Abstract
As Sign Language AI (SLAI) research experiences rapid growth, the field remains largely dominated by hearing, non-signing researchers. This demographic mismatch has resulted in systemic biases that prioritize computational convenience over linguistic accuracy and the lived experiences of the Deaf community. This paper presents a systematic review of 101 recent SLAI works, identifying a cycle of "technosolutionism"—the tendency to frame research as a "fix" for communication barriers while utilizing flawed datasets and reductive methodologies. We argue for a paradigm shift toward Deaf-led research agendas that prioritize linguistic grounding and representative data.
Introduction
Sign language recognition, translation, and generation are often presented as transformative technologies capable of bridging communication gaps. However, the current research landscape is frequently criticized by the stakeholders it claims to serve. When research is conducted without the leadership of Deaf and Hard-of-Hearing (DHH) individuals, it risks falling into the trap of technosolutionism: building tools to solve perceived deficiencies rather than addressing the actual needs of the community. This review examines the current state of the field to uncover how systemic biases are encoded into datasets, modeling choices, and research motivations.
Methodology
A team of five DHH researchers conducted a hybrid literature review and position paper to assess the current trajectory of SLAI. The corpus was established by retrieving 222 papers from arXiv published between January 2021 and November 2023 containing the term "sign language." After filtering for receptive models (those taking sign language as input) and excluding non-English works or reviews, 101 papers remained for systematic coding.
The authors developed a codebook to evaluate each paper across four critical dimensions:
- Motivation: How the research problem is framed (e.g., as a "bridge" for communication).
- Datasets: The demographics of the signers and the methods of data collection.
- Annotations: The labels used for training (e.g., glosses vs. translations).
- Modeling: The technical architectures and pre-trained models employed.
The Demographic and Dataset Disconnect
The review identified 43 unique publicly available datasets used across the 101 papers. A significant finding was the "Interpreter Gap": 41 out of 60 works focusing on continuous sign language relied on three legacy benchmarks—BBC-Oxford, RWTH-PHOENIX-Weather, and similar sets—that exclusively feature hearing interpreters.
Interpreted sign language, often referred to as "translationese," differs significantly from the natural signing used by native DHH signers in everyday contexts. Despite this, only 16 of the 43 datasets identified were composed solely of DHH contributors. Furthermore, 12 datasets failed to disclose the qualifications or hearing status of their contributors entirely. While newer benchmarks like ASL Citizen and Sem-Lex attempt to include more representative DHH signers, the field continues to struggle with defining fluency and avoiding "native signer bias."
Technical Limitations: Glossing and Transfer Learning
Two major technical trends contribute to the linguistic subordination of sign languages in AI:
The "Tyranny of Glossing"
Approximately 50% of the reviewed papers (51) rely on "glosses"—the practice of using a single spoken-language word to represent a sign. Glosses are not true translations; they are reductive labels that strip away the spatial, temporal, and non-manual components (such as facial expressions and body lean) that are essential to sign language grammar. By treating sign language as a word-for-word substitute for spoken language, researchers ignore its unique linguistic structure.
Inherited Biases in Transfer Learning
Many models utilize transfer learning from datasets like ImageNet (objects) or Kinetics (general human actions). Additionally, 34 papers utilized pose-based inputs from tools like MediaPipe that were not specifically trained on sign language. These models often fail to capture fine-grained handshapes or critical facial landmarks, leading researchers to adopt annotation schemes that prioritize what the machine can easily see rather than what is linguistically significant.
The Rhetoric of Technosolutionism
The review found that 64% of papers are motivated by the goal of solving "communication barriers." While seemingly noble, this framing often portrays Deaf individuals as "deficient" and ignores the complex, multimodal strategies DHH people already use to navigate the world. This perspective leads to "technological assimilation," where AI models trained on hearing interpreters or idealized signers may eventually force real-world Deaf users to adjust their natural signing to be understood by the software.
Conclusion: A Call for Deaf Leadership
The current trajectory of SLAI research risks creating tools that are linguistically inaccurate and socially exclusionary. To rectify these systemic biases, the authors propose a move beyond tokenism toward genuine Deaf leadership. This includes:
- Fostering DHH-Led Research: Creating institutional space for DHH researchers to set agendas and lead projects from inception.
- Positionality Statements: Requiring researchers to disclose their backgrounds, signing fluency, and biases as a standard practice in publications.
- Linguistic Grounding: Prioritizing representative data and complex linguistic features over the convenience of existing, flawed benchmarks.
By centering the Deaf community, the AI field can move toward "intentional" research that truly respects and accurately represents signed languages.
Summary updated