Lost in Translation: Understanding Autistic-Neurotypical Communication Style Differences in Job Postings
Huining Feng, Zinat Ara, Andrew Hundt, Slobodan Vucetic, John Joon Young Chung, Sungsoo Ray Hong · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3791853
Summary
This CHI 2026 paper investigates a concrete and consequential site of autistic-neurotypical (NT) communication mismatch: the written job posting. Autistic adults face an approximately 85% unemployment or underemployment rate (compared to 4% in the general US population), and prior work has pointed to communication style differences as a core driver of this disparity. However, no previous study had systematically documented which specific passages in written recruitment text cause autistic job seekers to feel 'lost in translation', or why. The authors built ANCAT, a custom web-based annotation tool implemented as a React single-page app with a Firebase/Firestore backend, that lets participants highlight any span of a job posting and assign one of six pragmatic-language categories - Unclear, Ambiguous, Incomplete, Inappropriate, Negative or Other - each with a written rationale. The categories were derived from foundational pragmatic-language work on autism by Tager-Flusberg and Landa. Twenty autistic adults (self-identified as DSM-5 Level 1, recruited via snowball sampling through university student support programmes, nonprofit vocational organisations and social media) each selected and annotated ten job postings from their own active job search, producing a public dataset of 683 annotations across 200 postings. The study used a three-stage protocol (onboarding interview, week-long annotation task, 60-72 minute closing interview averaging 1 hour 2 minutes) and analysed interviews with iterative qualitative coding. The team followed a disability-justice positionality statement, used identity-first language, included autistic co-authors in methodology and theme development, and treated participants as co-producers of knowledge rather than subjects.
Key findings
Surface lexical difficulties (long text, unfamiliar jargon) were only mild barriers; participants could usually resolve them by searching online. The real frictions came from implicit social arrangements and unstated expectations. Across the 683 annotations, Unclear was the most common category (35.1%), followed by Incomplete (16.6%), Ambiguous (15%), Negative (12.4%), Other (12.9%) and Inappropriate (7.9%). Five themes emerged. Non-social gaps included unmeasurable qualifications ('empathetic', 'ideal candidate'), unclear work routines (no description of a typical day), and frustrating abbreviations. Social-arrangement gaps surfaced the strongest reactions: phrases like 'team player', 'fast-paced', 'family-oriented' or 'outgoing' were widely read as red flags for overwork, conformity or manipulation - one participant called 'we're a family' 'manipulative or even invasive'. Accommodation gaps included hollow 'disability friendly' labels without specifics; participants strongly preferred concrete details (noise, lighting, meeting notes, break options). Navigation strategies combined search engines for lexical issues with trusted social contacts for social uncertainty; LLM use was mixed, with most participants mistrusting generative AI for factual or cultural interpretation. Suggested features included step-by-step application roadmaps, visual/bullet reformatting, day- and week-in-the-life routines, and tools that translate vague descriptors into measurable benchmarks.
Relevance
For accessibility practitioners and employers, the paper offers twelve concrete, employer-facing recommendations - including quantifying soft skills with examples, specifying daily and weekly routines, publishing salary ranges, avoiding coded culture phrases and naming actual accommodations rather than claiming to be 'accommodating'. For HR platform and AI tool designers, it makes the case for 'two-way translator' tools that can expand NT shorthand into explicit behavioural descriptions and flag vague or exclusionary language during posting drafting, while framing translation as an elective resource rather than a masking expectation. The released 683-annotation dataset is, to the authors' knowledge, the first public resource for studying autistic-NT communication-style differences in an employment context and is directly useful for training communication-accessibility tools and for empirical style-transfer research. Limitations include the US-only, 20-participant, DSM-5 Level 1 sample (findings may not generalise to autistic adults with higher support needs), the absence of an NT comparison group, and a focus on job postings as one gatekeeping stage - cross-neurotype communication mismatches extend throughout hiring, onboarding and daily work.
Tags: autism · autistic communication · communication style differences · autistic employment · double empathy problem · pragmatic language · workplace accessibility · natural language processing · large language models · data annotation · disability justice