millions of in copyright books from

play18:53

research Library collections was fair

play18:56

use when it was done for the purpose of

play18:58

indexing uh the contents and serving up

play19:03

Snippets in response to search queries

play19:04

so stability is going to be relying on

play19:09

these in similar cases to support its

play19:11

fair use defenses

play19:13

but there are a lot of people out there

play19:16

who really don't like the ingestion

play19:19

without permission and they can't really

play19:23

easily opt out they weren't paid for uh

play19:26

the value of their contributions uh and

play19:30

some part of the concern is that the

play19:33

images that are being generated through

play19:34

stable diffusion and others of these

play19:37

images

play19:38

generators is that they will compete in

play19:42

the market with the images that artists

play19:46

actually are doing and it's in some

play19:48

sense you're competing against yourself

play19:50

in some sense at least that's the way

play19:52

some of the artists think about it and

play19:54

this cartoon actually is kind of

play19:56

illustrates how some people think about

play19:58

generative AI

play20:00

um

play20:02

now they're a countervailing

play20:04

considerations

play20:05

generally speaking the people who

play20:07

develop these large language models are

play20:09

not really interested in the original

play20:12

expression in a work they're really

play20:14

interested in essentially understanding

play20:17

the facts and they kind of think of uh

play20:20

of of of documents and the like as raw

play20:23

material for computational uses that are

play20:26

actually not exploiting the expression

play20:29

and so that's sort of a way of kind of

play20:34

looking at it and also generative AI

play20:37

enables a lot of creative reuses of

play20:39

things and so what copyright law is

play20:42

supposed to be doing is promoting the

play20:45

progress of science or knowledge and

play20:47

culture and so the people who favor

play20:52

um

play20:54

generative AI are thinking that this is

play20:57

actually a good thing that that

play21:00

generative AI advances the purposes of

play21:03

copyright uh outputs as a derivative

play21:06

Works

play21:07

um so

play21:10

authors have exclusive rights to prepare

play21:13

derivative works but the course have

play21:16

basically said something can't infringe

play21:18

the derivative work right if you haven't

play21:20

just extracted some expression from an

play21:24

existing work and then put it into a

play21:26

second work so it's not enough just to

play21:29

be based upon an existing work you

play21:32

actually have to have

play21:35

essentially taken expression and reused

play21:39

it on the class action lawsuit against

play21:41

stability and claims that every output

play21:44

of stable diffusion is an infringing

play21:46

derivative work and I think that's

play21:48

actually hard to say under existing

play21:52

precedence the that that's a sound

play21:54

result but in general

play21:57

the text and images that are generated

play22:00

in response to user props are not going

play22:02

to be substantially similar in their

play22:05

expression to the works that were in the

play22:07

training data and if that's true through

play22:10

then that's unlikely to infringe the

play22:12

derivative work right

play22:14

and

play22:16

you know there are some examples that

play22:19

people have shown where you can

play22:21

essentially like Mickey Mouse and

play22:24

Superman you do a prompt with them

play22:26

you're going to get an image that's

play22:28

going to look like Mickey Mouse or

play22:30

Superman

play22:32

of the cases against stability

play22:35

um Getty is says Hey stability you

play22:38

ingested 12 million photographs and

play22:41

captions from our online database that's

play22:44

actually copyright infringement also

play22:48

you know there are some images they

play22:51

claim that are infringing derivative

play22:53

infringing of the derivative work right

play22:55

and also there's a copyright management

play22:57

information claim in this particular

play22:59

case and here's an example I don't know

play23:02

whether you can see it but uh

play23:05

this is getting images and this is the

play23:08

mangled uh stuff uh and so that's a

play23:11

basis for claiming uh that there is uh

play23:14

that that's uh that uh stable diffusion

play23:18

uh is at least violating the copyright

play23:21

