Generative AI’s Text-Only Design Excludes Deaf Signers, Conceptual Paper Argues

Generative artificial intelligence has transformed higher education, but its interaction model rests on an assumption that excludes millions of learners: that every user thinks and communicates through written text. A conceptual paper published in Discover Artificial Intelligence argues that the architecture of large language models, the design of prompting interfaces, and the framing of AI literacy as a text-based competency systematically disadvantage deaf signers whose first language operates in the visuo-spatial modality. The author, Bruno José Betti Galasso, introduces the concept of phonocentrism in generative AI as a theoretical lens and proposes a new framework, the Visuo-Spatial AI Interaction Model, for redesigning human-AI interaction from the ground up.

The scale of the problem is considerable. The World Federation of the Deaf estimates approximately 70 million deaf people worldwide, many of whom pursue higher education. For those whose first language is a signed language such as American Sign Language, British Sign Language, French Sign Language, or Brazilian Sign Language, interacting with a text-based AI system requires what the paper terms a double translation: from visuo-spatial thought in the signed first language to written text in a spoken second language, and from that general second-language writing to the specific register of an effective prompt. Each translation layer imposes cognitive load, and the cumulative effect compounds across the many iterations a single AI interaction typically requires.

The theoretical core of the argument adapts Jacques Derrida’s critique of phonocentrism, the privileging of speech and its written transcription as the primary vehicle of meaning. The paper is careful about this claim: it does not argue that large language models literally privilege sound, but that they inherit and operationalize the assumption that meaning is canonically carried by a linear, phonographically derived symbol stream. Three features of current systems support this reading. Training corpora consist overwhelmingly of phonographic writing that transcribes spoken languages, with signed languages almost entirely absent. Tokenization processes one-dimensional character strings, mirroring the temporal unfolding of speech but incompatible with the simultaneous, multi-articulatory grammar of signed languages. And the interface itself presupposes a literate user comfortable with extended written composition. Notably, notation systems for signed languages, such as SignWriting, are effectively absent from training pipelines, benchmarks, and interfaces alike, even though they are equally textual in the broad sense.

The paper attributes this bias to contingent historical and engineering choices rather than deliberate exclusion: digitized written corpora were abundant and cheap, tokenization over character strings was tractable, and dominant evaluation benchmarks were themselves textual. Yet the author cautions that framing the bias as merely technical risks naturalizing it. Contingent choices harden into infrastructure, and infrastructure encodes a norm; the effect on deaf signers is exclusionary regardless of intent.

Empirical evidence on the cognitive dimension remains partial, and the paper is explicit about its epistemic status. Two components are directly supported: research on American Sign Language-English bilinguals shows that sign language translations are automatically activated when deaf readers encounter single written words, even when task-irrelevant, and skilled second-language composition is known to depend on limited working-memory resources. Two further components are advanced as testable predictions rather than established findings: that sign co-activation persists during extended prompt composition, and that transcoding between modalities consumes working-memory capacity needed for metacognitive monitoring. No study has yet measured cognitive load in deaf signers composing AI prompts. The paper also draws on a survey of 80 deaf and hard-of-hearing participants, which found that while large language models offer benefits such as communication support and simplified text comprehension, the primary barriers are the lack of signed language support and insufficient understanding of Deaf cultural nuances.

The critical review that anchors the paper reveals what the author describes as the near-total invisibility of deaf learners across research, policy, and institutional practice. Systematic reviews of generative AI in higher education have examined academic integrity, assessment, institutional policy, and AI literacy, yet deaf and hard-of-hearing learners are virtually absent. Major policy documents fare little better. The UNESCO Guidance for Generative AI in Education and Research names inclusive use as a core principle but does not address signed languages specifically. The EU AI Act introduces accessibility requirements for high-risk systems, but does not engage the challenge of signed language interaction with language models. An analysis of generative AI policies at the top 50 U.S. universities found that while 94 percent had faculty guidelines, these focused overwhelmingly on academic integrity, and none specifically addressed the needs of deaf or signing students.

