GenAI Exceptionalism: A Legal Perspective

David Atkinson

​

​

Law, Ethics, and AI (LEAI)

• To learn more about the law and ethics of AI
• Free!
• Modifiable!
• Regularly updated!
• The most comprehensive document on the subject!

​

​

• For a more refined version, follow this free Substack
• Substack name: Intersecting AI
• With technical review by AI researcher Jacob Morrison

2

The Journey

• A taxonomy of GenAI legal risks
• What are the claims?

3

The Journey

• A taxonomy of GenAI legal risks
• What are the claims?
• The Gaps: Unsettled Law
• What are the open questions?

4

The Journey

• A taxonomy of GenAI legal risks
• What are the claims?
• The Gaps: Unsettled Law
• What are the open questions?
• The Unsettled Law Doctrine
• What is the role of GenAI in-house counsel?

5

The Journey

• A taxonomy of GenAI legal risks
• What are the claims?
• The Gaps: Unsettled Law
• What are the open questions?
• The Unsettled Law Doctrine
• What is the role of GenAI in-house counsel?
• GenAI Exceptionalism and the First Amendment
• GenAI outputs are not speech

6

The Journey

• A taxonomy of GenAI legal risks
• What are the claims
• The Gaps: Unsettled Law
• What are the open questions
• The Unsettled Law Doctrine
• What is the role of GenAI in-house counsel
• GenAI Exceptionalism and the First Amendment
• GenAI outputs are not speech
• GenAI Exceptionalism and Contract Law
• Bots can make terms of service binding

​

7

The Journey

• A taxonomy of GenAI legal risks
• What are the claims
• The Gaps: Unsettled Law
• What are the open questions
• The Unsettled Law Doctrine
• What is the role of GenAI in-house counsel
• GenAI Exceptionalism and the First Amendment
• GenAI outputs are not speech
• GenAI Exceptionalism and Contract Law
• Bots can make ToS binding
• GenAI Exceptionalism and Copyright Law
• Why nobody should have to pay for any content ever again

8

PART 1: A TAXONOMY OF LEGAL RISKS

9

A Taxonomy of Legal Risks

• Over 35 lawsuits

​

​

10

A Taxonomy of Legal Risks

• Over 35 lawsuits

​

​

11

Analysis of Cases

• Question:
• What is the greatest litigation risk for GenAI companies?

12

Analysis of Cases

• Question:
• What is the greatest litigation risk for GenAI companies?
• Methodology
• Gathered all complaints and tallied the claims made

13

The Claims

1. Direct copyright infringement
1. Reproducing, distributing, publicly displaying, etc.

14

The Claims

1. Direct copyright infringement
2. Vicarious copyright infringement
1. It allows for Party A to be found liable for the infringing acts of Party B if (i) Party A had the right and ability to control the infringing activity and (ii) Party A had a direct financial interest in the infringement.

15

The Claims

1. Direct copyright infringement
2. Vicarious copyright infringement
3. Contributory copyright infringement
1. A successful claim requires that (i) Party A makes a material contribution to the infringing activity, while (ii) having knowledge or a reason to know of the direct infringement by Party B.

16

17

The Claims

1. Direct copyright infringement
2. Vicarious copyright infringement
3. Contributory copyright infringement
4. Digital Millennium Copyright Act 1202(b)
1. Prohibits people from removing or altering copyright management information (sometimes called CMI, which includes information like the creator’s name and title of the work).
2. Requires the intent to remove the CMI, distributing CMI knowing that the CMI has been removed or altered without authorization, or distributing copies of works knowing the CMI has been removed or altered, provided that the person knew or had reason to know, that such removal or alteration of CMI would “induce, enable, facilitate, or conceal an infringement...”
3. Might also require exact copies of the works (Doe 1 v Github)

18

The Claims

1. Direct copyright infringement
2. Vicarious copyright infringement
3. Contributory copyright infringement
4. Digital Millennium Copyright Act 1202(b)
5. Unjust Enrichment
1. Occurs when Party A is in some agreement with Party B. Party A confers a benefit on Party B, but Party B does not comply with the terms of the agreement yet still retains the benefit conferred by Party A.
2. Some plaintiffs try to broaden the scope of this claim, though, to situations where there is no agreement, such that Party B receives a benefit from Party A without Party A’s assent or consent.