management information uh law uh and so

play23:25

with that uh you know these lawsuits are

play23:27

in very early stages

play23:30

um it's going to be many years actually

play23:33

probably until we know what the results

play23:35

are from the courts but it's actually

play23:38

good to remember this isn't the first

play23:40

disruptive technology that's ever been

play23:42

uh out there um so I like to remember

play23:46

that piano rolls were a disruptive

play23:48

technology of the day because the

play23:51

composers basically said hey that's my

play23:54

music but it didn't look like a

play23:57

copyright infringement to the courts and

play23:59

so it went up to the Supreme Court said

play24:01

no it's not a copy because it's part of

play24:03

a machine so copyright law has had to

play24:06

evolve to kind of recognize that that

play24:09

you know technology has to be responded

play24:13

to and sometimes you amend the law in

play24:15

order to protect copyright owners and

play24:18

sometimes you know you just let the

play24:20

technology happen so you know the

play24:22

recording industry hated

play24:24

um MP3 players but guess what

play24:27

um they're like they're legal so anyway

play24:29

I'll just leave that with the uh for a

play24:33

discussion and thanks very much for your

play24:35

time

play24:42

so I think we have microphones for

play24:44

questions so if you want to raise your

play24:45

hand I'll start things off thank you

play24:47

both for for those really stimulating

play24:50

comments

play24:51

um you you referred to a bunch of

play24:53

lawsuits uh that are ongoing and I and I

play24:56

was wondering is that how this is going

play24:58

to play out is it going to be the courts

play25:00

that figure this out versus kind of a

play25:02

prospective either regulatory approach

play25:04

or self-governance by the companies that

play25:08

are involved in I might be very

play25:09

interested in both of your thoughts

play25:10

about those alternatives to just having

play25:13

the courts make these decisions and

play25:15

using those precedents well the

play25:17

copyright office as I said has a clear

play25:19

intent to put a marker down in terms of

play25:23

the sort of the major questions that I

play25:25

was just raising and so they'll have to

play25:28

they'll have a say about do they have a

play25:30

good history of moving uh no

play25:34

um but

play25:37

um you know they have motivation they

play25:40

can't do anything by themselves right

play25:42

they can make recommendations to

play25:44

Congress to pass legislation

play25:48

and we all know how functional Congress

play25:52

is right now so um you know part of what

play25:55

happens is that when you have a

play25:57

dysfunctional

play25:58

Congress and you know the you know the

play26:01

Biden Administration whatever they might

play26:03

think about this

play26:05

um they can't do anything either okay so

play26:08

lots of things end up in court because

play26:13

there was a big dispute and there's no

play26:15

other entity that really can deal with

play26:17

this I should actually mention that um

play26:20

one of the things that is important is

play26:23

that fair use is a limitation on

play26:26

copyright in the United States there are

play26:28

a number of other countries that

play26:30

actually have fair use defenses in their

play26:33

copyright laws too but most of the other

play26:36

developed countries have special

play26:38

exceptions for text and data mining and

play26:42

so at least for the purpose of ingestion

play26:45

of copyright works for generative AI the

play26:50

text and data mining exceptions would

play26:52

actually kick in and at least for

play26:55

scientific and non-profit uses

play26:58

that cannot be overridden by contract

play27:02

and right now there's a question about

play27:05

whether people can opt out I remember

play27:07

the discussion on the first session

play27:10

about opting out and it's not clear you

play27:13

can opt out

play27:14

so Scott on the same issue you had the

play27:17

four key tenets of Open Source Code open

play27:20

inference code open data open and I

play27:22

wonder if you can comment on the data

play27:24

open I just wanted to hear a little bit

play27:26

more about how open and how this

play27:28

interacts with your thoughts about

play27:29

copyright and

play27:30

sure the when I say the data open I mean

play27:33

that the you know the all the the list

play27:36

