Online Resources on AI and Cancer Care Are Scarce, Poor Quality, and Omit Critical Safety Risks, Study Finds

Researchers from the Abramson Cancer CenterView company profile of the University of PennsylvaniaView company profile and Penn's Perelman School of Medicine have found that publicly available online information about artificial intelligence and its role in cancer care is limited, difficult to read, and frequently omits critical safety risks. The findings were presented at the 2026 American Society of Clinical OncologyView company profile (ASCO) Annual Meeting (Abstract 9000).

The study, led by Internal Medicine Resident Pearl Subramanian, MD, and senior author Henry Litt, MD, a Hematology-Oncology Fellow, screened the first 320 webpages and videos identified through Google and YouTube searches using common cancer- and AI-related keywords. After removing content not relevant to AI and cancer care or not intended for lay audiences, only 52 webpages (31% of Google search results) and 29 videos (19% of YouTube search results) were included in the final analysis.

"In the clinic, we hear from patients all the time, asking us about something an AI tool told them, so we know patients are using this emerging technology," said Litt. "Clinicians are used to educating patients about the risks of treatment, but not about the risks of misinformation that might come with using AI tools in the context of their cancer care."

Quality and Readability Fall Short of Standards

Using the DISCERN instrument—a validated tool that judges the quality of written consumer health information on a 5-point scale—the researchers classified resources scoring 4 or higher as high quality. Only 17 webpages (33%) and seven videos (23%) met this threshold. Mean DISCERN quality scores were 2.5 for webpages and 2.25 for videos, both well below the recommended score of 4 out of 5.

While the American Medical Association (AMA) and National Institutes of Health (NIH) recommend a 6th–8th grade reading level for consumer health information, all three validated readability indices used in the study determined that most webpages were written at a college reading level.

"This is a call to action to develop more accessible resources for all patients with cancer—particularly for those who may be less informed AI consumers but are using it to guide their own care without recognizing the potential risks," Litt said.

Critical Safety Gaps Identified

The research team evaluated content for discussion of four key AI safety concepts: risk for misinformation or hallucination, the need for clinician oversight, potential for bias and inequity, and transparency that AI had been used to generate conclusions. While a majority of webpages and videos mentioned clinician oversight and addressed transparency, fewer than 20% referenced misinformation or hallucination. Only 15% of webpages mentioned the risk of AI hallucinations specifically. Approximately half discussed the potential for bias or inequity.

"The lack of details about misinformation shocked me," Subramanian said. "It is scary how much people believe everything they read, partly because of the media momentum surrounding these tools. In some cases, patients ask to change a treatment protocol that has been developed and studied on thousands of patients, simply because one article said something with no caveat that could it be wrong."

The researchers noted that many resources discussing AI safety did so from the perspective of how clinicians and hospitals incorporate AI tools, rather than addressing risks patients face as direct users of the technology. Subramanian offered a concrete example: if a patient asks a chatbot whether a treatment side effect is normal, the AI may provide an accurate, well-researched answer—or it may hallucinate information that could lead a patient to ignore a potentially serious side effect.

A Call for National Standards

The authors urged health systems, cancer centers, and national oncology organizations—including ASCO, the National Comprehensive Cancer Network, and the NCI—to partner with patient advocacy groups to develop high-quality, plain-language resources that integrate patient education into AI implementation strategies.

"Given that just one in four items from our search were deemed relevant to patients, and among those, only one in three was high-quality, patients may struggle to find useful, plain-language information," said co-author Ronac Mamtani, MD, who holds the David J. Vaughn MD Professorship in GU Oncology and served as faculty advisor on the study. "As AI is further integrated into oncology, patient education should be prioritized as a key part of AI implementation strategies."

Litt added that the clinical encounter itself may need to evolve: "Most oncologists are probably waiting for patients to bring up the fact that they found some piece of information on the internet. As these tools become more widely used, we may need to be more proactive and identify a standard educational intervention that we can give to all patients—or at least a screening question that we ask—to see if patients are using AI."

Study Limitations and Future Directions

The researchers acknowledged several limitations, including the single time point analysis conducted in August 2025 despite a rapidly evolving AI landscape, the fact that the DISCERN instrument—launched in 1999—may not sufficiently capture AI-specific quality concerns, and the exclusion of content from platforms such as Instagram, TikTok, and Reddit. The use of English-only resources also limits generalizability.

Future research could examine resources published in languages other than English and quantify the downstream effects of patients' AI use on clinical decision-making. "We definitely want to learn more about how this information is influencing patient behavior," Subramanian said. "It will be important to learn more about how these resources are influencing patient decisions."