19

The Claims

1. Direct copyright infringement
2. Vicarious copyright infringement
3. Contributory copyright infringement
4. Digital Millennium Copyright Act 1202(b)
5. Unjust Enrichment
6. Negligence
1. Someone is harmed and someone else must be liable even if the action that caused the harm wasn’t in breach of a contract or in violation of a criminal law.
2. Violator must have acted unreasonably and been the proximate cause of the injury (but-for and foreseeable)

20

The Claims

1. Direct copyright infringement
2. Vicarious copyright infringement
3. Contributory copyright infringement
4. Digital Millennium Copyright Act 1202(b)
5. Unjust enrichment
6. Negligence
7. Unfair competition
1. “any unlawful, unfair or fraudulent business act or practice and unfair, deceptive, untrue or misleading advertising...”

21

The Claims

1. Direct copyright infringement
2. Vicarious copyright infringement
3. Contributory copyright infringement
4. Digital Millennium Copyright Act 1202(b)
5. Unjust enrichment
6. Negligence
7. Unfair competition
8. And many more

22

23

Direct Copyright Infringement Claims

Vicarious Copyright Infringement Claims

Contributory Copyright Infringement Claims

Do Mitigations Matter?

• Some definitely do
• Using public domain; licensing material

24

Do Mitigations Matter?

• Some definitely do
• Filtering inputs and outputs may help
• Doesn’t work for open source (can’t control filtering other parties use)
• Judge in NYT v. OpenAI hints that frequency of infringing outputs may not matter

25

Do Mitigations Matter?

• Some definitely do
• Filtering inputs and outputs may help
• Deduplicating may help
• Reduces memorization; less memorization = less regurgitation
• If outputs don’t suggest copyrighted stuff was used, you’re less likely to be sued

26

Do Mitigations Matter?

• Some definitely do
• Filtering inputs and outputs may help
• Deduplicating may help
• De-linking images from creator and copyrighted names
• Don’t include artist names (e.g., “Andy Warhol”) or copyrighted names (e.g., “Mickey Mouse”) along with the image
• Makes it harder for model to do “in the style of” outputs

27

Do Mitigations Matter?

• Some definitely do
• Filtering inputs and outputs may help
• Deduplicating may help
• De-linking images from creator and copyrighted names
• Takeaway?
• Mitigations help! But they may not be a panacea even if they work flawlessly (except maybe the first bullet above)
• Mitigating injuries to plaintiffs can lower damages GenAI companies might have to pay

28

PART 2: UNSETTLED LAW

29

What are the Open Questions for GenAI

• Trying to identify dozens of unsettled and often overlooked legal questions
• I’ve underlined questions I find particularly interesting

30

Current Litigation Just Scratches the Surface

1. Copyright Law
1. Copyright’s Purpose
1. Whether Article I Section 8 of the Constitution is furthered by extending current copyright law to GenAI.
2. Fair Use Factors
• Whether the fair use factors should be revised and clarified.
3. Whether GenAI should receive special fair use privileges that don’t apply to humans.
4. Data Provenance
• Whether the provenance of data matters.
• Whether it matters if developers know or should know that their training dataset includes unauthorized data, such as books, that were taken without the copyright owner’s authorization. 
• Whether fair use absolves every controversial source of data. 
• Whether strict liability should only apply when a developer knows or should know the dataset includes data collected without the copyright owner’s authorization.
• Whether data intended to be publicly available should be treated the same as data that is publicly available but wasn’t intended to be by the data creator. 
5. A New Secondary Liability?
• Whether a new secondary liability analysis is necessary and whether it should include a negligence component.
6. How Strict
• Whether the frequency of infringing outputs matter, or if a single output is sufficient for infringement.
• Whether how many times a model must be prompted before it makes an infringing output matters.
• Whether the technique required to elicit an infringing output matters.

