Understanding Human-AI Misalignment in LLM-Based Job-Seeking Support for Neurodivergent Users
Kaely Hall, Marcus Ma, Xinyue Zhang, Vedant Das Swain, Jennifer G Kim · 2025 · ASSETS 2025: 27th International ACM SIGACCESS Conference on Computers and Accessibility · doi:10.1145/3663547.3746361
Summary
This paper examines how misalignments manifest between neurodivergent job-seekers and a GPT-4-powered career support chatbot deployed by Mentra, a neuroinclusive employment platform with over 46,000 neurodivergent users. The researchers analysed 348 real-world chat logs from 271 unique users (60% autistic, 69.7% co-diagnosed with ADHD/ADD) and conducted semi-structured interviews with 15 neurodivergent participants. The chatbot offered four conversation topics — career advice, cover letter writing, interview preparation, and professional email writing — and was given a job-coach persona with access to users' profile data including work history, skills, interests, strengths, and diagnosis information. The study uses misalignment as an analytical lens to examine where and how the chatbot's outputs diverge from users' values, self-understanding, and goals. The analysis identified three major categories of misalignment: inauthentic characterisation of skills and experiences, failure to capture implicit preferences from user data, and misalignment stemming from user ambiguity about their own goals or the system's capabilities. The research contributes a bi-directional alignment perspective, arguing that alignment must address not only how AI adapts to human needs but also how humans can better understand and direct AI systems.
Key findings
The chatbot frequently misrepresented users' qualifications — in one case fabricating a Bachelor's degree in Graphic Design and claiming proficiency in skills the user did not possess when generating a cover letter. The system imposed neurotypical job-seeking language such as "quick learner," "works well under pressure," and "thrives in fast-paced environments" that participants found inauthentic and alienating, as these phrases describe work environments neurodivergent users specifically avoid. Despite having access to detailed user profiles, the chatbot made only surface-level matches, failing to interpret implicit preferences like preference for freelance work or non-hierarchical career paths. When users provided vague or emotionally charged inputs — common among neurodivergent individuals experiencing job-seeking anxiety — the system defaulted to generic, neurotypically-framed advice rather than prompting for clarification. Notably, participants who were more technically literate about AI could redirect the chatbot through explicit prompting to achieve better-aligned outputs, while less AI-literate users could not, creating an equity gap. Many participants lacked the confidence to correct the system when it responded inaccurately, sometimes internalising the misalignment as their own shortcoming rather than recognising it as a system limitation.
Relevance
This research has critical implications for the growing deployment of LLM-based tools in sensitive accessibility domains. The finding that neurodivergent users may internalise AI misalignment — believing the chatbot's inauthentic portrayal of their skills reflects their own inadequacy — highlights a significant psychological risk that extends beyond usability to well-being. For practitioners building AI-powered accessibility tools, the study underscores the danger of deploying general-purpose LLMs in specialised contexts without addressing their inherent neurotypical bias in training data. The proposed bi-directional alignment framework offers practical design strategies: making the AI's reasoning transparent so users can verify whether suggestions are grounded in their actual data, supporting collaborative correction where users can approve or modify AI interpretations, and using scaffolded dialogue to help users with unclear goals articulate their needs without defaulting to generic advice. These insights apply broadly to any AI system serving marginalised users whose experiences fall outside normative training distributions.
Tags: neurodivergence · large language models · employment · AI alignment · autism · ADHD · chatbot · job-seeking · AI bias · human-AI interaction