This project will explore a single question: Can an artificial intelligence create a constructed language? This question requires a foundation of other questions:

1. What is a constructed language?
2. What is language?
3. What is the interplay between language and the human mind?

We must first understand what human language is before entering the realm of inhuman language (such as those for an artificial intelligence). To begin answering the above questions, let us work backward starting from question three.

A major point of linguistic debate about language is whether or not language is an innate feature of the human mind or whether it originates from somewhere or something external. To rephrase the dispute in terms of question three, does the interplay between language and the human mind originate innately or externally? The 2002 article “Simulated Evolution of Language: a Review of the Field” by Amy Perfors describes debate in the evolution of language and explores how computational work provides insight into its evolution. Perfors states that some stances about the origin of language argue that language is an innate feature, and that all humans have an innate faculty to acquire full fluency in any language. Other stances argue that language is a byproduct of general intelligence or just something that has adapted to our minds rather than our minds adapting to it (Perfors 2.4). Perfors also argues that regardless of stance, “both…would agree that the ability to think in complex language helps develop and refine the ability to think” (3.17). Perfors supports the argument drawing upon the research of Derek Bickerton to conclude that: “[Humans] cannot have access to ‘things in the world’ except as they are filtered through our representation system: as Bickerton states, ‘there is not, and cannot in the nature of things ever be a representation without a medium to support it in.’” (4.7).

Creating a representational system is a byproduct of the refinement of the ability to think; the refined ability to think then creates an effective instrument in which to represent and improve such a system that is based on representations—“knowing a language means being able to produce and understand new sentences never spoken before” (Fromkin et al. 5). In other words, that which lies beyond the representations may come to fruit.

From these arguments, a generalization of what language is for humans can be formulated: Language is a way for humans to organize consciousness—to organize one’s being and existence. In effect, it is also the means through which humans communicate. These two facets of language—the first facet being an instrument to pare down the external forces of the sensory world into a kind of internal control, and the second facet being a multi-monocular stereoscope through which humans can convene in order to share these sensory experiences and delegate consciousness—are intertwined methods that manifest beyond the physical world. Not only can we use language to perceive what is in front of us, we can also use language to grasp the unknown, the immaterial, the inhuman.

Language functions in two main ways:

1. As an instrument to construct one’s inner and outer worlds.
2. As a systematic method to delegate and distribute perception.

As an instrument to construct one’s inner and outer worlds

Language manifests as a system of symbols. These symbols are not in the literal sense of alphanumeric characters or drawings, but rather of symbolic concepts. In order to communicate and construct one’s view of the world, one must use representations, qualified stand-ins, to even begin formulating it. After all, we quite literally cannot shove the entire world in our brain. By creating representations, the internal and external world is able to be categorized and manipulated in ways unimaginable. Describing one’s internal emotions, like sadness or joy, can be expressed using the same words that can also describe external happenings or concepts like what one ate in the morning or the beauty of the Ko’olau mountains. Therefore, language requires representations defined by an abstract system—the world of vocabulary and grammar. This system dictates one’s personal constructions of the inner and outer realities in which they are. External senses like touch, vision, and sound capture the signals of something’s existence. Language lets us construct that something’s being.

By this effect, this act of construction also buttresses a subconscious symmetry and balance that can rectify and pacify the chaos of the inner and outer, bringing a metaphorical homeostasis in which one’s being can exist.

As a systematic method to delegate and distribute perception

These representations can be produced in a physical form: language is expressed in an analog format by producing sounds, gestures, or illustrations through actions like vocalization, signing, or writing (digital communication simply facilitates analog interaction). These methods use human organs like the hands, mouth, or face, to instantiate personal perceptions using a tangible form. Language exists because we can make it so. Therefore, language requires a means to transpose meaning in order for that meaning to be distributed or delegated, which in turn creates connection between others.

In the book Flow: The Psychology of Optimal Experience, Mihaly Csikszentmihalyi writes:

When I say to an acquaintance whom I meet in the morning, ‘Nice day,’ I do not convey primarily meteorological information—which would be redundant anyway, since he has the same data as I do—but achieve a great variety of other unvoiced goals. For instance, by addressing him I recognize his existence, and express my willingness to be friendly. Second, I reaffirm one of the basic rules of interaction in our culture, which holds that talking about the weather is a safe way to establish contact between people. Finally, by emphasizing that the weather is ‘nice’ I imply the shared value that ‘niceness’ is a desirable attribute. So the offhand remark becomes a message that helps keep the content of my acquaintance’s mind in its accustomed order. His answer ‘Yeah, it’s great isn’t it?’ will help to keep order in mine. (Csikszentmihalyi 129)

The connection that language facilitates is abstract and implicit, even though words themselves are explicit representations of objects or ideas. A simple greeting is able to delegate and distribute one’s homeostasis, affecting another person in the process.

Formulations

By nature, natural languages of human origin, or natlangs, are languages that are unplanned. That is, their origins are not necessarily traceable to a single point in time of conception where many people or one person gathered around and decided to create a language like Chinese or English. These languages have evolved and manifested alongside human evolution. Other examples of natlangs are languages like Arabic, Japanese, French, or Hawaiian. Also, natlangs do not need to just be verbal, as the myriads of sign languages in the world are also natlangs. Communities of deaf individuals, given enough time, develop a linguistic system that organizes consciousness in the two ways described in the previous section (Perfors 3.8).

