Same Question, Different Answer? Measuring and Mitigating Prompt Privilege for Equitable AI Access

Synthesis: Jin and colleagues (2026) identify a systematic Accessibility bias they call prompt privilege: users who can phrase requests skillfully get markedly better output from large language models than users who express the same underlying intent less adroitly. To make this measurable and fixable, they propose a Prompt Equity Score (PES) that quantifies performance consistency across user populations and a Prompt Equity Transformer (PET) agent that rewrites a user's request into semantically equivalent, accessibility-oriented wording before it reaches the model. Demonstrations on the MedQA benchmark show statistically significant performance gaps between low-literacy and expert-prompting cohorts that PET eliminates while preserving meaning — advancing a system-centered view of equitable AI access in which the burden of effective Prompt Engineering shifts from the learner to the AI itself.

Key Findings

1. Prompt privilege is real and measurable. On the MedQA benchmark, model accuracy rose monotonically with prompt sophistication: the Low Literacy cohort scored lowest (82.4%), followed by the Original Question, Everyday User, and Educated User cohorts, while the expert Prompt Engineer cohort scored highest (83.4%) — a statistically significant low-literacy versus expert gap.
2. The Prompt Equity Score (PES) formalizes the problem. PES is defined as one minus the coefficient of variation of accuracy across user cohorts, yielding a normalized measure of performance consistency (baseline PES = 0.9959) in which higher values indicate greater prompt equity.
3. PET is an accessibility-oriented prompt-normalization agent. Placed between user and model, the Prompt Equity Transformer rewrites requests into semantically equivalent, clearer prompts while preserving intent, shifting the burden of Prompt Engineering from the user to the system.
4. PET eliminates the statistically significant disparity. After prompt normalization, every pairwise confidence interval overlapped zero and the Low Literacy cohort's accuracy rose to match the expert upper bound (82.4% → 83.4%), so no accessibility disparity remained statistically significant.
5. The mechanism is semantic-preserving reorganization. PET improves equity by enhancing the presentation of existing information — e.g. rewriting informal "3 HIV medicines" as "three-drug antiretroviral regimen" and restructuring long clinical narratives into standard sections — without adding or altering clinical content.

Prompt privilege as an overlooked accessibility gap

The paper reframes prompt sensitivity. Existing work on prompt robustness focuses on adversarial attacks, prompt injection, and prompt optimization — all concerned with whether an attacker or expert can extract more from a model. Jin et al. instead ask the inverse question: whether equally legitimate requests phrased differently receive different-quality responses simply because of phrasing. They call this accessibility challenge "prompt privilege" — users with greater prompting expertise systematically obtain better performance despite identical intent. This is fundamentally an equity problem: as LLMs are woven into healthcare, education, and public services, anyone with weaker prompt literacy is silently disadvantaged, compounding the Digital Divide.

The authors situate the concern within human-centered and trustworthy interaction design: because conversational AI increasingly mediates access to information, a user's ability to phrase an effective query becomes a new, largely invisible axis of access. Prior trustworthiness and debiasing work treats equitable, reliable model behavior as something that must be actively engineered rather than assumed — a premise the paper extends from security robustness into the non-adversarial, everyday setting where users are not trying to manipulate the model yet still receive systematically unequal outcomes.

Making the gap measurable: Prompt Equity Score (PES)

To move prompt privilege from anecdote to measurement, the authors propose the Prompt Equity Score (PES), a quantitative Benchmark metric that evaluates how consistently a model performs across different user populations (e.g., low-literacy vs. expert-prompting cohorts). Formally, PES equals one minus the coefficient of variation of per-cohort accuracy, providing a normalized, confidence-aware measure of performance consistency. A low PES signals that performance depends on who is asking — exactly the kind of disparity an equitable system should eliminate. This connects to the knowledge base's broader concern with fair, valid evaluation of AI systems and with how confidence-aware metrics expose hidden bias in AI behavior, echoing the logic of rigorous AI evaluation and psychometrically grounded measurement.

The controlled design generates, for each original MedQA question, four additional semantically equivalent prompt variants representing distinct user cohorts — Low Literacy, Everyday User, Educated User, and Prompt Engineer — holding the underlying clinical reasoning task and model (GPT-5.4-mini) fixed. Because only linguistic characteristics such as vocabulary complexity, terminology, and sentence structure are modified, any observed performance difference can be attributed to prompt formulation rather than task difficulty or information content.

