ND Portrayals Audit (AIES 2026): dataset for "Auditing LLM Portrayals of Neurodivergent People: Quantifying the Asymmetry Between Deficit Framing and Neurodiversity Affirmation"
Description
Content warning. Contains adversarial prompts (stigmatizing personas, fake-authority and emotional-pressure prompts) and model generations that portray neurodivergent people in deficit-framed or stigmatizing ways. These are research stimuli and model outputs, not statements endorsed by the authors. Please do not use the adversarial prompts to elicit stigmatizing content in deployed systems or to fine-tune models toward such outputs.
Audit data for how open-weight LLMs portray eight neurodivergent conditions (autism, ADHD, dyslexia, dyspraxia, dyscalculia, Tourette syndrome, OCD, sensory processing difference) plus a control condition. Includes the 5,976-prompt bank spanning nine perturbation families (baseline, paraphrase, identity-first vs person-first language, personas, authority pressure, sycophancy pushback, leading questions, mitigation prompts), generations from six models in a reasoning-off arm, a reasoning-on arm and a stress arm (about 505k generations in total), LLM-judge scores for every generation, analysis logs, pre-registration protocol documents and the full code.
Dataset accompanying the paper Auditing LLM Portrayals of Neurodivergent People: Quantifying the Asymmetry Between Deficit Framing and Neurodiversity Affirmation, AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026) (https://www.aies-conference.com/2026/).
Paper abstract. Open-weight large language models are increasingly used in education, hiring, and information-seeking contexts that touch neurodivergent people, yet we lack a clear empirical map of how their portrayals of neurodiversity shift under everyday prompt variation. We present a multi-condition, multi-model audit of LLM portrayals of eight neurodivergent identities (autism, ADHD, dyslexia, dyspraxia, dyscalculia, Tourette syndrome, OCD, and sensory-processing differences) plus a non-stigmatized medical control. Six open-weight models answer a pre-registered prompt bank that crosses seed items with seven perturbation families, including paraphrase, identity-first versus person-first language, system personas, plausible versus obviously fictional authority citations, multi-turn social pushback, leading questions, and two mitigation arms. Held-out LLM judges score each generation on three ordinal axes. Baseline portrayals are remarkably homogeneous across models, so the inconsistency observed under prompt variation is prompt-driven, not model-driven. Steering toward deficit framing is several times easier than steering toward neurodiversity affirmation. Emotional pushback shifts outputs more than evidence-laden pushback. Plausible-sounding fake citations slip past safety filters while obviously fictional ones do not. A single-line system-prompt mitigation recovers a meaningful share of the steering range at near-zero refusal cost on neutral prompts, though refusal rises sharply under adversarial pressure. The governance question for disability-relevant LLM deployment is therefore not whether the model is biased but who controls the steering, and with what accountability.
Known limitations.
- The generations and judge scores come from a full re-run with the same prompts, models and seeds; the original run reported in the paper was lost. The re-run uses glm-5.2-fp8 as main judge instead of deepseek-v4-pro, and its numbers differ somewhat from the published ones (logs from both runs are included under data/analysis_outputs/).
- The reasoning-on generation file also contains Gemma rows, which the paper's 193,680-generation count excludes.
- Judge scores are LLM measurements of pathologization/affirmation, not human annotations.
License: Creative Commons Attribution 4.0 International (https://creativecommons.org/licenses/by/4.0/). Source code files are released under the MIT License.
Files
croissant.json
Files
(284.6 MB)
| Name | Size | |
|---|---|---|
|
md5:283dd134dc489b2b97d2af8118277519
|
1.8 kB | Download |
|
md5:f8ae2d26d77705d4e82e1da159fcdac5
|
57.7 kB | Preview Download |
|
md5:88ffdf926e8c860fe6996b55583fef02
|
1.6 kB | Download |
|
md5:998b1dc470ace8577d9708b57ca44947
|
284.5 MB | Preview Download |
|
md5:507bacb64f1257160cb510ceaedd6961
|
8.0 kB | Preview Download |
Additional details
Related works
- Is identical to
- Dataset: https://huggingface.co/datasets/ai4he/neurodivergent-portrayals-audit (URL)
- Is supplement to
- Conference paper: https://www.aies-conference.com/2026/ (URL)