Gaps in patient-facing information on artificial intelligence in cancer care: A cross-sectional analysis.

Abstract

9000 Background: Artificial intelligence (AI) is increasingly integrated into cancer care and is also widely accessed by patients seeking information about their diagnoses and treatments. While AI has the potential to improve care delivery and patient engagement, inadequate or misleading patient education may introduce safety risks. We assessed the availability, readability, and quality of publicly available online patient-facing information about AI in cancer care. Methods: We conducted a cross-sectional analysis of online patient-facing information related to AI in cancer care. Using common cancer- and AI-related keywords identified using Google Trends, we searched Google and YouTube on August 6, 2025. The first 170 webpages and 150 videos were screened for relevance and patient-facing intent; scientific or industry-facing content was excluded. Eligible webpages and videos were independently evaluated by two reviewers with discrepancies resolved by a third. Webpage readability was assessed using validated indices (Flesch–Kincaid [FK], Gunning Fog [GF], and SMOG). Content was evaluated for discussion of key AI safety concepts including clinician oversight, transparency, bias, and hallucination or misinformation risk. Quality of consumer health information was assessed for both webpages and videos using the DISCERN instrument, with scores ≥ 4 (of 5) indicating high quality. Descriptive statistics were used to summarize findings. Results: Of the 170 webpages screened, 52 (31%) met inclusion criteria. Most content focused on breast cancer (n=30, 58%) or was pan-tumor (n=22, 42%). Median readability corresponded to a college-level reading standard, with median grade levels of 12.8 (IQR: 11.6-14.1), 14.8 (IQR: 13.5-16.5), and 14.2 (IQR: 13.6-15.6) based on FK, GF, and SMOG indices, respectively. Most webpages discussed clinician oversight (n=41, 79%) and/or transparency (n=41, 79%), and over half addressed bias (n=29, 56%); however, few discussed hallucination or misinformation risk (n=8, 15%). Only 33% (n=17) of webpages met criteria for high-quality information. Of the 150 videos, 29 (19%) met inclusion criteria. Median view count was 127 (IQR: 23-1000). Few (n=11, 38%) were classified as high quality based on DISCERN scores. Conclusions: Publicly available, patient-facing information about AI in cancer care is limited, difficult to read, and often of low quality. Inadequate discussion of AI-related risks, particularly hallucinations or misinformation, may leave patients poorly informed as they encounter or independently use AI-based tools. These findings highlight the need for accessible, high-quality educational resources that clearly explain the clinical role of AI in oncology and provide guidance for safe patient engagement.