of images or texts all the all the

play27:39

documents that are fed into the training

play27:40

algorithm is a published list that

play27:43

anyone can can inspect it's and there's

play27:46

actually a proposal in the European

play27:48

Union that um as part of the AI act that

play27:52

they're considering uh is to have a

play27:54

requirement that generative AI systems

play27:57

have to disclose uh what copyright Works

play28:00

um they ingested so that's kind of

play28:03

consistent with your kind of preference

play28:06

for open data in that way yeah so and

play28:09

then once if if you have a list you can

play28:12

look and see are are you on it um and

play28:15

you know copyright lies certainly too

play28:17

complicated for me to understand I have

play28:19

no idea uh you know I what what's going

play28:22

to happen and um uh but I I do believe

play28:26

that there is you know some some list of

play28:29

uh you know there's tons of uh public

play28:32

domain imagery out there there's tons of

play28:34

information out there there there's all

play28:37

kinds of Licensing available and I

play28:39

believe there is a list of data there is

play28:42

a data set uh that would be uh sort of a

play28:46

useful and and legal uh and you know we

play28:51

should we should pursue it you know

play28:53

together so like let's let's look at the

play28:55

list and uh

play28:57

come up with a consensus it's actually

play28:59

also important to know especially for

play29:01

any of you out there who are trying to

play29:03

develop some of these models the

play29:05

Creative Commons official position

play29:07

is that using

play29:10

a Creative Commons works as training

play29:14

data is actually not a violation of the

play29:17

Creative Commons licenses now this is in

play29:19

contrast to the the GitHub lawsuit in

play29:25

which the four programmers have sued

play29:29

GitHub open Ai and Microsoft

play29:33

for breaching open source licenses

play29:36

because the generative

play29:38

system that was built on top of a model

play29:42

that was ingested the open source code

play29:45

is actually there's claiming it's a

play29:48

breach of license because there's no

play29:50

attribution right and so that's actually

play29:54

an issue that where they at least taking

play29:58

that position now again

play30:01

who knows what the courts will do with

play30:03

that I did want to take some questions

play30:04

from the audience do we have

play30:08

I saw this hand first

play30:12

thank you very much for your

play30:14

thought-provoking comments uh one of the

play30:17

points you made were about uh truth and

play30:21

Beauty right

play30:23

crude

play30:24

coming from a scientific background you

play30:26

kind of think about it as objective

play30:28

evidence something which is backed by

play30:30

objective evidence but beauty is more

play30:32

subjective

play30:33

it can be interpreted by the Iowa to

play30:36

Beholder right so can you elaborate on

play30:39

that in in terms of how can we make sure

play30:43

that

play30:44

we don't get into the world of

play30:47

alternative facts

play30:48

but maintain one notion of truth if it

play30:53

makes sense

play30:56

uh you know there's certainly scientific

play30:59

uh truth is maybe uh the the easier part

play31:04

like at least we have like you know the

play31:07

scientific method that we have a way

play31:08

that we can agree a process to figure

play31:11

out what that is obviously it's executed

play31:13

by people and imperfect

play31:15

um but there's you know there's there's

play31:18

all kinds of truths uh you know there's

play31:20

historical truths uh and there's you

play31:24

know personal truths and I just really

play31:27

don't see us uh coming up with a single

play31:31

definition for that and so that's why we

play31:36

need you know the ability for people to

play31:39

fine-tune their own versions of the

play31:41

models and

play31:44

ultimately the you know

play31:47

Define their Define their own truths and

play31:51

yeah those could be if you disagree with

play31:53

them you you might uh call it

play31:56

misinformation or alternative facts I

play31:58

think uh

play32:00

you know uh but I you know I I just

play32:04

don't see uh any way of uh you know uh

play32:08

uh regulating a single answer there and

play32:12

I think it's better to uh you know allow

play32:16

um you know diversity of those terms A

play32:21

diversity of Truth and Beauty

play32:23

um it's so that's that's uh that's