​

31

Current Litigation Just Scratches the Surface

1. Privacy
1. Patchwork of Laws
1. Whether a federal privacy law would be helpful or is necessary.
2. Linking Data
• Whether datasets may allow a form of search or intrusion upon seclusion that would otherwise be illegal.
3. Unknown Content
• Whether datasets should be subject to scrutiny by regulators or auditors to protect privacy.
4. Inability to Comply with GDPR
• Whether an AI entity provides a remedy that does not satisfy the letter of the law but mostly meets the spirit of the law is sufficient.
• Should GenAI developers be allowed to choose which privacy rights they do and don’t have to adhere to?
• Whether regulators should prohibit the deployment of technology that cannot comply with laws as written (taking a firm stance on what is a “right,” such as a right to privacy, versus what is a “privilege”).
1. Or, whether the laws should adapt to technical realities.

32

Current Litigation Just Scratches the Surface

1. Torts
1. Backward-Facing Laws
1. Whether strict liability should apply to GenAI to account for the greater potential for unknowable risk by the user. 
2. Whether allowing “in the style of” prompts or allowing AI replicas of real people without consent should be legal.
2. Liability Formula
• Whether the standard should shift from “foreseeable” to “known or knowable,” setting a higher bar for liability for GenAI developers who can’t possibly foresee all potential harms? 
• Whether companies should have to follow the precautionary principle.
• Whether a set of minimum standards should be set for what’s reasonable and failing to satisfy that criteria shifts the standard from negligence to strict liability.
3. Product Design Liability
• Whether developers should be allowed to deploy models even if they are unable to sufficiently mitigate or prevent foreseeable harms.
4. Counterfeit People
• Whether AI should be required to reveal it is AI and not a human when interacting with a human.
5. Safe Harbor
• Whether there should be a “reasonable effort” threshold/safe harbor for entities making good faith efforts to prevent harmful outputs.
6. Data Laundering
• Whether it matters if a nonprofit gets an exemption (either a per se exemption under a law or a fair use exemption) to do something a for-profit can’t, but then they license it to the for-profit, so the end result is the same.

33

Current Litigation Just Scratches the Surface

1. Contract Law
1. Terms of Service
1. Under what circumstances should a bot scraping content from virtually every public website be contractually bound by a site’s terms of service?
2. Whether terms of service apply the same to bots as it does to humans visiting a site, regardless of how many times the bots visit (and therefore are likely “aware” of any terms. 
3. Whether there should be minimum standards to make browsewrap applicable so it doesn’t have to be litigated on a case-by-case basis. 
4. Whether a site can claim any rights just by putting them in a terms of service. For example, can it create property rights just by saying it’s so?
2. Dataset Licensing
• Whether it is legal to unilaterally change the license terms of a dataset. 
• Whether any illegality in changing license terms is truly enforceable.
• Whether intent should play a role in determining liability for changing license terms. 
• Whether a party should be liable for not confirming the proper license is connected with the appropriate dataset, and how any license could reasonably be enforceable if the data can be modified and recombined with other data or into other formats.

34

Current Litigation Just Scratches the Surface

1. Criminal Law
1. Intent
1. Whether GenAI developers should be effectively immune from criminal charges for virtually all actions of the GenAI models.
2. Whether actions by GenAI agents that closely (or maybe even identically) resemble intent in traditional criminal cases should make GenAI developers criminally liable.
2. Criminal Negligence
• Whether a criminal negligence standard would make sense for GenAI.

35

Current Litigation Just Scratches the Surface

1. Property Law
1. Personal Property and Digital Artifacts
1. Whether some or all property rights that attach to physical personal property should also apply to digital property.
2. Robots.txt
• Whether and to what extent following robots.txt is important when considering legal liabilities.
• Whether for-profit entities should be responsible for ensuring they don’t acquire data through any means when robots.txt tells them to not scrape a page, or only if the entity is barred from scraping the original site itself.
3. Publicly Available
• Whether it matters if the information that is publicly available was placed in the public space without authorization.
4. Opt-out
• Whether and to what extent, if at all, is allowing people to opt-out of datasets important. Should this be balanced against the cost or technical burden of complying with an opt-out request?
• Whether opt-out requests must always be respected. If so, is it retroactive, or only for current datasets and future models?
• Whether GenAI entities should only be able to use opt-in data.

