Editor’s Note: Hema Thakur is Manager of Skill Development at Cactus Communications, where she trains academic editors and supports early-career researchers in navigating scholarly publishing. She graduated with first-class honours in Banking and Finance from the University of London International Programmes and has served as an alumni ambassador for the university.

Simplifying language for people with cognitive disabilities isn’t just a stylistic or compliance exercise; it’s a design decision with ethical and cognitive implications. This is especially true when using generative AI to adapt complex material for users. In particular, it can be challenging to process and interpret peer reviewer feedback without support, given its dense, technical, and context-specific nature. But what happens when the simplifying tool distorts meaning in the process?

My recent research explored this question through a targeted experiment on the use of AI to improve the accessibility of peer review feedback. What I learned is important for every member of the scholarly communications and publishing community.

I evaluated GPT-4’s ability to simplify peer reviewer feedback in the academic field of finance. This domain was chosen not only for its highly technical language but also because its concepts, such as causality, risk modeling, or endogeneity, are closely tied to specific analytical methods. In such contexts, even small shifts in wording can distort connotations or undermine a reviewer’s intent. With a background in finance myself, I saw this as an ideal test case for examining whether AI can simplify with both clarity and conceptual integrity. I’d encourage researchers in other subject areas to run similar experiments within their own fields, where the stakes and semantics may differ, but the accessibility challenges remain just as urgent. To ensure complete transparency and reproducibility, I have shared transcripts of my conversations with GPT in my paper.

By prompting the model to rewrite feedback — once using a general plain-language prompt and once referencing cognitive accessibility needs — I examined how well it could preserve core meaning while improving readability. I selected ten technically dense reviewer comments and ran each prompt scenario twice, allowing me to assess not only the quality of simplification but also the model’s consistency across outputs.

I found that the model often missed the mark, not due to lack of clarity but due to loss of meaning. GPT-generated simplifications frequently altered the interpretation of technical language, glossed over methodological nuance, and introduced semantic inconsistencies across outputs. This raises critical concerns about the reliability of AI tools in accessibility contexts, especially when the goal is not just comprehension but equitable participation in scholarly discourse. What I found should concern anyone who disseminates content with both accuracy and inclusion in mind.

Take, for example, a reviewer’s comment about using a difference-in-differences approach to explain the timing and magnitude of market reactions to central bank announcements. The original text alluded to causal inference, a core concept in econometrics. GPT-4’s simplification turned this into a line about “how fast and how strongly the market reacts” rather than why markets react that way. That rephrasing softened the point from a methodological recommendation to a vague observation, effectively stripping away the purpose of using the technique in the first place.

In another case, the model described “endogeneity” as “hidden effects.” While this sounds more accessible on the surface, it dangerously dilutes the technical specificity of the term. Endogeneity is not just about things we can’t see — it refers to situations where explanatory variables are statistically compromised, often requiring advanced modeling strategies to correct. A more accurate definition would be “When something in your model affects both the cause and the outcome, making it hard to tell what’s really driving the results.” Oversimplifying it can lead a reader to overlook critical flaws in their research design. For instance, imagine a paper studying the effect of education on income. If the author interprets endogeneity as just “hidden effects,” they may focus only on missing data or measurement errors — things that are literally “hidden” — and might ignore the fact that factors like innate ability or family background influence both education and income, biasing the results.

The model also faltered with conceptual language. It simplified “bounded rationality” as “limited thinking ability” — a phrase that not only misrepresents the concept, but risks sounding offensive. Bounded rationality refers to decision-making under constraints like incomplete information or time pressure, not an inherent cognitive deficit. In content design for accessibility, we’re taught to center dignity and precision; GPT-4 failed on both counts here.

Even visual accessibility suffered in some outputs. For example, a reviewer’s concern about the illegibility of figure legends in a regression plot was simplified into a comment about “labels being hard to read.” While factually correct, this rewording dropped the context of why the legends mattered — namely, their role in interpreting group-wise effects in income data. For users depending on simplified content, this loss of purpose is no small matter.

Across multiple runs, GPT-4’s inconsistencies added to the problem. In one version, it retained key terms like “sensitivity analysis” but failed to explain them. In another, it replaced them with vague metaphors like “double-checking your methods.” The fluctuation wasn’t just stylistic — it was semantic. Sometimes the message changed entirely depending on how the model interpreted the same comment on a different day.

These results point to a core tension in AI-based content simplification: Efforts to clarify can unintentionally blur the underlying logic of a concept. This tradeoff is not acceptable in accessibility-focused design. When we rewrite for inclusion, we must preserve purpose, not just polish form. Especially in domains like finance, science, or law, the price of a misunderstanding isn’t confusion or flawed wording — it’s misdirection or flawed modeling of meaning.

Large language models like GPT-4 don’t “understand” the intent or function of a comment the way a human reviewer might; they predict likely phrasings based on statistical patterns in language. As a result, they often miss the why behind a piece of feedback, especially in content where accuracy and reasoning are the message itself. This makes them ill-suited, in their current form, to adapt content that demands both clarity and conceptual rigor. In accessibility contexts, where simplification must be done with care and respect for user autonomy, these failures are systemic rather than incidental. And they point to a need for more deliberate design in how we use AI for inclusive communication.

So, what can publishers, software developers, and AI practitioners take from this? First, prompt design is not a substitute for domain knowledge. Even accessibility-aware prompts failed to consistently guide GPT-4 toward better simplification. Second, verification matters. Each AI-generated simplification should be reviewed for technical fidelity and cognitive respect. And third, we need better alignment between LLM training and accessibility goals—not just in language style, but in conceptual accountability.

Designing for inclusion means more than making things easier to read. It means ensuring people receive the intended message rather than just a readable one. Since we have seen that GPT-4, in its current form, isn’t there yet, here are a few ways we might begin to develop tools that respect both meaning and accessibility.

One path forward is to develop AI tools fine-tuned to domain-specific academic content — such as models trained on preprints, peer reviews, and methodological appendices — so that simplification retains conceptual fidelity, not just surface clarity. Early-stage use could also include feedback loops: rather than expecting ongoing human verification, tools can learn from initial rounds of correction by editors or accessibility experts, improving their semantic accuracy over time.

Additionally, interface-level innovations — such as allowing reviewers or editors to label the purpose of a comment (for example, whether it relates to methods, structure, or language) — can help AI understand why a piece of feedback was written. This added context can guide the model to simplify more intelligently, preserving the intent behind the comment rather than just rewriting the words.

If we want accessibility to mean equity — not erasure — then these are the initiatives that every content strategist, accessibility advocate, and AI developer should take seriously.