Large language models (LLMs) are increasingly being used and studied in healthcare for support in clinician decision and patient education. In the setting of low back pain, these tools are used to answer questions about diagnosis, treatments, and when to seek medical attention. While board-certified pain physicians have the expertise to manage these conditions, LLMs may serve as a supplementary resource when clinicians need rapid access to guideline-based recommendations, clarification of uncommon presentations, or clear explanations for patients. Hence, the accuracy and completeness of this information are crucial given the complexity of decision making in spine disorders.

Prior work has shown that LLMs has medical knowledge and can work to a reasonable level in respect to structured medical reasoning, provided it is prompted adequately [1, 2]. As access to LLMs is becoming more popular and easily available, it becomes important to understand the effect of prompts and responses [3]. Multiple studies have demonstrated that LLMs’ performance in medicine relies on how information is elicited from clinicians. When prompts are written for clinicians the model outputs tend to be more detailed and accurate. In contrast, simplified prompts often yield shorter responses with less accurate information. These findings suggest that performance differences may reflect how information is elicited rather than limitations in the LLMs [2, 4].

Within the field of pain, incomplete clinical reasoning can lead to procedural consequences. Inadequate discussion of diagnostic uncertainty, red flag symptoms, anticoagulation status, or imaging correlation may compromise patient selection and procedural safety. Established spine and pain guidelines stress the importance of a thorough clinical assessment before proceeding with intervention [5, 6]. Variability in the detail and clarity of generated responses therefore has meaningful implications for clinician decision-making.

At the same time, interest is rising in using LLMs to provide explanations for patients. Simplifying medical information is often necessary to address health literacy and support patient understanding [7]. The National Institutes of Health recommends that patient education materials be written at approximately an eighth-grade reading level to improve accessibility [8]. However, prior work on text simplification has shown that efforts to reduce language complexity can result in the loss of important information [9]. Across pain practices, this can be problematic if simplified explanations do not clearly convey diagnostic uncertainty, warning signs, or guidance for appropriate management.

Recent studies have examined how readability constraints affect LLMs’ generated medical content. These evaluations suggest that enforcing low reading levels can reduce clinical completeness, largely due to the omission of important contextual details [4, 10]. In most of these studies, it has only been considered as a response generated based on simplification. It does not examine whether omitted clinical information can be recovered through reframing or follow-up prompts. As a result, it remains unclear whether simplification leads to irreversible information loss or instead reflects a prompt driven suppression of detail.

The current evaluation metrics for medical AI systems also seem to focus primarily on metrics related to accuracy [11]. Although accuracy is an important factor, it may not fully capture clinically important details that would be expressed in open ended questions. Additionally, most studies do not use multiple prompt conditions within the same clinical question.

The distinction between actual loss of information and mirroring based on the prompt has important implications for both provider and patient facing applications of LLMs. If simplified prompts result in irreversible loss of clinical content this will raise concerns about the safety of patient-oriented AI tools. In contrast, if omitted information can be recovered by clinician up prompting this would suggest that LLMs retain clinically relevant knowledge despite constraints.

In this study, we used a controlled comparative design to evaluate the effects of prompt simplification and upscaling on LLMs generated responses to low back pain questions relevant to pain practice. By comparing professional, fourth grade reading level, and upscaled prompts, we are trying to determine whether the drop in accuracy and clinical detail seen with simplification reflects true information loss. We also assessed clinical completeness and linguistic measurements to provide a more comprehensive evaluation of model performance across prompt conditions.