Prompt engineering, often framed as a new foundational literacy, comes under particular scrutiny. The paper argues that framing it as a universal skill assumes equivalent access to written text, when for deaf signers it operates in a secondary modality and demands command of written academic language, precise lexical choice, and iterative reformulation, all in a second language. The review also flags a compounding risk: studies have shown that GPT-based text detectors misclassified 61 percent of TOEFL essays by non-native English speakers as AI-generated while correctly identifying over 90 percent of native-speaker essays as human-written, a bias that could penalize deaf writers whose second-language texts exhibit similar features. The paper acknowledges a countervailing translanguaging literature noting that generative models tolerate non-standard input, but argues that this flexibility operates within the textual modality and cannot substitute for interaction across modalities.

In response, the paper proposes the Visuo-Spatial AI Interaction Model, organized around three dimensions: Input, Processing, and Output. The input dimension envisions visual prompting through diagrams, concept maps, and spatial layouts as primary communicative acts, direct signed language input without intermediary text, and spatial-gestural prompting that exploits the spatial grammar of signed languages. The processing dimension calls for spatial semantic representations, drawing on techniques such as graph neural networks, alongside cultural-linguistic contextualization incorporating Deaf cultural knowledge, and reduced cognitive mediation between intended meaning and machine interpretation. The output dimension prioritizes diagrammatic and spatial responses, signed language generation through avatar technologies such as SignLLM, and multimodal complementarity that strategically combines text, image, and sign. Three principles ground the model: modality equivalence, grounded in neurolinguistic evidence for the neurobiological equivalence of signed and spoken languages; cognitive sovereignty, drawn from the Deaf Gain framework and decolonial epistemology; and design from the margins, following the disability justice principle that accessibility designed for the most marginalized benefits all users.

The paper is candid about feasibility and limitations. Signed-language data remain scarce and expensive to collect and annotate; video-based sign capture raises acute privacy risks because a signer’s face and body are biometric identifiers; and current sign language production models remain unidirectional, text-to-sign, and suited mainly to short, announcement-style communications rather than extended academic interaction. Existing initiatives such as the Gallaudet University AI, Accessibility and Sign Language Center and Syracuse University’s pilot of an AI-powered American Sign Language interpreting service are welcomed, but characterized as treating deaf-AI interaction as a translation challenge rather than a design challenge. The model itself has not been empirically validated, and the author calls for comparative studies of prompt effectiveness across modalities, cognitive load measurements, and participatory design research with deaf communities positioned as co-designers and evaluators rather than test subjects.

The broader implication extends well beyond deaf learners. The paper focuses deliberately on early signers, for whom the mismatch between a visuo-spatial first language and a text-based interface is most acute, and notes that generalization to postlingually deafened adults, oral deaf learners, hard-of-hearing users, and DeafBlind individuals requires further investigation. But the design-from-the-margins principle suggests that systems built to accommodate visuo-spatial cognition would be more flexible, more multimodal, and more responsive to diverse ways of thinking, benefiting multilingual students, visual thinkers, and users whose cognition exceeds the boundaries of linear text. As a conceptual contribution, the framework offers researchers and developers a structured agenda: the phonocentric paradigm, the paper argues, constrains not only deaf users but everyone whose thinking does not fit the prompt box.

Subject of Research: Phonocentric bias in generative AI interaction and a proposed visuo-spatial interaction model for deaf signers

Article Title: Visuo-spatial cognition and the redesign of human-AI interaction in generative artificial intelligence

Article References: Galasso, B. J. B. (2026). Visuo-spatial cognition and the redesign of human-AI interaction in generative artificial intelligence. Discover Artificial Intelligence, 6(1), Article 950. https://doi.org/10.1007/s44163-026-02040-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02040-2

Keywords: generative AI, large language models, deaf education, signed languages, phonocentrism, prompt engineering, accessibility, Universal Design for Learning, Deaf Gain, visuo-spatial cognition, human-AI interaction, higher education