The New England Journal of MedicineBenefits, Limits, and Risks of GPT-4 as an AI Chatbot for Medicine

Examining the capabilities, limitations, and potential risks of GPT-4 in medical applications, and its implications for clinical practice and future AI developments.

The article explores the various applications of GPT-4, an advanced AI chatbot developed by OpenAI, in the field of medicine. It discusses the system’s potential to assist in tasks such as medical note-taking, consultation, and education while addressing the inherent risks and limitations associated with its use. The piece underscores the necessity of vigilance and continuous improvement in AI technology to ensure its safe and effective integration into healthcare settings.

Key Points:

• AI Chatbot Technology:
  • GPT-4 combines a general-purpose AI system with a chat interface.
  • Users interact with the chatbot by entering prompts in natural language.
  • The system generates natural-language responses, creating a conversational experience.
• Prompt Engineering:
  • Effective use of GPT-4 requires careful development and testing of prompts.
  • The quality of responses depends significantly on the wording of the prompts.
  • Future AI systems may be less sensitive to prompt specifics.
• Hallucinations and Error Detection:
  • GPT-4 can produce incorrect or misleading responses, known as “hallucinations.”
  • These errors can be dangerous in medical contexts.
  • GPT-4 can also identify and correct its own mistakes in subsequent sessions.
• Medical Applications:
  • GPT-4 assists in various medical tasks, including documentation, data interoperability, diagnosis, research, and education.
  • It was not trained specifically for healthcare but shows competence in medical applications due to its general cognitive capabilities.
  • The AI system was tested on scenarios like medical note-taking, USMLE questions, and curbside consults.
• Medical Note-Taking:
  • GPT-4 can generate medical notes from transcripts of physician-patient encounters.
  • It can produce notes in standard formats, such as SOAP, and include billing codes.
  • The AI can be prompted to perform additional tasks, such as extracting prior authorization information and generating lab and prescription orders.
• Innate Medical Knowledge:
  • Despite training on openly available information, GPT-4 performs well on medical exams like the USMLE.
  • It can provide detailed explanations and reasoning for its answers.
• Consultation and Education:
  • GPT-4 can assist with medical consultations by providing useful responses to clinical questions.
  • It can summarize research articles, perform technical analyses, and engage in informed discussions.
• Risks and Limitations:
  • GPT-4’s general intelligence poses challenges in defining acceptable performance standards.
  • Continuous monitoring and improvement are necessary to mitigate risks.
  • Trust and verification of AI-generated outputs remain critical.

“Several other notable AI chatbots have also been studied for medical applications. Two of the most notable are LaMDA (Google) and GPT-3.5,8 the predecessor system to GPT-4. Interestingly, LaMDA, GPT-3.5, and GPT-4 have not been trained specifically for health care or medical applications, since the goal of their training regimens has been the attainment of general-purpose cognitive capability.”

More on Chatbots