The research team selected ChatGPT, Gemini, DeepSeek, and Doubao, four publicly accessible chatbot interfaces chosen for their consumer-facing availability during the study period. On 20 June 2026, from Jiangsu, China, the investigators submitted 36 standardized patient-facing questions about osteoporotic vertebral compression fractures to each chatbot in independent single-turn sessions under matched conditions, generating 144 responses for the primary between-chatbot analysis. Seven days later, on 27 June 2026, they repeated the identical procedure with the same 36 questions, producing another 144 responses for a within-model stability analysis. The questions spanned the full clinical pathway of vertebral compression fracture care, covering symptoms and warning signs, diagnosis and imaging, conservative management, medication safety, rehabilitation, vertebral augmentation, fracture prevention, and special clinical scenarios. All prompts were entered in English using default settings, with no browsing tools, uploaded files, custom instructions, follow-up prompts, or readability-oriented instructions, and each question was submitted in a new independent chat session to minimize contextual carryover.

Five clinical raters, two spine surgeons, two specialists in osteoporosis or metabolic bone disease, and one rehabilitation physician, with 10 to 15 years of clinical experience, independently evaluated every response while blinded to chatbot identity and generation round. Safety was prespecified as the primary outcome and rated as binary, with a response classified as potentially unsafe if it could plausibly contribute to delayed urgent evaluation, inappropriate self-management, unsafe medication use, excessive immobilization, premature preference for invasive procedures, unsafe mobilization or exercise, failure to recognize contraindications, or misleading reassurance. Secondary outcomes included factual accuracy, empathic communication, treatment-related information quality using the DISCERN instrument, patient-information quality using the EQIP instrument, transparency using the JAMA benchmark criteria, and overall quality using the Global Quality Score. Readability was assessed automatically using six established indices, including the Flesch-Kincaid Grade Level, Gunning Fog Index, and Flesch Reading Ease Score. Before formal scoring, raters completed a calibration exercise using sample responses excluded from the final analysis, and the study was prepared in accordance with the CHART reporting statement for chatbot health advice studies.

The headline safety finding is that every chatbot produced potentially unsafe responses. ChatGPT generated 3 of 36 potentially unsafe responses, or 8.3 percent, DeepSeek produced 5 of 36, or 13.9 percent, Gemini produced 6 of 36, or 16.7 percent, and Doubao produced 7 of 36, or 19.4 percent. However, the overall paired difference in safety was not statistically significant, with Cochran’s Q equal to 3.621 on 3 degrees of freedom and P equal to 0.305. A qualitative review of the 21 responses classified as potentially unsafe identified recurrent concerns involving inadequate warning-sign triage, insufficient medication contraindication safeguards, overly broad indications for vertebral augmentation, prematurely prescriptive exercise advice, and recovery recommendations that did not adequately reflect clinical context. In practical terms, a chatbot response may appear broadly correct while still being potentially unsafe because it underemphasizes neurologic red flags, overstates the role of vertebral augmentation, or provides insufficiently qualified advice regarding medication, mobilization, or exercise.

Factual accuracy told a different story. Median accuracy scores were tightly clustered at 4.00 for all four chatbots, and although the overall difference reached statistical significance at P equal to 0.009, the effect size was small, with Kendall’s W equal to 0.108, indicating that the magnitude of the between-chatbot difference was modest and should not be interpreted as evidence of clinically meaningful superiority of any individual chatbot. Empathy, by contrast, varied substantially more, with P less than 0.001 and Kendall’s W equal to 0.817. ChatGPT and Gemini had median empathy scores of 4.00, whereas DeepSeek and Doubao had median scores of 3.00, suggesting that the models differed not only in informational content but also in communication style. This distinction matters because patients with vertebral compression fractures may experience acute pain, fear of neurologic injury, functional limitation, anxiety about future fractures, and uncertainty about activity restrictions, all of which call for supportive tone combined with clear safety boundaries and advice to seek clinician assessment when individualized care is required.

Information quality, transparency, and overall quality also differed significantly among chatbots, with all P values less than 0.001 and Kendall’s W values ranging from 0.151 to 0.334. ChatGPT had the highest observed median DISCERN score of 66.90 and the highest median EQIP score of 80.00, whereas Doubao had the lowest observed medians for DISCERN, EQIP, JAMA benchmark, and Global Quality Score. Transparency remained limited across all chatbots, with low or inconsistent JAMA benchmark performance likely reflecting the limited authorship, source attribution, disclosure, and currency information typically visible in unprompted chatbot responses. For osteoporotic vertebral compression fracture information, transparency is especially important because management decisions depend on fracture acuity, neurologic status, imaging findings, pain severity, comorbidities, medication history, and patient preferences, and future patient-facing systems may benefit from verifiable source links, dates, explicit uncertainty statements, and safeguards against unsupported medication, procedural, or rehabilitation recommendations.