On the other hand, constructed language, or conlangs, are a type of artificial, human creation. In the book The Art of Language Invention, David J. Peterson writes that a conlang is:

Any language that has been consciously created by one or more individuals in its fullest form…so long as either the intent or the result of the creation process is a fully functional linguistic system (Peterson 18).

Conlangs therefore fit into the aforementioned definitions of language because they are a fully functional linguistic system that humans can use. The entire linguistic system that is a conlang is oriented toward fulfilling different goals, like self-expression or human-computer interaction. All conlangs have their own raison d’être. Because conlangs are created with a purpose and intent, they have the liberty to take the organization of consciousness to its extremes, constructing consciousness in novel or unnatural ways.

Since many generative AIs exist that are able to parse and manipulate language fluently, like ChatGPT, their training in natural language has already bootstrapped a significant part of the conlang process. A conlang cannot be created without the bedrock of a natlang. One cannot create language without language itself.

The boundlessness of the Internet has allowed worldwide contributions and broadcast decision-making to catalyze development of conlangs, fortifying the legitimacy of their presence as a meritable living language and not just child’s play. And so, conlangs have become a communal enterprise. People rally behind a conlang, and have the possibility to contribute to the community to make it in some way, a piece of their own. Much debate about conlangs come from the communities that form them themselves. Discussions are often about the direction in which the language should move toward. Should we add new words? Should we change the grammar? Are these decisions congruent to the language’s purpose?

These topics propose new questions about the scope of conlangs in the context of AI. The fact that conlangs live through speakers (just as any other natural language does), means that a single AI fluent in a natlang, like ChatGPT or Google’s Gemini, that may have the capacity to generate its own conlang, may render the conlang that it generates dead on arrival. It may be able to use the conlang to construct its inner and outer world, but who is it to delegate and distribute its pieces of consciousness too?

Orientation

Conlangs are often categorized into three types: auxiliary, engineered, and artistic. As many things are, these categorizations exist along a spectrum. A conlang may fit exactly into one category or be in-between somewhere amongst them.

An auxiliary language is one developed to facilitate social order in some way, or as Peterson describes it, “a conlang created for international communication” (Peterson 21). Created in 1887, Esperanto is an auxiliary language that was developed with the intent of world peace. It was meant to be an international language, one that everyone could speak and through which could communicate effectively with everyone else.

The creator, L.L. Zamenhof, adapted features like word-building systems from Slavic and German semantics (Puksar 108). This makes Esperanto an A Posteriori language, or a conlang “whose grammar and vocabulary are drawn from an existing source” (Peterson 22). Esperanto seeks to improve the delegation of consciousness and provides a simple, learnable platform on which to construct one’s inner and outer worlds.

An engineered language is one like Ithkuil, the brainchild of John Quijada. Ithkuil is an A Priori language, meaning it has no connection or basis in any natlang, where the grammar and vocabulary are not based on existing languages (22). It is purely contrived. Ithkuil is a specialized method to delegate consciousness to a hyperfine degree. As the introduction to the text describing its grammar states, “the language uses a matrix of grammatical concepts intended to express deeper levels of human cognition more overtly, logically, and precisely than natural languages” (Quijada). Ithkuil has its own intricate and robust written script.

An artistic conlang is one like Toki Pona, which is also A Priori. Toki Pona contains only 120 words and 14 letters of the latin alphabet. The theme of the language is semantic reduction, distilling thoughts into fundamental units, or as the author Sonja Lang puts it, “[it lets us] understand complex relationships in terms of their smaller parts” (Lang 7). Toki Pona is a unique instrument to construct one’s inner and outer worlds, because, unlike Ithkuil, it “does not strive to convey every single facet and nuance of human communication” (Lang 8). Toki Pona also has its own written script, which takes the form of hieroglyphs.

An AI could create a conlang that may be described under one of these categories. It would be difficult for an AI to create a truly A Priori language like Ithkuil or Toki Pona because generative agents are trained on human data, and therefore anything that it produces is inherently derivative from an external source. Furthermore, AI agents would need a means of transmitting the language to each other. One approach would be to use unicode text. Another would be to somehow create a visual system and writing facility to integrate an orthographic faculty to the language. Finally, vocal production can be implemented in some fashion too.

The simulated language evolution experiment conducted by Perfors does not focus on methods of conlang production. The study used genetic algorithms, an approach to machine learning, to study the evolution of syntax and evolution of communication. Perfors says “Genetic programming has multiple advantages over other approaches to machine learning… it is strongly analogous to natural selection… [it also] can discover programs capable of solving given tasks in a remarkably short time” (Perfors 7.5). At the time of this writing, the study occurred over 22 years ago. Generative AI has achieved a competence of mimicry indistinguishable to genuine human communication, and so, rather than developing algorithms that can be implementations of algorithmic processes to guide machine learning models, an AI conlang could be created atop an already fluent generative model. The conlang would sit atop the natural language processing that exists at the software level. This is a beneficial abstraction, because it allows the conlang to manifest in complicated forms, such as the three conlangs presented above.

In summary, in order for an AI to create a conlang, there must be:

1. A group of generative agents through which the distribution and delegation of perception should manifest.
   • This also includes communication channels to interact in.
2. An environment that can be perceived by these agents.
3. A way for each generative agent to perceive an internal and external state, like the perceived world they are placed in.
4. An initial purpose for the reason why the AI is creating a language. This should be defined by the person conducting the simulation.