36

Current Litigation Just Scratches the Surface

1. First Amendment
1. Regulation and Scrutiny
1. Whether GenAI outputs are constitutionally protected speech. 
2. CDA 230
• Whether a company’s GenAI outputs make it the speaker or publisher, and therefore make the company potentially liable for the outputs.

37

PART 3: THE UNSETTLED LAW DOCTRINE

38

Law, not Ethics

• Professional responsibility = due diligence, commingling of funds

​

​

39

Lawyers Are Non-Ethical

• NOT unethical. Well, not usually.
• Ethics agnostic:
• Child molesters, election deniers, climate change deniers, vaping companies who want to market to children
• Why??

40

Taken to Extremes

• It’s the only alternative, or
• It was an accident, or
• The chemicals aren’t that bad, or, maybe,
• The chemicals are actually good for the environment

41

It’s still probably the best approach, imo

AI and Law

• Copyright Overview
• Creators automatically get exclusive rights:
• Reproduce, distribute, perform in public, etc.
• OpenAI and Meta…pirates?
• Books3
• Argument?
• No substantial similarity
• Backup: fair use affirmative defense
• Copies? What copies?

​

• THIS IS UNSETTLED LAW TERRITORY

​

​

42

Unsettled Law and Ethical Deciders by Default

• Whether non-profits and/or research organizations using legal exemptions from laws for copyright or privacy are committing data laundering by then licensing that data to for-profit entities.
• Whether gathering data from across the web (web scraping) to train for-profit AI models is ethical.
• Whether it is important to follow robots.txt.
• Whether generating profits from LLMs trained on copyrighted materials or private data without compensating the source is ethical and is not a form of unjust enrichment.
• Whether the provenance of the training material matters.
• Whether allowing “in the style of” prompts or allowing AI replicas of real people (again, without consent) is ethical in addition to not being illegal.
• Whether people should have the right to opt out of datasets; or, should only opt-in be allowed?

​

43

The Unsettled Law Doctrine

1. When the legal issue is unsettled and there are massive ethical questions, such as whether using pirated material to train AI is permissible, the discussion should be driven by ethical arguments, not legal arguments.
2. If a company can’t make a strong ethical argument to take a controversial action in an area of unsettled law, then their counsel should not make a legal argument in support of the controversial action.

44

When Lawyers Take the Lead

• When a legal team settles on a response to an issue, they are not only saying, “We believe we have a strong legal position,” but are often also, by default, implying, “and we believe this position is the one we should take.”

​

• Example:

45

Doesn’t Legal = Ethical?

• The outcome of many judicial decisions is less a factor of the strength of the arguments and more about the access to resources.

​

• Example:
• Any of the MAMAA* companies can easily throw more money into building a legal position (via research, lawyers, investigators, forensic analysts, etc.) than virtually any individual plaintiff or even the entire Federal Trade Commission (FTC) because each member of MAMAA generates more revenue in a few days than the entire annual budget of the FTC.
• This also means MAMAA companies can hire more experienced lawyers who can use more advanced tools to seek out every minute legal advantage.

46

*MAMAA = Microsoft, Alphabet, Meta, Amazon, Apple

Implication

1. By the time the issues are adjudicated, the AI will be deeply ensconced in society, difficult or impossible to uproot.
2. The tech companies will determine what is ethical and acceptable, not the society, or administrative agencies, or the judicial system.

47

Lawyers, Generally, Aren’t Ethicists

• For in-house lawyers, their loyalty generally lies with the provider of their paychecks, not with what’s best for society.

