Introduction
Do first amendment theory and doctrine protect the rights of speech developed using contemporary artificial intelligence (or AI)? In this paper, I will assert that they should. Current implementations of AI using natural language may augment and supplant human-developed speech, but their development is still driven by human intervention and protected by first amendment theory and doctrine. The term "protected speech" indicates speech protected under the First (and the 14th Amendment to clarify federal oversight on state laws) to the United States Constitution. Whether a written speech or published work, AI produces content engineered within the parameters provided by a human author who owns the completed product. But there are looming challenges to this premise. The iterative design of AI may lead to a system of narrowly defined solutions lacking the transparency of trusted sources that are typically and intentionally exposed in modern question-and-answer systems. The manner with which for-profit corporations scrape information for training and reuse as a manner of reinforcement learning (RL) is a divisive mix of open-source content and intellectual property; masking established ownership that's traditionally identified in search systems, even under scrutiny. Ultimately, while ownership of speech currently rests with the human creator, the boundaries of authorship, ownership designation, and who is the recipient of free speech protection are blurring. The speed of changes to AI systems and the obscurity of how AI technology generates content for speech production forces a reexamination of regulations, theory, and doctrine in arguments in democracy and self-governance, the marketplace of ideas, and autonomy theories alongside the umbrella of First Amendment rights in the United States.
What is AI?
AI has grown significantly from its first programming development in 1956 as Logic Theorist. From the development of Feigenbaum systems in the 1980s to the creation of Deep Blue, AI systems were more narrow in their ability to execute singular tasks and logical understanding. Dragon System's speaking technology began to change in the 1990s with consumer-level AI interaction, and the new millennium saw the rise of social robots like Kismet. IBM's Watson won Jeopardy in 2011, the same year virtual assistants like Siri, Google Assistant, and Alexa emerged. After Massaro & Norton's 2016 paper and 2017 follow-up, AI progressed more rapidly: DeepMind's AlphaGo and Facebook AI dialog agents showcased advanced AI capabilities. In 2018, Alibaba's language processing AI and Stanford Reading highlighted the progress of AI research. Finally, we have witnessed the development of what some claim as the foundations of AGI in GPT-3 OpenAI 2020, Gato 2022, ChatGPT 2022, Midjourney 2022, and DALL-E 2022.
What is the difference between AI and AGI? Artificial General Intelligence (AGI) (better known as strong AI) possesses the understanding and cognitive abilities comparable to human intelligence:
"Artificial General Intelligence (AGI), that is, AI that matches human-level performance across all relevant human abilities."
In contrast, Artificial Intelligence (AI), better known as weak AI, is a system designed for a specific task and does not learn beyond its established programming. Regardless of outlandish claims, we have not achieved AGI (in relation to challenges such as Turing Test and the Chinese Room). We may never do so due to the primacy of implicit skills unique to human beings. In the meantime, we do our best to mimic humans, and that raises some of our concerns: with a lack of transparency about the mechanics of AI systems, it can be challenging to explain how AI makes a decision, what is the source of AI information, where human involvement begins or ends, and ultimately, if that information provided is reliable.
There are concerns about lack of transparency with AI that uses large language models (LLM) with natural language understanding (NLU) to communicate in human-like interactions. LLM are AI systems trained under supervision using large quantities of text with the ability to predict and reproduce solutions. In the past, LLMs required expertise to effectively engage and produce a reliable response. Utilizing NLU, LLM is better able to understand the complexity and adaptability of human speech. The results are systems such as OpenAI's GPT-2 through GPT-4, and the public variation, ChatGPT: able to parrot large quantities of information, attempt to predict answers, accept a wide variety of speech from people, and output responses by mimicking human speech.
What do we mean by speech?
First and foremost, when we say speech, we mean both traditional and nontraditional content. In this case, I argue that speech augmented by AI or wholly developed by AI with a human guide is protected by free speech theory and doctrine. Contemporary AI can develop a variety of speech in a written form to be expressed by humans such as political speech, entertainment, hate speech, vulgarity, blasphemy, false speech, prose, documented corporate expenditures, and truthful or non-misleading commercial speech. These methods of speech established with technology are distinctive from traditional performative tasks that utilize technology as a means to augment speech or garnish developed works. Here is a "speech product" is simply an end medium of speech, whereas a "communication tool" does not deliver speech but facilitates task notifications. Both are performative tasks that lack the complexity of AI:
"Generally, we can distinguish software that serves as a 'speech product' from that which is a 'communication tool.' Communication tools fall into the categories just described: they primarily facilitate the communications of another person, or perform some task for the user. In contrast, speech products are technologies like blog posts, tweets, video games, newspapers, and so on, that are viewed as vessels for the ideas of a speaker, or whose content has been consciously curated."
