Generations of Generating a Cytology Report

Cytology reports have evolved over the last 40 years and will continue to do so as technology advances (Figures 1 and 2). While the use of generative artificial intelligence (AI), such as ChatGPT, can potentially aid in the generation of reports and training of residents, appropriate use of AI also requires training and understanding of the advantages and pitfalls of the technology. With the appropriate guidance from subject matter, digital pathology, and AI experts, we need to embrace the rapidly evolving technology that has the potential to contribute to training future pathologists and improve patient care, despite our concerns about how this might change our careers. The cytology reports in Figures 3 and 4 were generated in approximately 30 s using OpenAI's freely accessible ChatGPT web application (Figures 3 and 4). ChatGPT is an AI chatbot that uses natural language processing to engage users in human-like conversations. OpenAI's latest models allow for text, image, and audio input. The large language models (LLMs) that underlie ChatGPT and other commercially available, general-purpose AI chatbots (e.g., Google's Gemini, Anthropic's Claude) are trained on diverse datasets, enabling them to generalize across a broad spectrum of tasks. The performance of these general-purpose tools on pathology tasks has been explored in the literature and is promising [2]. However, tools tailored to pathology, such as Modella AI's PathChat, a domain-specific generative AI copilot for human pathology, have been shown to outperform general-purpose AI chatbots in pathology tasks [3]. Such domain-specific generative AI shows great promise to augment the evaluation of pathology images, the suggestion of differential diagnoses and ancillary testing, and the generation of pathology reports [4]. The cytology reports in Figures 3 and 4 highlight the importance of crafting intentional and specific inputs (so-called prompt engineering) and the need for subject-matter expert oversight when utilizing generative AI. Prompt strategies include writing clear instructions that request an expert-level response and the desired format and detail of the response [5]. Some people recommend adopting a persona, such as including, “You are an ACVP board-certified clinical pathologist.” Providing specific references may minimize hallucinations, but acceptance of the output requires a subject-matter expert. The report in Figure 3 was generated without providing a reference. The cytologic description drifts into content that would more appropriately apply to histologic descriptions. Some terminology may be more applicable to human medicine. Some tests, such as ALP staining, are not suggested. The prompt used to generate the report in Figure 4 includes a specific veterinary clinical pathology resource for guidance, provides report formatting instructions, and uses a newer, more capable model (GPT-5 thinking) than the model used to generate Figure 3 (GPT-4o) [1]. Expert oversight is needed to ensure that all the described cytologic features are present on the actual sample slide and the level of diagnostic confidence is appropriate for the specific case. It is important to note that repeated use of identical prompts will generate somewhat different reports each time. For a report being generated for a specific case, including patient history, cytologic findings, and representative photomicrographs would likely further enhance utility. One should exercise caution when entering case details or uploading files into generative AI tools, doing so only in accordance with institutional data-governance policies, client/patient confidentiality requirements, and copyright/licensing restrictions on source materials; this landscape is evolving, and many institutions are adopting enterprise deployments with stronger protections (e.g., limited data retention and opt-out settings that prevent prompts/outputs from being used to train vendor models). Care was taken in the generation of Figure 4 to use a reference with amenable copyright terms (and a reference from a website ChatGPT frequently cites without specific guidance when answering veterinary clinical pathology inquiries based on the authors' experience) and an OpenAI account with strict data-control settings. Artificial intelligence will be an integral part of our future as pathologists. We need to learn how to use it effectively and efficiently, and find ways to incorporate it into our training programs so that the next generation of pathologists is aware of the challenges and opportunities. As Annie reminds us in “Tomorrow” (as well as the 8 a.m. meeting in our work calendars), tomorrow is both exciting and rapidly approaching [6]. The authors declare no conflicts of interest.