Kihyuk Lee
6 min
This paper by Kihyuk Lee investigates whether large language models (LLMs) like Gemini 2.5 Flash can reliably generate exercise prescriptions—the same way a doctor might recommend workouts tailored to a patient's health. Exercise prescriptions aren't just generic advice; they must precisely specify Frequency (how often), Intensity (how hard), Time (duration), and Type (what exercises), known as the FITT principle. In clinical settings, especially for patients with comorbidities like diabetes, inconsistency could lead to unsafe or ineffective plans. The study tests intra-model consistency: does the same LLM give similar outputs when fed the exact same prompt repeatedly?
They generated 120 prescriptions across six scenarios (clinical patients vs. healthy adults) with 20 repeats each. Key insight: LLMs are semantically stable (text reads similarly) but wobble on precise numbers, highlighting why AI can't yet replace human experts without safeguards.
Semantic consistency was measured using SBERT (a sentence embedding model) to compute cosine similarity between repeated outputs. Scores were high: mean 0.879–0.939 across scenarios. Imagine re-rolling a D&D character sheet 20 times—the backstory and class stay consistent, even if exact stats vary slightly. Clinically constrained scenarios (e.g., diabetes patients) showed higher stability than open-ended healthy adult cases, suggesting LLMs anchor better to medical context. This builds intuition: LLMs compress knowledge into latent spaces where similar prompts activate overlapping neural pathways, yielding textually stable outputs.
Structure was evaluated using 'AI-as-a-judge': another LLM scored outputs against FITT rubrics. Frequency (e.g., '3x/week') was rock-solid across repeats. But intensity (e.g., '% max heart rate' or 'light/moderate') varied, especially in resistance training where 10–25% of outputs had unclassifiable descriptions like vague 'challenging weights'. Time and Type were middling. Why? LLMs excel at patterns but struggle with quantitative precision without explicit constraints—think of it as a chef who nails recipes qualitatively but eyeballing measurements inconsistently.
Safety warnings (e.g., 'stop if dizzy') appeared in 100% of outputs—great! But the volume varied wildly: clinical scenarios averaged far more safety sentences than healthy ones (Kruskal-Wallis H=86.18, p<0.001). This reflects LLMs' training bias toward caution in medical contexts, but variability means the same patient might get 2 vs. 10 warnings across runs, eroding trust.
High semantic stability suggests LLMs could draft consistent narratives, but quantitative gaps (intensity variability) demand structured prompts, expert review, and hybrid systems for clinics. This echoes broader LLM limits: great at language, finicky on numbers. For digital health, it advocates 'prompt engineering + validation' pipelines. Why care? Scalable exercise Rx could democratize fitness for underserved areas, but only if reproducible. Future: constrain outputs to FITT templates or ensemble multiple generations.
Background: Large language models (LLMs) have been explored as tools for generating personalized exercise prescriptions, yet the consistency of outputs under identical conditions remains insufficiently examined. Objective: This study evaluated the intra-model consistency of LLM-generated exercise prescriptions using a repeated generation design. Methods: Six clinical scenarios were used to generate exercise prescriptions using Gemini 2.5 Flash (20 outputs per scenario; total n = 120). Consistency was assessed across three dimensions: (1) semantic consistency using SBERT-based cosine similarity, (2) structural consistency based on the FITT principle using an AI-as-a-judge approach, and (3) safety expression consistency, including inclusion rates and sentence-level quantification. Results: Semantic similarity was high across scenarios (mean cosine similarity: 0.879-0.939), with greater consistency in clinically constrained cases. Frequency showed consistent patterns, whereas variability was observed in quantitative components, particularly exercise intensity. Unclassifiable intensity expressions were observed in 10-25% of resistance training outputs. Safety-related expressions were included in 100% of outputs; however, safety sentence counts varied significantly across scenarios (H=86.18, p less than 0.001), with clinical cases generating more safety expressions than healthy adult cases. Conclusions: LLM-generated exercise prescriptions demonstrated high semantic consistency but showed variability in key quantitative components. Reliability depends substantially on prompt structure, and additional structural constraints and expert validation are needed before clinical deployment.
Alex: So even with the same prompt, it might say "light weights" one time and skip the math the next. That could push someone from safe to risky without realizing.
Sam: Precisely. For patients with conditions like diabetes, that dose variation matters—a low walk versus moderate lift could mess with blood sugar or injury risk. The paper points to this as why quantities drift: clinical rules narrow the big picture, but numbers stay loose without stricter guidance.
Alex: Those differences across patient scenarios—like diabetes versus healthy strength goals—show up clearly. How did they confirm which ones varied most?
Sam: They used a statistical tool suited for small groups and uneven data: first, a test that checks if scores from several scenarios differ overall, without assuming the numbers form a perfect bell shape. If differences pop up, they follow with pairwise checks adjusted to cut false alarms. The paper suggests clinical cases, such as knee issues or post-cancer recovery, had tighter word matches than healthy goals like building muscle.
Alex: For the exercise structure itself, did patterns hold steady within those groups?
Sam: In clinical scenarios, frequency stuck to aerobic work three or four days a week plus resistance twice, in about half the plans. Aerobic intensity stayed low nearly always, like gentle walks fitting cautious patients, while resistance was mostly low but some lacked enough detail for sorting. Healthy cases shifted to moderate or high intensities, with longer sessions, reflecting fewer limits.
Alex: And those unclassifiable resistance plans—mostly vague on weights? Did safety mentions follow the same split?
Sam: Every single plan included all four safety types—warnings, precautions, monitoring symptoms, and no-gos—but clinical ones packed in far more sentences overall. Stats confirmed clinical outputs emphasized safety more, with significant gaps between complex patients and simple fitness aims. This points to prompts steering toward caution in riskier scenarios, yet the sheer volume varies.
Alex: So if words line up but the actual workout doses don't, that could mean the same patient gets plans that look safe but deliver different physical stress.
Sam: Exactly. Plans might shift aerobic effort from something easy to a tougher level, which feels similar in description but changes the body's response enough to risk injury or poor results. High word matches don't guarantee that kind of dose match.
Alex: The paper worries about rolling this out to clinics yet?
Sam: It does. They point out resistance effort wobbled most, with some plans too vague to categorize even after asking for weight percentages. Frequency shifted in healthy cases too, no single pattern dominating. Plus, limitations like testing just one AI model mean these patterns might not hold for others. Safety checks only counted mentions, not if they're spot-on for the patient. The judge AI wasn't compared to human experts.
Alex: So semantic steadiness is a start, but numbers need controls.
Sam: The conclusion lands there: words converge under constraints, but quantitative bits like intensity drift without extras. For clinics, that means structured steps beyond raw AI output. A solid call for cautious steps forward, with prompt tweaks and expert review.
Alex: Makes sense—this balances promise with real caution. Thanks, Sam, for breaking it down so clearly.
Sam: My pleasure, Alex. Thanks for listening to ResearchPod.