San Francisco – OpenAI's newly released GPT-5 model is exhibiting "almost no hallucinations" and high factual accuracy in biomedical sciences and healthcare, according to early testing by Derya Unutmaz, MD, a prominent human immunologist and biomedical scientist. Dr. Unutmaz, who has had early access to the model, announced his findings on social media, stating, > "I’ve tested GPT-5 with hundreds of prompts, mostly in biomedical sciences & healthcare, over the past week & compared it to o3. It is highly factual with almost no hallucinations, & even when I try to push it to hallucinate or ask it to imagine, it remains relatively grounded." This assessment highlights a significant leap in AI reliability crucial for sensitive applications.
GPT-5, officially launched in August 2025, represents OpenAI's latest advancement, aiming to unify advanced reasoning and multimodal capabilities into a single system. OpenAI CEO Sam Altman has described it as a comprehensive system integrating previous technologies, including the "o3" model, which Dr. Unutmaz previously likened to a "seasoned expert." The model is designed to streamline complex workflows and significantly reduce erroneous outputs compared to its predecessors.
Dr. Unutmaz is a Professor at The Jackson Laboratory, where his research focuses on human immunology, aging, and cancer immunotherapy, with a strong emphasis on the application of artificial intelligence. His background includes extensive experience evaluating AI models, having received a ChatGPT Pro grant and previously offering detailed assessments of OpenAI's "o1" and "o3" models. His consistent engagement with and evaluation of cutting-edge AI tools lend significant weight to his observations.
The reported near-zero hallucination rate is particularly critical for the healthcare and biomedical sectors, where inaccurate AI outputs can lead to severe consequences such as misdiagnoses, inappropriate treatments, and compromised patient safety. Previous studies have highlighted the risks of AI hallucination, with models sometimes generating plausible but factually incorrect information. GPT-5's enhanced factual grounding addresses a major barrier to broader AI adoption in clinical and research settings.
This development could accelerate the integration of AI into medical research, drug discovery, and clinical decision support systems, fostering greater trust among healthcare professionals. The ability of GPT-5 to remain grounded even when prompted to imagine or extrapolate suggests a robust internal consistency vital for scientific inquiry and precision medicine. It marks a pivotal step towards more reliable and impactful AI applications in fields where accuracy is paramount.