Mitigating it: Prompt Equity Transformer (PET)

The proposed remedy is architectural rather than instructional. The Prompt Equity Transformer (PET) is an LLM-based agent placed between the user and the foundation model; it automatically transforms a user's request into a semantically equivalent, accessibility-oriented prompt while preserving intent. PET thereby shifts the burden of Prompt Engineering from the end user to the system, functioning as an intelligent accessibility layer. The authors argue this is preferable to asking low-literacy users to learn better prompting — a strategy that further privileges the already-skilled — and instead makes equitable output a property of the system. This aligns with system-centered accessibility and universal design thinking, particularly universal design for learning: rather than requiring the user to adapt to the tool, the tool adapts to the user.

Formally, PET seeks to maximize downstream task performance subject to a semantic-preservation threshold between the original and transformed prompts. Its qualitative behavior confirms this intent: in one example it upgrades informal phrasing ("3 HIV medicines") to clinically standard terminology ("three-drug antiretroviral regimen"), and in another it reorganizes an unstructured clinical narrative into History, Associated Features, Current Medications, and Physical Examination sections — improving linguistic precision and the salience of diagnostically relevant evidence without introducing new medical information.

Experimental evidence on MedQA

Experiments on the MedQA benchmark confirm the problem and the fix. Before PET, accuracy improved monotonically with prompt sophistication, and only the Low Literacy versus Prompt Engineer comparison reached statistical significance (a one-percentage-point gap with a 95% bootstrap confidence interval excluding zero). After applying PET as a preprocessing step, all user cohorts benefited — most the Low Literacy cohort, which rose from 82.4% to 83.4% — and no pairwise comparison remained statistically significant, with every post-PET confidence interval overlapping zero. The authors interpret this as demonstrating that prompt-quality should be treated as an accessibility variable rather than solely a user responsibility, and that accessibility-oriented normalization can substantially improve equitable access to medical and general-purpose AI.

What this means for practice

• Instructors. Audit the AI tools you assign for prompt privilege before assuming they serve students equally: with the underlying clinical task held fixed, accuracy rose monotonically from 82.4% for low-literacy phrasing to 83.4% for expert phrasing, so a student's wording — not their question — changes the answer they get.
• Instructors. Do not make better prompting a prerequisite for good AI help. Teaching students to phrase requests more skillfully further privileges the already-skilled; prefer tools that normalize input, or supply model phrasings, so that prompt literacy is not the gateway to quality output.
• Administrators. Measure performance consistency across your actual student populations, not just expert users. A single headline accuracy number hides phrasing-driven disparity that a metric such as the Prompt Equity Score (baseline PES = 0.9959) makes visible, and low PES should block deployment of a tool in an inclusive classroom.
• Designers. Engineer accessibility normalization into the system layer. The Prompt Equity Transformer lifted the low-literacy cohort from 82.4% to 83.4% — matching the expert upper bound — and removed every statistically significant pairwise gap without adding or altering clinical content, so vendor evaluation should ask whether the tool adapts to the user's phrasing rather than only to their intent.
• Researchers. Treat prompt formulation as an accessibility variable in evaluation, alongside the usual debiasing and validity checks: the disparity here was invisible to accuracy averaged over a single phrasing, and only a cohort-contrast design surfaced it.

Limitations

• The demonstration is a healthcare benchmark (MedQA) evaluated on a single model family (GPT-5.4-mini); generality across other domains and models is untested, and no classroom outcome was measured.
• User cohorts were generated by controlled prompt rewriting rather than sampled from real users with authentic literacy and communication differences, so the reported gaps are simulated rather than observed.
• The baseline disparity is small: only the low-literacy versus prompt-engineer pair reached significance, at a 1.0 percentage-point difference with a 95% CI of [−1.9, −0.1], and the baseline PES of 0.9959 already indicated largely consistent behavior across phrasings.
• PET currently operates as a static normalization agent: adaptive capabilities such as clarification questions or personalized rewriting strategies — and broader human-in-the-loop design — are left to future versions.

Citation

Jin, L., Hu, L., Shen, B., Cai, H., & Xin, Y. (2026). Same Question, Different Answer? Measuring and Mitigating Prompt Privilege for Equitable AI Access.