​

48

• Importantly, it’s still not clear they should. Perhaps it would be unethical to not train on the data.
• The data makes the LLMs more useful, more well-rounded, and, hallucinations aside, more accurate than the LLMs would be without the data.
• LLMs have proven helpful at a myriad of tasks and their potential upside seems limitless. Would it be unethical to deprive society of such advances for the sake of copyright law?

• Well…perhaps!

​

What’s a Lawyer to Do?

• Again, this is not a place for lawyers to take the wheel.
• Ethics committees: impartial third-party experts with diverse backgrounds; transparent
• Defensible positions
• A small number of people in a single location should not determine what’s best for everyone everywhere.

49

What People Think

50

Ethical

Unethical

Legal

Illegal

Simple: Just do what falls into this area

Simple: Don’t do what falls into this area

Reality

51

Illegal

Not Illegal

Ethical

Unethical

Easy calls

No such thing as “legal” just “not illegal”

​

Always fact-dependent

Easy calls

Defacing site that facilitates pedophilia

Many of the restricted uses in ChatGPT terms

Unsettled law

Ignoring robots.txt (?)

Training on SciHub corpus (?)

GenAI Exceptionalism

GenAI Exceptionalism Defined

• The belief held by GenAI companies that society should grant GenAI special exemptions that society either
• (a) doesn't allow for humans (see: Unfair Learning), or
• (b) only allows for humans (see: Intentionally Unintentional)

53

PART 4: GENAI EXCEPTIONALISM AND THE FIRST AMENDMENT

54

The First Amendment

• Protects lots of stuff, including freedom of speech:
• “Congress shall make no law respecting an establishment of religion, or prohibiting the free exercise thereof; or abridging the freedom of speech, or of the press; or the right of the people peaceably to assemble, and to petition the Government for a redress of grievances.”

55

Examples of Speech

• Code
• Music
• Arm bands
• Writings
• Speeches
• Paintings
• Cakes, sculptures, etc.

56

Examples of Speech

• Code
• Music
• Arm bands
• Writings
• Speeches
• Paintings

57

Q: WHAT DO THEY ALL HAVE IN COMMON?

Examples of Speech

• Code
• Music
• Arm bands
• Writings
• Speeches
• Paintings

58

Q: WHAT DO THEY ALL HAVE IN COMMON?

​

A: Intentionality by a human

Importance

• If GenAI output is speech, it would likely prohibit treating AI as a product or attaching liability to harmful outputs.
• It could also give AI companies free rein to collect and use personal information as data inputs for their algorithmic (constitutionally protected) outputs.
• It could also limit many forms of consumer protection and other regulatory objectives.

59

Are GenAI Outputs Speech?

60

NO

Are GenAI Outputs Speech?

61

NO

• Free speech protections have never extended beyond humans (it’s been litigated).
• Even gorillas that are sentient, self-aware, and can intentionally communicate ideas do not receive free speech protections, and people have no First Amendment right to receive gorilla communications.

Are GenAI Outputs Speech?

62

• Whose speech would it be?

Are GenAI Outputs Speech?

63

• Whose speech would it be?
• The developer’s?
• They don’t know what the model will say to every given input (anyone at Google expect Google to suggest glue on pizza?)

Are GenAI Outputs Speech?

64

• Whose speech would it be?
• The developer’s?
• No
• The model’s?
• No intentionality, not sentient, not self-aware
• Q: What if humans have no free will, so nothing is really intentional?
• A: If we have none, then the First Amendment would be the least of our legal problems. Other intentional crimes and torts would not be enforceable:
• Murder, rape, robbery, arson, battery, etc.

Are GenAI Outputs Speech?

65

• Whose speech would it be?
• The developer’s?
• No
• The model’s?
• No
• The prompter’s?
• The prompter doesn’t control what the model will output (anyone expect Google to suggest glue on pizza?)
• Notably, prompters also don’t own copyrights in the outputs

What about the right to receive speech?

66

• If there is no speech, there is no right to receive it.

What about the right to receive speech?