A car alarm does not garner first amendment protection, unless, perhaps, it is used within the creation of creative work (e.g., Olivia Rodrigo's song Driver's License has the seat belt chime a B3 and B4 in B♭ Major and is copyrighted). Yet, these examples are distinct ways in which technology embellish human created works, with no question of ownership, authorship, and speech recognition. As we shall see, while Wu's distinction provides a valuable foundation, it is an oversimplification to how AI integrates into contemporary speech creation without clear lines of delineation in human creation.
The usage of AI in establishing speech
One way to approach the topic is to exemplify the contemporary usage of AI in place of abstraction. There are limitless potential scenarios where AI may be used to develop speech as a means of public consumption—I will focus on a limited grouping to establish a preliminary overview and criticism. It is worth noting that while AI in the past has carried similar potential in their creation of works, no prior point in human history has LLM with NLU been made available under broad public consumption. Natural language (directly intended to obscure the designation of computer interaction and comfort the user into practical engagement as human as possible) in prompt and engagement makes it more challenging to assess the distinction between where human-created speech begins and how much of a human retains authorship, ownership, and agency in completed works. Let us imagine three primary degrees of AI involvement used to develop speech, which I will call alteration, delegation, and self-starting:
- Alteration: Speech works initially conceived and written by a human, but AI is used as an assisted instrument to polish or proof created works. This position utilizes a traditional human author and owner who uses AI technology as a tool to improve conceived work (e.g., Grammarly).
- Delegation: Speech works initially conceived by a human, but complex written speech is produced solely using an author-engineered series of prompts to the AI. This position is nontraditional, where a human author is developing speech using an AI system that obscures the source of material and the decisions surrounding how speech works are subsequently produced concerning output and completed work (e.g., ChatGPT).
- Self-Starting: Speech works are determined and executed through AI without being prompted in subsequent developments by a human user. While such execution may be the subject of strong AI (true, independent, artificial intelligence), it may also be the decisions and production of related work previously executed by associated technology. This position is also nontraditional, where a human is not engaged in any form of development beyond established guidelines (e.g., scripted systems alongside GoogleAds serving as regenerative content syndication platforms).
In our scenarios, Tyler from San Francisco, creating an opinion article for publication, may utilize contemporary AI through varying degrees of development. Tyler may conceive of an opinion article and write a preliminary draft. Tyler would then open ChatGPT and enter the following command prompt into the user interface:
"I want you act as a proofreader, please. I will provide you texts and I would like you to review them for any spelling, grammar, or punctuation errors. Once you have finished reviewing the text, provide me with any necessary corrections or suggestions for my approval."
ChatGPT would respond with a prompt requesting the article, Tyler would copy/paste the article into the user interface, and ChatGPT would generate a modified version for Tyler to reject or approve. The final speech work would be an alteration, where AI serves as an instrument (similar to complex typing assistants) to support the completion of Tyler's final work—whereas the conception of the idea and preliminary writing was produced by our human, Tyler. There is little question of authorship and ownership.
Alternatively, Tyler may conceive of an article, but instead of writing a preliminary draft, open ChatGPT and enter the following command prompt into the user interface:
"I want you to write me an article. Please write a 650 word opinion piece on why toxic speech is wrong in social media. Please take the position that toxic speech is dangerous to democracy in the United States."
ChatGPT would respond with an opinion article using exactly 650 words for Tyler to use, and Tyler may prompt ChatGPT further to address any errors in ChatGPT's logic. Once Tyler is satisfied with the changes and no longer prompts the AI, they may use copy/paste the work into their final format. That final speech work would be delegation, where AI produced a final written product solely through the prompt of a human. This decision-making level obscures how the AI forms written opinions, what source material it draws from, and whether the content is legally available for output. More concerning, the depth of human involvement in writing and creating the final work is speculative to an audience. There is no clear answer to authorship and ownership in the final produced work.
Finally, Tyler may publish articles through content syndication. While rules for publications would be enabled by Tyler, developing content would be a product of iteration without human intervention. Tyler would establish ground rules using automation software for publication, such as prompting desktop or web-based systems using coded webhooks in the following order:
- Connect to ChatGPT
- Write a 650 word opinion piece on why toxic speech is wrong in social media using opinion articles from public media in the United States over the past 30 days.
