ARTIFICIAL PARAPRAXES: FREUDIAN SLIPS, LATENT COMPETITION, AND UNINTENDED EXPRESSION IN GENERATIVE AI
The phenomenon of unintended expression occupies an important position in the psychological study of language, cognition, and behavioral regulation. In classical psychoanalytic theory, parapraxes, including slips of the tongue, were interpreted as potentially meaningful deviations arising from competition between intended actions and other psychological tendencies. Subsequent psycholinguistic research demonstrated that speech errors can also arise through ordinary mechanisms of lexical activation, semantic interference, production monitoring, and response selection. The emergence of sophisticated generative artificial intelligence raises a related but conceptually distinct question: can artificial systems produce parapraxis-like errors in which competing internal representations causally influence outputs that deviate from specified communicative objectives? This paper proposes artificial parapraxes as a provisional functional category for investigating such phenomena without attributing human intentions, unconscious desires, psychological repression, or subjective experience to artificial intelligence. A distinction is developed between ordinary generation errors and causally established expressive intrusions. The paper introduces the Latent Competition–Expression–Control Model (LCECM), which examines three interacting processes: the activation of competing representations, their influence on output generation, and the mechanisms through which unwanted expressions are inhibited, selected, or corrected. Five forms of parapraxis-like output are distinguished: lexical intrusion, semantic contamination, contextual carryover, regulatory leakage, and partial self-repair. These categories are interpreted as observable candidates for mechanistic investigation rather than as established psychological phenomena. The framework integrates insights from psychoanalysis, experimental psycholinguistics, spreading-activation theories, computational language production, mechanistic interpretability, and research on artificial self-correction. Six hypotheses are formulated concerning latent competition, contextual priming, regulatory inhibition, representational persistence, repair fidelity, and causal specificity. An empirical program is proposed using controlled semantic interference, matched generation tasks, activation interventions, monitoring manipulations, and comparisons between behavioral and mechanism-informed explanations. The theoretical contribution lies in establishing falsifiable criteria for determining when an apparent artificial slip can be explained through identifiable competition among internal representations rather than decoding variability, ordinary factual error, or retrospective anthropomorphic interpretation. The proposed framework does not claim that artificial intelligence possesses a Freudian unconscious, hidden desires, or involuntary psychological intentions. Instead, it offers a research program for investigating how computationally active alternatives may influence generated expression under conditions of incomplete regulatory control.
Authors
- Roman Shaban (ORCID: https://orcid.org/0009-0009-5259-6102)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24427565
- Primary Topic
- Neurobiology of Language and Bilingualism
- Type
- preprint