Readability emerged as an important limitation across all four chatbots. Mean Flesch-Kincaid Grade Level values ranged from 12.12 for Gemini to 17.54 for Doubao, suggesting high textual complexity for patient-facing materials, and Doubao had the highest mean scores on the Automated Readability Index, Coleman-Liau Index, Flesch-Kincaid Grade Level, Gunning Fog Index, and Simple Measure of Gobbledygook, along with the lowest mean Flesch Reading Ease Score of 20.28, indicating the greatest estimated reading difficulty. ChatGPT, Gemini, and DeepSeek had lower grade-level estimates, although pairwise differences among these three chatbots varied by index. These findings are particularly relevant because osteoporotic vertebral compression fractures primarily affect older adults, a population in which some individuals may have limited health literacy, visual impairment, pain-related distress, or difficulty processing dense medical information, and readability formulas estimate textual complexity but do not directly measure comprehension, trust, recall, or actionability.

The study also examined within-model generation stability by comparing the two independent response-generation rounds conducted seven days apart. All 36 questions were successfully regenerated in the second round, yielding 144 matched chatbot-question response pairs for the stability analysis. Test-retest stability was generally good across models, with safety agreement ranging from 88.9 percent for Doubao to 97.2 percent for ChatGPT and Gwet’s AC1 values ranging from 0.838 to 0.966. Across the 12 non-safety outcomes, median intraclass correlation coefficient values were 0.904 for ChatGPT, 0.870 for Gemini, 0.820 for DeepSeek, and 0.781 for Doubao, and all 12 outcomes met the prespecified threshold of 0.75 for ChatGPT and Gemini, compared with 11 of 12 for DeepSeek and 10 of 12 for Doubao. Median absolute Cohen’s dz values were small across all models, ranging from 0.000 to 0.018, and no non-safety outcome showed significant systematic drift after Holm correction, suggesting that the observed within-model stability was generally good but did not eliminate variability across generations.

Inter-rater agreement was high across all rater-assessed outcomes, lending credibility to the evaluation. Safety ratings showed substantial agreement with a Fleiss kappa of 0.818, and the two-way random-effects, absolute-agreement, single-measure intraclass correlation coefficient estimates were 0.816 for factual accuracy, 0.820 for empathy, 0.820 for DISCERN, 0.820 for EQIP, 0.821 for JAMA benchmark, and 0.818 for Global Quality Score, all with P values less than 0.001. The study had several strengths, including a disease-specific and safety-sensitive question set, standardized single-turn evaluation conditions, five blinded clinical raters across multiple quality dimensions, and a paired design with supplementary materials that enhanced transparency and auditability. However, the investigators also acknowledged important limitations, noting that the primary between-chatbot comparison was anchored to a single initial access date and location, that backend model routing, safety filters, interface-level processing, and software updates may be undisclosed and may change over time, and that the exact computational configuration could not be independently verified or reproduced.

The bottom line for patients, caregivers, and clinicians is clear. Although all four chatbot interfaces generated responses with relatively high clinician-rated factual accuracy, each also produced some responses classified as potentially unsafe, with no statistically significant difference detected in the primary safety comparison, and differences were more pronounced in empathic communication, information quality, transparency, overall quality, and textual complexity than in factual accuracy. Given the occurrence of potentially unsafe responses across all evaluated chatbot interfaces and the limited readability for general audiences, the study concludes that large language model-generated information about osteoporotic vertebral compression fractures should not be used as a standalone patient education resource. If such tools are incorporated into patient education workflows, they should be considered only as supervised adjuncts requiring clinician review, safety safeguards, and further validation, and future patient-facing applications should incorporate clear escalation guidance, evidence transparency, plain-language communication, and appropriate clinician oversight. The findings are specific to the evaluated access location, language, single-turn conditions, and generation dates, and should not be assumed to represent stable performance across settings or future model updates, a caution that underscores how quickly this technology evolves and how carefully it must be evaluated before it can be trusted with the health questions of vulnerable older adults.

Subject of Research: Safety and quality of four large language model chatbots answering patient questions about osteoporotic vertebral compression fractures

Article Title: Performance of four large language model chatbots in responding to patient-facing questions about osteoporotic vertebral compression fractures: a cross-sectional paired evaluation

Article References: Mao, T., Bian, F., Liu, W., Zhou, X., & Zhang, G. (2026). Performance of four large language model chatbots in responding to patient-facing questions about osteoporotic vertebral compression fractures: a cross-sectional paired evaluation. Frontiers in Musculoskeletal Disorders, 4, Article 1935149. https://doi.org/10.3389/fmscd.2026.1935149

Image Credits: AI Generated

DOI: 10.3389/fmscd.2026.1935149

Keywords: large language models, chatbots, osteoporotic vertebral compression fractures, patient education, health information safety, ChatGPT, Gemini, DeepSeek, Doubao, clinical evaluation, readability, artificial intelligence