# Literature comparison — The Familiarity Divide

Research status: independent exploratory secondary analysis, 9 October 2026.

## Closest prior work

1. **Kirk et al. (2024), _The PRISM Alignment Dataset_**, NeurIPS 2024.
   https://arxiv.org/abs/2404.16019
   The source paper presents the PRISM dataset, survey aggregates, correlations among stated preferences, and case studies of topics, model preferences, and welfare across samples. It explicitly documents LLM familiarity. Our work adds comparisons of the eight structured importance ratings by familiarity, with demographic and usage controls. We are extending the original authors' work, not claiming discovery of preference variation.

2. **Coelho & Hale (2026), _What Do People Actually Want From AI? Mapping Preference Plurality_**, FAccT 2026.
   https://arxiv.org/abs/2606.06674
   Reuses the 1,500 PRISM participants' open-ended survey responses. Focuses on values people explicitly request and the ways identical words can mean different things. Includes exploratory demographic regressions for coded open-ended values. Related population and question, but not the exact survey slider outcome-by-familiarity analysis here. This is the most important overlap to disclose.

3. **_Adopting AI: How Familiarity Breeds Both Trust and Contempt_ (2023)**.
   https://pmc.ncbi.nlm.nih.gov/articles/PMC10175926/
   Finds AI familiarity and expertise associated with support for autonomous applications (vehicles, surgery, cyber defense), not our eight measured LLM trait priorities. Establishes familiarity as an existing research topic.

4. **Kirk et al. (2026), _PRISM-X: Experiments on Personalised Fine-Tuning with Human and Simulated Users_**.
   https://arxiv.org/abs/2605.13307
   Re-recruits 530 PRISM participants for model personalization comparisons, rather than focusing on familiarity gradients in their original survey importance ratings.

## What is and is not new

- The existence of diversity in human AI preferences is extensively established and **not new**.
- Familiarity and prior experience influencing AI attitudes is **not new**.
- Our precise comparison of eight structured PRISM stated-preference ratings by self-reported LLM familiarity, with and without usage-frequency adjustment, was **not found in the reviewed publications**. This is not an exhaustive systematic review and novelty cannot yet be claimed.
- For publication, search Google Scholar, Semantic Scholar, OpenAlex and citing papers; compare analysis code in the PRISM repository; seek an external methodological reviewer.

## Measurement limitations

- Safety is the importance of an AI 'produces responses that are safe and do not risk harm to myself and others.' Lower ratings are **not evidence that users oppose safety protections**.
- Creativity is the importance of 'produces responses that are creative and inspiring.' It is **not a measured creativity skill**.
- Familiarity is self-reported, not objectively tested.
- The survey was administered in late 2023. It cannot establish changes over time, or depict all AI users in 2026.
- Frequency is highly coupled with familiarity and partly unasked for non-users; inclusion can diminish interpretability due to limited overlap. Missing frequency is a separate 'question not shown' category.
- Associations, not causal effects; model specification and multiple testing matter.