play32:26

that's uh how I that's my position

play32:30

we have one two and then if we're really

play32:32

lucky three

play32:36

um thank you so much for uh your

play32:38

comments I have a question that sort of

play32:41

um touches on both aspects of uh

play32:44

generative AI from the artist's

play32:46

perspective as well as when you look at

play32:48

copyright infringement and so I don't

play32:50

know if everyone is familiar with uh the

play32:53

new trend of Music covers by generated

play32:57

AI machines

play33:00

um but more so coming from the consumer

play33:03

um listening to a piece of music and

play33:05

wondering what it would sound like if a

play33:07

different artist sang it

play33:09

um and I know Universal Music put out a

play33:12

statement saying they would pursue legal

play33:14

action against those types of music

play33:17

covers

play33:18

um but when another human makes a cover

play33:20

of a musical piece it's not illegal

play33:22

right there they have the right to

play33:25

creatively Express themselves

play33:27

um on a piece of music and so I'm I'm

play33:30

wondering from the artist's perspective

play33:32

as well as from a legal perspective how

play33:35

can we always even possible to regulate

play33:37

this intersection of generative AI meets

play33:42

consumers that are paying to have access

play33:43

to certain types of music as well as

play33:48

copyright and intellectual property from

play33:51

the artist's perspective

play33:55

degree because we do have a panel in the

play33:57

afternoon that addresses that too so

play33:58

thank you so much for the great question

play34:01

well I think that to the extent that

play34:04

some software tool

play34:06

enables people to you know do their

play34:11

karaoke as a as a kind of

play34:15

generate the music that you want to hear

play34:17

that's going to be something that

play34:19

Universal probably won't know about and

play34:21

won't care about

play34:22

it's kind of like fan fiction kind of

play34:25

stuff which mostly has been either

play34:28

considered fair use or been tolerated by

play34:32

people but you try to go you try to go

play34:36

public with that you try to

play34:37

commercialize it and Universal I

play34:40

guarantee Universal will be there and

play34:43

one thing that actually is going to be I

play34:45

think more significant

play34:48

in the generative AI space then it has

play34:52

been in other contexts and that is right

play34:55

of publicity so there is this law it's

play34:59

really state law which basically says

play35:02

that people have a right to control the

play35:07

use of their names and likenesses and

play35:10

some for kind of commercial purposes so

play35:13

if somebody takes a picture of you and

play35:15

then uses that as an advertisement for

play35:19

some other product that's actually a

play35:21

violation of right of publicity and that

play35:25

the the Drake song that went viral is

play35:30

something where I think there was a

play35:31

right of publicity issue in that case I

play35:34

mean I didn't turn into litigation

play35:36

because it just got taken down really

play35:38

fast because I can tell you that

play35:41

Universal makes its voice heard really

play35:45

loud with anybody that's hosting stuff

play35:48

like that so it got taken down pretty

play35:50

fast but it went viral really really

play35:52

fast too and it's certainly not going to

play35:55

be the only time when a sample of

play35:59

someone's voice is used to

play36:01

then make a music that sounds like the

play36:05

that's them when it's not so they're

play36:07

also laws against false representation

play36:10

and so there are kind of some existing

play36:13

laws other than copyright that um that

play36:16

may come may have some bearing on that

play36:19

we are now entering the lightning round

play36:21

all right really quickly I I really

play36:24

appreciate um your your response got to

play36:26

the question about scientific truth and

play36:28

I think it's really important to push

play36:29

back on this privileging of scientific

play36:31

truth because it's you know I think from

play36:33

this audience it is just as ambiguous as

play36:36

the definitions of beauty I think I

play36:39

wanted to ask you quickly if you have

play36:41

you've been really unique in giving away

play36:43

a huge amount of your your work and even

play36:46

without asking for credit or attribution

play36:48

so I'm curious if you have any regrets

