Clin Orthop Relat Res. 2024 Apr 01. 482(4): 578-588
BACKGROUND: The lay public is increasingly using ChatGPT (a large language model) as a source of medical information. Traditional search engines such as Google provide several distinct responses to each search query and indicate the source for each response, but ChatGPT provides responses in paragraph form in prose without providing the sources used, which makes it difficult or impossible to ascertain whether those sources are reliable. One practical method to infer the sources used by ChatGPT is text network analysis. By understanding how ChatGPT uses source information in relation to traditional search engines, physicians and physician organizations can better counsel patients on the use of this new tool.
QUESTIONS/PURPOSES: (1) In terms of key content words, how similar are ChatGPT and Google Search responses for queries related to topics in orthopaedic surgery? (2) Does the source distribution (academic, governmental, commercial, or form of a scientific manuscript) differ for Google Search responses based on the topic's level of medical consensus, and how is this reflected in the text similarity between ChatGPT and Google Search responses? (3) Do these results vary between different versions of ChatGPT?
METHODS: We evaluated three search queries relating to orthopaedic conditions: "What is the cause of carpal tunnel syndrome?," "What is the cause of tennis elbow?," and "Platelet-rich plasma for thumb arthritis?" These were selected because of their relatively high, medium, and low consensus in the medical evidence, respectively. Each question was posed to ChatGPT version 3.5 and version 4.0 20 times for a total of 120 responses. Text network analysis using term frequency-inverse document frequency (TF-IDF) was used to compare text similarity between responses from ChatGPT and Google Search. In the field of information retrieval, TF-IDF is a weighted statistical measure of the importance of a key word to a document in a collection of documents. Higher TF-IDF scores indicate greater similarity between two sources. TF-IDF scores are most often used to compare and rank the text similarity of documents. Using this type of text network analysis, text similarity between ChatGPT and Google Search can be determined by calculating and summing the TF-IDF for all keywords in a ChatGPT response and comparing it with each Google search result to assess their text similarity to each other. In this way, text similarity can be used to infer relative content similarity. To answer our first question, we characterized the text similarity between ChatGPT and Google Search responses by finding the TF-IDF scores of the ChatGPT response and each of the 20 Google Search results for each question. Using these scores, we could compare the similarity of each ChatGPT response to the Google Search results. To provide a reference point for interpreting TF-IDF values, we generated randomized text samples with the same term distribution as the Google Search results. By comparing ChatGPT TF-IDF to the random text sample, we could assess whether TF-IDF values were statistically significant from TF-IDF values obtained by random chance, and it allowed us to test whether text similarity was an appropriate quantitative statistical measure of relative content similarity. To answer our second question, we classified the Google Search results to better understand sourcing. Google Search provides 20 or more distinct sources of information, but ChatGPT gives only a single prose paragraph in response to each query. So, to answer this question, we used TF-IDF to ascertain whether the ChatGPT response was principally driven by one of four source categories: academic, government, commercial, or material that took the form of a scientific manuscript but was not peer-reviewed or indexed on a government site (such as PubMed). We then compared the TF-IDF similarity between ChatGPT responses and the source category. To answer our third research question, we repeated both analyses and compared the results when using ChatGPT 3.5 versus ChatGPT 4.0.
RESULTS: The ChatGPT response was dominated by the top Google Search result. For example, for carpal tunnel syndrome, the top result was an academic website with a mean TF-IDF of 7.2. A similar result was observed for the other search topics. To provide a reference point for interpreting TF-IDF values, a randomly generated sample of text compared with Google Search would have a mean TF-IDF of 2.7 ± 1.9, controlling for text length and keyword distribution. The observed TF-IDF distribution was higher for ChatGPT responses than for random text samples, supporting the claim that keyword text similarity is a measure of relative content similarity. When comparing source distribution, the ChatGPT response was most similar to the most common source category from Google Search. For the subject where there was strong consensus (carpal tunnel syndrome), the ChatGPT response was most similar to high-quality academic sources rather than lower-quality commercial sources (TF-IDF 8.6 versus 2.2). For topics with low consensus, the ChatGPT response paralleled lower-quality commercial websites compared with higher-quality academic websites (TF-IDF 14.6 versus 0.2). ChatGPT 4.0 had higher text similarity to Google Search results than ChatGPT 3.5 (mean increase in TF-IDF similarity of 0.80 to 0.91; p < 0.001). The ChatGPT 4.0 response was still dominated by the top Google Search result and reflected the most common search category for all search topics.
CONCLUSION: ChatGPT responses are similar to individual Google Search results for queries related to orthopaedic surgery, but the distribution of source information can vary substantially based on the relative level of consensus on a topic. For example, for carpal tunnel syndrome, where there is widely accepted medical consensus, ChatGPT responses had higher similarity to academic sources and therefore used those sources more. When fewer academic or government sources are available, especially in our search related to platelet-rich plasma, ChatGPT appears to have relied more heavily on a small number of nonacademic sources. These findings persisted even as ChatGPT was updated from version 3.5 to version 4.0.
CLINICAL RELEVANCE: Physicians should be aware that ChatGPT and Google likely use the same sources for a specific question. The main difference is that ChatGPT can draw upon multiple sources to create one aggregate response, while Google maintains its distinctness by providing multiple results. For topics with a low consensus and therefore a low number of quality sources, there is a much higher chance that ChatGPT will use less-reliable sources, in which case physicians should take the time to educate patients on the topic or provide resources that give more reliable information. Physician organizations should make it clear when the evidence is limited so that ChatGPT can reflect the lack of quality information or evidence.