67

• If there is no speech, there is no right to receive it.
• Suppose you found the gurgling of a stream to be profound. It causes you to have a transcendent religious experience that is radically important to you.

What about the right to receive speech?

68

• If there is no speech, there is no right to receive it.
• Suppose you found the gurgling of a stream to be profound. It causes you to have a mystical religious experience that is radically important to you.
• Does that make the stream a speaker?

What about the right to receive speech?

69

• If there is no speech, there is no right to receive it.
• Suppose you found the gurgling of a stream to be profound. It causes you to have a mystical religious experience that is radically important to you.
• Does that make the stream a speaker?
• Does that make the gurgling speech?

What about the right to receive speech?

70

• If there is no speech, there is no right to receive it.
• Suppose you found the gurgling of a stream to be profound. It causes you to have a mystical religious experience that is radically important to you.
• Does that make the stream a speaker?
• Does that make the gurgling speech?
• If the government wanted to dam the stream it would have to satisfy high levels of First Amendment judicial scrutiny?

What about the right to receive speech?

71

NO

And yet…

72

• The majority of elite legal scholars seem to believe GenaI outputs are protected speech.
• (Prominent ones also get paid by tech giants to consult…)

​

• If so, this would be the only non-human entity to have such rights

And yet…

73

• The majority of elite legal scholars seem to believe GenAI outputs are protected speech.

​

• If so, this would be the only non-human entity to have such rights
• (corporations have speech rights not because they are people under the law, but because they are made up of people. It’s the rights of those people the courts recognize, not the rights of the corporation itself.)

​

THIS IS GENAI EXCEPTIONALISM

We treat it as mystical because it’s complex.

And that is B-A-N-A-N-A-S!

PART 5: GENAI EXCEPTIONALISM AND CONTRACT LAW

74

Browsewrap v. Clickwrap

75

• Very demure
• Not typically enforceable

Browsewrap v. Clickwrap

• Ostentatious
• Enforceable

76

Notice

• No notice
• Inquiry Notice
• Constructive notice
• Actual notice

​

77

Scraping Bots

• Bots are not fully autonomous. They do what they are programmed to do.
• So, the deployer is liable for anything the bots do
• If your bots hit a server too frequently and the server crashes, there is no “oopsie, it was my bot’s fault, not mine” defense.

78

Scraping Bots

• Bots are not full autonomous. They do what they are programmed to do.
• So, the deployer is liable for anything the bots do
• Bots often scrape every page on a website they can access
• This would include the legal terms page
• “Seeing,” downloading, and storing the terms would all indicate actual notice
• This would make the terms enforceable

79

Scraping Bots

• Bots are not full autonomous. They do what they are programmed to do.
• So, the deployer is liable for anything the bots do
• Bots often scrape every page on a website they can access
• This would include the legal terms page
• “Seeing,” downloading, and storing the terms would all indicate actual notice
• This would make the terms enforceable

If a site has a “no AI” clause (i.e., not allowed to scrape or not allowed to use data to train AI), then the GenAI company would be in breach of contract if it used the site’s info to train a GenAI model

​

80

Would This Logic Only Apply to GenAI?

• No.
• But GenAI is notable for the orders of magnitude difference in the scale of its scraping
• Llama 3 trained on 15 trillion tokens
• There is a reason GenAI companies use their own scraping bots and not just public datasets like Common Crawl

81

Exceptionalism

• If we instead say, “Well, sure, a human would be liable, but a bot deployed by a human doing the same actions is not liable”…

82

Exceptionalism

• If we instead say “Well, sure, a human would be liable, but a bot deployed by a human is not liable”…

THAT WOULD BE GENAI EXCEPTIONALISM

83

Exceptionalism

• If we instead say, “Well, sure, a human would be liable, but a bot deployed by a human is not liable,”…

​

THAT WOULD BE GENAI EXCEPTIONALISM

​

• It would also have the dubious logic of saying, “you can subvert the law if you have a piece of software do the same thing a human would have done”

