The Crungus Effect: Hallucinations or Neologisms?

The interface between human language and machine-generated text has never been more porous. As humans continue to extensively participate and incorporate Large Language Models (LLMs) like ChatGPT into tasks and the regular meaning-making processes, a peculiar phenomenon has emerged: the generation of lexemes that do not correspond to any known linguistic inventory. These "hallucinated" words, often dismissed as statistical noise or system errors, are beginning to demand reconsideration not merely as outputs of malfunction, but as nascent linguistic phenomena. In digital ecologies primed for novelty and memeification, such machine-generated ‘novelty’ is not just noticed but is also adopted, circulated, and can thus sometimes be canonized too. Is this neologism or mere hallucination?

Linguistic hallucinations occur when an AI model outputs a word that has no precedent in any known linguistic corpus. These, albeit typographical mistakes, can also be phonologically and morphologically plausible constructions that mimic the formal characteristics of real words. For instance, the term crungus, which surfaced from a DALL·E prompt, exemplifies this. Although devoid of semantic history, its morphological construction is evocative. Users then anthropomorphised it, circulated memes, and embedded it within fictional tales, thus effectively semiotizing a null signifier into a ‘meaningful’ sign.

This human response reveals an interesting dynamic that AI outputs, even those deemed erroneous, are recontextualized and can be semiotically activated by human users. In the vernacular churn of internet subcultures, hallucinated tokens become the loci of collective creativity. Its ontological status can easily shift from error to ‘artefact’.

Innovations or Errors?

The coinage of new words has always been intrinsic to linguistic evolution. Human speech communities constantly adapt their lexicons to reflect cultural, economic, political, sociological, and even technological transformations. Neologism can be of different kinds – Proper Neologism is when a new lexical item is created for a new concept, Transnomination occurs when a new lexical item is combined with an already existing concept, and finally, Semantic Innovations, wherein a new concept takes the form of an already existing lexical item. These neologisms emerge because of distinct cultural patterns and waves that lead to the deliberate creation of a word that is then, upon societal acceptance and usage, considered a neologism. 

What distinguishes the AI generation from human creation is the sheer lack of intentionality. Language models do not ‘invent’ in a teleological sense, and their neologisms are merely statistical byproducts of enormous training data. These lexical units also weirdly include name generation models, particularly baby name generation websites that flood the internet. Experiments have shown that these baby names are ‘novel’ and are generated from outside the trained set of baby names. This ontological ambiguity destabilises the distinction between ‘invention’ and probabilistic emergence. If linguistic legitimacy is conferred through usage and shared interpretation, then hallucinated lexemes have the potential to become real. Lexicographic institutions like the Oxford English Dictionary(OED) routinely incorporate words that were once considered ephemeral internet slang.

Neologisms needed in the human language?

It is a particularly intriguing epistemological development that, while AI models provide hallucinated linguistic outputs, researchers concurrently recognize the exigency of deliberately creating neologisms to mediate human-machine communication. Hewitt, Geirhos, and Kim argue that extant human lexicons, which are optimized for intra-human discourse, are structurally insufficient for encapsulating the divergent conceptual frameworks operative within AI systems. This lexical inadequacy precipitates communicative friction because machines and humans instantiate and operationalize knowledge differently.

To remediate this, the authors advocate for the systematic creation of neologisms capable of precisely indexing human concepts for machine interpretation, and vice versa (Fig.1). This endeavor is posited not as a peripheral novelty but as an infrastructural necessity for ensuring operational transparency in AI systems. Crucially, this programmatic creation of linguistic signifiers stands in stark contrast to the stochastic hallucination of words by AI models, the latter, as mentioned, lacking intentional semantic grounding or pragmatic orientation. 

Mimicking Meaning

Although the widely accepted Chomskian concept of the Universal Grammar inherent in humans contests the notion of machine cognition mimicking human cognition, there has lately been a significant body of work that states otherwise. In Can AI Mimic the Human Ability to Define Neologisms?, author Georgiou explores the degree to which AI can replicate the human skill of semantic comprehension in the case of neologisms involving blends (combinations of parts of two or more words), compounds (merging two whole words to create a single concept), and derivatives (new words formed by adding affixes to root words). Using an experimental protocol with Greek lexical formations, Georgiou compares human consensus definitions with those generated by ChatGPT.

The findings reveal nuanced differentials: high concordance in blends and derivatives, but significant divergence in compounds. This asymmetry underscores AI’s limitations in parsing complex morphological syntax and context-sensitive semantics. While neural models are adept at mimicking surface-level interpretive patterns, they falter when confronted with deeper etymological and syntagmatic relations.

Conclusion

Despite of its abilities, hallucinated lexicons pose epistemic and ethical risks. When invented terms are mistakenly perceived as established due to uninformed user usage, it can have severe repercussions, especially in medical, legal, and other socio-political contexts. A fabricated term, when presented to users who perceive AI as ‘trustworthy’, can simply misinform, mislead, and cause real-world harm. Moreover, hallucinations may unwittingly reproduce or amplify latent biases embedded in training data. A seemingly innocuous neologism could carry classist, casteist, racist, sexist, xenophobic, or other evil connotations. In such instances, the hallucination functions not merely as noise but as a vector of harm.

AI systems, even while devoid of semantic intent, act as inadvertent lexicographers. Their outputs are indexed, shared, and popularised by human agents. In this feedback loop, there is a clear chance of the machine speaking and users assigning meaning. As AI-generated language continues to creep into human language, discourses, and cultural production, users must approach such content with informed criticality. Not every novel word is meaningless, and not every hallucination is harmless. Understanding the mechanics, intentions, and risks behind machine-generated language is essential not only for researchers and computationalists but more crucially for everyday users consuming data in an increasingly automated linguistic environment. In the evolving ecology between human and machine, awareness is the first step toward agency.