- Copy, paste, and store the articles in Tyler's database
- Publish the most recent article to all availably authenticated social media
- Repeat at random weekly intervals
Similar instances are available with existing software to automate online advertisements and the creation of fake blog posts, but modern implementations are intentionally deceitful. Tyler's initially developed article based on their webhook would be an example of delegation, but each subsequent iteration would be a refreshed opinion piece and self-starting; no longer conceived in full or engineered through prompt by a human author. The resulting pieces may bear the initial markers of human involvement, but each iterative article is a patchwork of opinions grafted from public works. There is no authorship or ownership by an identifiable human.
I should think only alteration would achieve authorship and ownership from a human, and only alteration would garner first amendment rights and protection under free speech theory and doctrine. My perspective is driven by the view that speech is a product of human language, forged within the context of cultural evolution that sees language as a uniquely human creation and, subsequently, laying claim to speech as human ownership. Yet, this thinking is circular; the act of creation and revision indicate ownership. That ownership may be seen as justification for protection. Such thinking is easy when comparing the speech of humans to animals. In today's time, speech can be reproduced through the complexity of AI systems designed to obscure how humans play a role in its authorship and ownership. Can a rational person tell the difference? No, and yet, the distinction matters as a means of authenticity in speech. While unsophisticated syndications devised under self-starting were easy to identify for a person experienced using the internet in the past, LLM with NLU makes content indistinguishable from a human author to a human reader. Regardless, as we will discuss, distinguishing human involvement in the creation of speech does not matter in current First Amendment protection, although it matters in relation to speech theory. While legal theory and philosophy are different, they can mutually inform one another and show potential gaps or blind spots when new technologies disrupt the status quo on which these laws were founded.
Theory & Doctrine that Protect AI Speech
Various theories and doctrines protect the rights of speech developed using contemporary AI. Each of these stands under the umbrella of First Amendment rights. There is a variety of legal precedents that validates nontraditional content, which we will discuss within our argument. But precedence aside, the continued progression of AI's place in speech is an increasingly unsettled view. Arguments in democracy and self-governance suggest that freedom of speech should extend to AI speech if it contributes to the democratic process. Alexander Meiklejohn notes the crux of such speech:
"What is essential is not that everyone shall speak, but that everything worth saying shall be said."
Speaker identity is irrelevant; if the speech aids the democratic process, it should be worth saying. In this view, the freedom to hear speech "secures cultural democracy as well as political democracy." If this is true in the age of AI, it may seem that defining works as alteration, delegation, or self-starting is immaterial when value is driven by cultural participation—influencing and protecting political democracy. But from Meiklejohn's 1961 perch, accessible public communication was necessary due to a dearth of mass media. Today, our problem is "no longer scarcity of access to mass communications, it is scarcity of human attention." Nevertheless, if the modern town hall is uncensored public discourse without limitations, AI is a valuable tool to develop and propagate self-determination regardless if nonhuman origins produce the speech.
The marketplace of ideas follows a similar premise, valuing expressions based on the enlightenment of the listener. The metaphor borrows from a free market analogy, where a rational person will strive to attain the best available product; in an open market of available goods, and with full and complete information to make an informed decision, the "best" product will always succeed. Comparably, the marketplace of ideas envisions good ideas will prevail, and bad ideas will naturally fall by the wayside. If so, it would seem content created by AI is inconsequential: the market will naturally separate falsehood from fact regardless of the owner, author, or means of developing speech. If we believe in a marketplace of ideas, this approach supports the First Amendment value for AI-generated speech as long as it contributes to exchanging ideas, disseminating knowledge, and aids a listener's discovery of truth. However, AI seems specifically vulnerable to the traditional argument that the marketplace of ideas works on false assumptions: that all persons have access to the market, truth is somehow objective and discoverable, truth always survives, and we are all rational beings. Validity of the marketplace aside, the ability to generate vast quantities of content quickly and without ownership risks the same scarcity of human attention previously noted. Like a needle in a rapidly multiplying haystack, it is easy to imagine a time when AI outpaces its human engineers, and we see more advanced systems extend the trend of toxic speech in a marketplace that looks to increasingly suppress the truth. AI might also push an insular set of ideas, eliminating the diversity of ideas without a fight. It would be far easier to destroy the truth by providing millions of people with a new definition of something framed as fact and never telling you where that idea was established. Whether speech is generated through alteration, delegation, or self-starting, they fit the spirit of the marketplace of ideas, even if such speech amplifies marketplace dilution.