84

PART 6: GENAI EXCEPTIONALISM AND COPYRIGHT LAW

85

Copyright Law

• Protects original express when fixed in a tangible medium

86

Copyright Law

• Protects original express when fixed in a tangible medium
• Bestows exclusive rights:
1. to reproduce the copyrighted work in copies or phonorecords;
2. to prepare derivative works based upon the copyrighted work;
3. to distribute copies or phonorecords of the copyrighted work to the public by sale or other transfer of ownership, or by rental, lease, or lending;
4. in the case of literary, musical, dramatic, and choreographic works, pantomimes, and motion pictures and other audiovisual works, to perform the copyrighted work publicly;
5. in the case of literary, musical, dramatic, and choreographic works, pantomimes, and pictorial, graphic, or sculptural works, including the individual images of a motion picture or other audiovisual work, to display the copyrighted work publicly; and
6. in the case of sound recordings, to perform the copyrighted work publicly by means of a digital audio transmission.

87

Purpose of Copyright Law

• Copyright law was created pursuant to Congress’s express constitutional power…

“To promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries”

88

Purpose of Copyright Law

• Copyright law was created pursuant to Congress’s express constitutional power…

“To promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries”

89

• So, the key question when analyzing copyright should be…

Would the outcome promote or inhibit the progress of science and the useful arts?

Fair Use Factors

1. The purpose and character of the use, including whether such use is of a commercial nature or is for non-profit educational purposes;
2. The nature of the copyrighted work;
3. The amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
4. The effect of the use upon the potential market for or value of the copyrighted work.

90

GenAI Outputs Might Be Fair Use

• The outputs typically don’t resemble the inputs.
• Show me the source that ChatGPT copied when it instructed how to remove a peanut butter sandwich from a VCR in the style of the King James Bible
• Model outputs don’t usually compete with training data (e.g., even if trained on a novel, ChatGPT can’t produce a similar novel)

91

Why It’s Important

• If it’s fair use, then GenAI companies don’t have to pay for the songs, movies, images, videos, books, essays, code, and other material it trained on.
• Fair use = no copyright infringement
• Copyright Act preempts contract claims in some circuit courts

92

But What About Humans?

• EVERY argument that works in favor of GenAI works at least as well for humans
• Transformative?
• Humans are far more likely to innovate!
• Ever hear of calculus? Cubism? Evolution? General relativity? Honky tonk?
• How many new scientific theories has GenAI developed? New fields of study? New genres of music or movies?
• How many by humans? ALL OF THEM!

93

But What About Humans?

• EVERY argument that works in favor of GenAI works at least as well for humans

​

• Competitive?
• Know anyone who can recite a chapter from Harry Potter from memory?
• How many times have you sold an album of songs that sounded substantially similar to Taylor Swift?
• Write any books substantially similar to Stephen King’s recently?

​

94

Idea-Expression Dichotomy

• The last refuge of GenAI exceptionalism arguments
• Copyright law protects expression, not facts

95

Idea-Expression Dichotomy

• With GenAI, the dichotomy is a fiction
• There is no way to argue that GenAI only uses the ideas of works it’s trained on, but humans only use the expressions
• If GenAI and I both write a short story using characters, mannerism, style of dialogue, settings, and more of a popular book…
• How can we possibly say that when I write it, I relied on expressions but when GenAI generated it, it was only based on ideas?

96

Idea-Expression Dichotomy

• With GenAI, the dichotomy is a fiction
• The argument also presumes we know how our neurons process information.
• From book pages, to photons hitting our eyes, to electrochemical signals along our neurons and across synapses we somehow only retain expressions
• And when GenAI is fed tokens it somehow only retains ideas yet is able to regurgitate memorized source material
• (this is why there are privacy, copyright, contract, and other claims!)

97

The GenAI Should Be Allowed to Learn Like Humans Argument

• Humans pay for stuff we use either directly or indirectly
• Also…
• GenAI trains on many orders of magnitude more data
• GenAI processes data many orders of magnitude faster
• GenAI is more likely to memorize and regurgitate training data
• GenAI companies have far more market power than individual humans
• GenAI has the potential to massively disrupt marketplaces, unlike individual humans

​

Saying GenAI is like humans is a false equivalency

98

What Does it Mean If GenAI Companies Are Right?