play36:51

or advice on that and uh and really

play36:54

quick to Pam if we have time but uh is

play36:56

it has this recent Warhol versus

play36:59

Goldsmith's decision changed anything in

play37:02

in regard to AI

play37:05

um I I have uh I don't have any regrets

play37:08

I mean I've lived a Charmed existence

play37:10

I've been incredibly lucky

play37:12

um and it's it's hard to untangle you

play37:14

know like what what would have happened

play37:17

um but so I think it's it's it's I think

play37:20

it's working great I would recommend it

play37:22

and um

play37:25

yeah I I don't know uh

play37:28

let's go let's go to the next one thank

play37:30

you just

play37:32

I'm sorry

play37:34

yeah go so short version

play37:38

um the United States Supreme Court just

play37:40

decided that that a commercial license

play37:44

of an image that

play37:47

Andy Warhol created in 1984

play37:51

so the commercial license in 2016

play37:54

was

play37:56

not uh a fair use of that image they

play38:03

didn't reach the question about sort of

play38:06

how far and how whether the works when

play38:10

they were created were infringing or not

play38:14

they just basically said that's the that

play38:16

Goldsmith had abandoned those claims and

play38:18

so really really narrow issue as a

play38:22

commercial license

play38:23

of that particular image Court didn't

play38:27

say it was uh it was necessarily

play38:29

infringing but said that it wasn't the

play38:31

transformative enough use of the image

play38:34

so you know there's some dicta in that

play38:36

opinion that a lot of people who hate

play38:38

fair use are going to just love but for

play38:41

them for the most part really narrow

play38:43

narrow ruling

play38:48

we do have some questions from the from

play38:50

the audience remotely and but many of

play38:52

the issues have at least been touched

play38:54

remote uh and glancing blows

play38:57

um I wanted to just ask you Pam just uh

play38:59

to speculate or sort of what your

play39:00

assessment is of the um viability of

play39:03

fair use claim given that the fourth

play39:05

factor is uh maybe increasingly like uh

play39:10

in question uh

play39:12

since the AI companies now have created

play39:14

this booming market for data and the

play39:18

value of that data now is sort of widely

play39:20

recognized and a lot of companies who

play39:22

have access to a lot of data are

play39:24

creating licensing agreements for the

play39:28

use of that data in training AI does

play39:31

that change your you know guess I guess

play39:35

about how courts will fall on the fair

play39:37

use question well I think

play39:39

I think that uh Getty Images lawsuit

play39:43

against stability

play39:46

um that's a that's a that's a more kind

play39:49

of

play39:50

um important consideration because

play39:54

um getting images says I don't I I'm

play39:57

completely happy for you to use uh my

play40:01

images and the captions uh as uh

play40:04

training data but I have a licensing

play40:06

program for that and so it's not fair

play40:08

use the class action lawsuit I think is

play40:11

quite a different thing because the the

play40:15

three artists that claim to be the class

play40:19

plaintiffs

play40:20

are trying to represent all of the

play40:23

artists of all the visual images in the

play40:25

world that

play40:28

um that were ingested you can't get a

play40:30

license there so it seems to me that

play40:33

that the fair use claim is going to be

play40:36

stronger where it's infeasible to

play40:40

um to do it I don't know that the

play40:42

existence of look every time somebody

play40:44

brings a lawsuit they're basically

play40:46

saying I

play40:48

lice I have licenses okay or I'm willing

play40:52

to license this particular thing so the

play40:54

fact that you want to license the fact

play40:56

that you have a licensing program

play40:57

doesn't actually necessarily mean that

play40:59

it's not fair use I am not going to

play41:01

predict here but certainly the Getty

play41:05

Images case has a a stronger response to

play41:09

the fair use claim than the the Anderson

play41:13

class action case

play41:15

well I want to thank the panelists and

play41:17

thank the uh for the questions

play41:18

[Applause]

Rate This

5.0 / 5 (0 votes)

Related Tags
Generative AIOpen SourceCopyright LawArtificial IntelligenceCreative ExpressionLegal ChallengesInnovation ImpactTech EthicsIntellectual PropertyFair Use