Finally, it is in autonomy theories where additional cover for AI-developed speech resides. Modern legal personhood includes unions, municipalities, corporations, and objects protected by corporate governance, such as buildings or ships, to establish legal personhood. First National Bank of Boston v. Bellotti in 1978 defined the right for corporations to contribute to state ballot initiatives, granting a nonperson First Amendment rights for the first time in the United States. Citizens United v. FEC in 2010 allowed corporations, non-profits, unions, and other associations to advocate for a candidate (known as independent expenditure), granting further nonperson First Amendment rights. There is a history in the United States of protecting nontraditional speakers and their freedom to express their associated views through acts of speech. Antonin Scalia's famous statement interpreting the First Amendment as "speech, not speakers" makes plain that no category of speaker is excluded from free speech protection.
While autonomy theory may emphasize the autonomy of human listeners, there is a significant challenge to the idea that AI garners free speech protection: the view that personhood is a necessary trait for fundamental rights. It is difficult to explain why moral, human, personhood matters when free speech rights have been freely given to non-corporeal entities such as ships and buildings, which lack consciousness and personal identity. There is a palpable tension between accepting speech from a non-human lacking what we perceive as consciousness and a human in the creation of works. Much of this is grounded in the belief that human beings are novel agents producing and directing thought—manifested through the creation of speech. More importantly, that moral responsibility is wielded, and attributed, to human beings as the torch bearer for moral agency. Nonpersons such as corporations are created and held under human oversight, but the humans involved in supervision are subject to change, and elements of corporate speech are developed in collaborative and technological ways. Humans dictate corporate expenditures, but their monitoring, management, allocation, and processing are automated. Corporate expenditures, or spending money for a speech purpose, are used to support political expression and are protected by first amendment rights. These comfortably tuck in alteration, delegation, or self-starting as an example of AI-created speech, blanketing their produced content in the same protections as nontraditional nonpersonhood. This is not so readily accepted as a consciously-enabled moral agent developing things of opinion, statements of value, or creations for listeners.
In Closing
I agree that current implementations of AI, which can augment and compose human-developed speech are protected by First Amendment theory and doctrine at this moment. However, that is mainly from the view that AI is a tool prompted and engineered by humans who tell it what to do. Current implementations of AI using natural language may augment and supplant human-developed speech, but their development is still driven by human effort. These works are protected by first amendment theory and doctrine regardless if that speech is conceived, prompted, or automated with artificial intelligence. Even in cases where personhood is challenged as an establishment of rights, AI speech garners recognition through contemporary legal personhood status. Additionally, such speech would be protected by First Amendment rights unless it is used to target unprotected categories.
Nevertheless, the lack of transparency defining what prior sources AI has used to establish produced content, and their intellectual property status, is a concern surrounding ownership and authorship. These questions introduce further scrutiny if the work produced is expressive at scale, where the level of human involvement is opaque. However, this view may be driven by my fear of what AI would do to current speech development, and my lack of imagination of what AI may be able to do in the near future. History is littered with technology driving people to panic about the proliferation of accessible speech:
"The art of printing can be of great service in so far as it furthers the circulation of useful and tested books; but it can bring about serious evils if it is permitted to widen the influence of pernicious work."
AI may or may not have the same profound impact as the printing press did (nor drive people to destroy and censor them), but it shares a history of producing more speech with less work. Still, that history always carried specific engagements of personhood far less nuanced than AI development. Ultimately, current free speech doctrine may even privilege the creation of works with nonhumans over human speech to avoid risk for speakers. A speaker's intent is a condition to impose liability for harmful speech—if we allow amoral agents to be protected from speech, they become a loophole for corporations and humans alike to avoid liability.
Changing regulations to impose the requirement of a human as creator, owner, and responsible party for AI-developed works invites the risk that government will suppress and deprive listeners of potentially valuable expressions. Such legal oversight would also extend to accountability, copyright control, and the liability distinguishing between speech-related and non-speech-related harms under existing regulations. Ultimately, it may come down to a re-evaluation of the role of a human (not just a legal person) as a receiver of protection for free-speech doctrine and theory. I am not putting that message online in case future robots read this essay and come after me.