99

What Does it Mean If GenAI Companies Are Right?

You never have to pay for any content ever again!

100

BELIEVING THAT GENAI SHOULD BE ALLOWED TO USE ANY COPYRIGHTED MATERIAL BECAUSE IT’S FAIR USE BUT NOT APPLYING THE SAME PRINCIPLE TO HUMANS IS…

GENAI EXCEPTIONALISM

101

Appendix

102

Exceptions to Exceptionalism

• There should be a general waiver for nonprofit scientific research
• To prevent data laundering:
• Probably must release under non-commercial license
• Probably must release for scientific use only
• Probably must gate access (verify identities of people who want to download the models/datasets)

103

Potential Copyright Penalties

• Statute allows $150,000 per willful infringement
• NYT alleges OpenAI willfully infringed on over 10 million articles
• 150,000 x 10,000,000 = $1.5,000,000,000,000 ($1.5 trillion)

104

Why a Person May Want Their Content to be Public While Not Wanting GenAI Bots to Scrape Them

• To generate revenue
• Subscriptions
• Licenses
• Pay for premium content
• Order merchandise
• Pay for services (e.g., consultation)
• Affiliate marketing
• Ad revenue

105

• To connect with visitors
• To build relationships with like-minded folks
• To share interests with fellow hobbyists
• To raise awareness of a topic
• To share an opinion and hope to convince others of a position

​

​

• To build a brand or reputation
• To gain attention of possible employers or an audience
• To gain attention of possible partners
• To create a fan base or following
• To establish their credibility

Exceptionalism’s Effects on Society and Democracy

• Copyright and Contract: Chilled speech and other expression
• People may be less willing to create content (less incentive to do so)
• People may be less willing to share content (may feel like an unpaid cog in the GenAI machine)
• First Amendment: GenAI outputs don’t necessarily further democracy
• GenAI has no intention, so it can’t try to persuade anyone of a viewpoint
• Overwhelming communication channels with AI-generated content dilutes the ideas/speech of humans
• Enhancing misinformation makes even the best possible outcomes from GenAI challenging. Even if GenAI came up with a brilliant solution to a problem, it would take humans to implement it (e.g., RNA vaccine was a breakthrough, but even pre-ChatGPT misinformation scared people from it)

106

First Amendment FAQs

1. Why don’t you provide examples of various design choices that might make the output expressive and discuss why the design choices employed by models like GPT-4 fail to achieve expressiveness?
1. The design choices do not matter. No outputs by the model is speech as recognized by the 1A.
2. The "stochastic parrots” argument, while interesting, is applied to the weaker argument (whether a generative AI model is a "speaker") rather than the stronger argument (whether generative AI outputs are "intentional"). Why not explore whether a "stochastic parrot" has "expressive intent" for First Amendment purposes?
• Even when outputs are intentional, they are still not speech if it’s from a non-human. Examples: gorillas and sign language, dogs begging, animals warning other animals of a predator approaching, etc.
3. Why is there no method or framework to address critical questions like the line between intentional and unintentional expression? Why not address the more challenging "close calls" like generative models, going beyond extreme examples of intentional and unintentional speech?
• Intentionality has prerequisites, including sentience and self-awareness. No amount of GenAI outputs will ever satisfy intentionality until the models are self-aware, so there are no close calls.

​

​

107

Ideal Policy Outcome

• Content creators (artists, writers, directors, actors, singers, etc.) receive both compensation and credit when their works are used.

​

• Why?
• The current economy is built around creators getting some benefit
• Without a benefit, the incentive to create and share is diminished
• Undermines democracy (only functions if public is well-informed and participates in the exchange of ideas)
• Impoverishes society (hinders the promotion of science and the useful arts)
• Also, GenAI can only thrive if people continue to create and share
• Primarily training on synthetic data has so far led to “model collapse” not model improvement

108