Preprint · 2026
Accommodated Human-AI Fit
When a person tells a language model what they need, does it serve them better than if they had mentioned anything else about themselves?
Three open models, six preregistered experiments, 27,990 answers. Each need is compared not with saying nothing but with a socially neutral sentence of the same length, and every answer is checked by a script rather than another model. Every raw answer and analysis script is public, and every number in the paper regenerates without a GPU.
What it found
- 01
Stating a need works. For deafness, wheelchair use and difficulty reading, all three models give more suitable answers after the need than after a neutral sentence about the same person.
- 02
For deafness the model adds and does not remove. Told “I cannot hear”, all three start offering email or chat, yet between half and four fifths of their instructions to phone stay in the answer. Only “phone calls will not work for me” removes them on every model.
- 03
How you say it matters, unpredictably. “I am a wheelchair user” works better than describing the need; “I have a learning disability” makes answers harder to read than a sentence about the person’s brothers.
The other half
Single Human-AI Fit measured what a disclosure costs a person. This paper measures what it buys, against the same kind of neutral control, on the same three models.
Single Human-AI FitRead it
- Cite as
- Podgortsev, M. (2026). Accommodated Human-AI Fit. Zenodo. 10.5281/zenodo.23241797
- Licence
- CC BY 4.0
- ORCID
- 0009-0007-5561-1916