Background

Over the last decade, there has been a growing interest in the use of artificial intelligence (AI) to streamline health care delivery and this technology is now being applied to areas that were previously thought to be only the jurisdiction of human experts. One AI technology that has caught the attention of the medical community and lay public alike is OpenAI’s large language model (LLM)-based chatbot ChatGPT (generative pretrained transformer). ChatGPT is a natural language processing technology that can generate conversational, human-like text using a deep learning machine learning algorithm that was trained on 175 billon tokens and has been heralded as the best AI chatbot ever released for public consumption (Figure 1).^1-3 ChatGPT has garnered significant attention for its ability to provide articulate responses and carry out tasks with a wide range of sophistication.

Figure 1. Sample prompt to create an appeals letter in ChatGPT, with sample output.

ChatGPT and other LLM-based chat AI represent a promising time-saving tool for physicians in an era with increasing administrative burdens.^4,5 A recent study using AUA Census data found that 37% of urologists report signs of burnout, with the highest workplace dissatisfiers being electronic health record tasks and limited personal/family time.⁶ Given that the preponderance of physician writing follows predictable and templated formats, natural language processing AI such as ChatGPT can make an immediate impact on physician workflow.

We describe the strengths and weaknesses of ChatGPT and encourage physicians to utilize this technology to free up more time for face-to-face interaction with patients and alleviate burnout by aiding in routine, low-stakes compositions.

ChatGPT in Practice

Physicians can sign up to use ChatGPT through OpenAI’s website (chat.openai.com) and submit requests through its user-friendly interface.

There are numerous applications for this technology along the care continuum (Figure 2). Some specific examples include drafting emails, letters to insurance companies, patient-facing discharge or medication instructions, brainstorming research ideas, event or appointment scheduling, and designing queries for research articles.⁷ These tasks have predictable outputs that can be quickly reviewed and fine-tuned, saving considerable time in the initial writing process. Additionally, ChatGPT can recall inputs from within the same session and thus one can ask the chatbot to rephrase, reconfigure, or expand upon previous inputs. ChatGPT has demonstrated that it is able to cope with ambiguity from requests. However, to get the most out of its time-saving capabilities, it is advantageous for physicians to become acquainted with the terminology and structure of requests to get a desired output. There are several repositories (eg, EmergentMind) in which users have published prompts (some serious, some comical) and resultant outputs to guide others.⁸ Furthermore, there are tools that utilize ChatGPT’s API (eg, Doximity’s DocsGPT) to provide doctors with premade prompts for clinical care purposes.⁹

Figure 2. The future of generative artificial intelligence in health care. EHR indicates electronic health record.

Limitations of ChatGPT

Physicians should be cognizant of ChatGPT’s numerous limitations. Since ChatGPT is trained using text obtained from the Internet, its outputs are prone to bias and inconsistencies.¹⁰ ChatGPT does not exhibit indecisiveness and may cite publications that do not exist or make factually inaccurate statements with certainty. As evidenced by events that have occurred following its beta integration with Bing, ChatGPT is also prone to hallucination—generation of meaningless (and sometimes unsettling) answers based on bugs in the algorithm.¹¹ At best, it can function as a starting point for low-value written content from which one can revise.

ChatGPT currently does not scrape the Internet for new data, so any output that it provides will be based on information that it learned prior to 2021. ChatGPT and other LLMs are not HIPAA (Health Insurance Portability and Accountability Act of 1996) or GDPR (General Data Protection Regulation) compliant. Physicians should not enter protected health information.

ChatGPT is currently free to use during times of low traffic to their website with full capabilities. For some users, the $20 monthly subscription, which provides faster responses, availability during high traffic times, and earlier access to new features as they become available, may be a valuable investment.

Although ChatGPT is being used to streamline researcher workflows including the brainstorming of research ideas, literature review, and peer review, there has been increasing use of this technology in the drafting of manuscripts. JAMA Network, Science, Nature, the World Association of Medical Editors, and the Committee on Publication Ethics have issued statements prohibiting the listing of AI as an author because “AI cannot take responsibility for submitted work … or assert presence or absence of conflicts of interest.”^12-14 Researchers who use AI tools in manuscript writing must disclose this in the methods or acknowledgments of the paper. Several groups have designed software that can detect AI-generated content, and in the future, there will likely be a watermark embedded in all AI-generated content to ensure transparency of its origin.

Conclusions

When taken together, ChatGPT can be an essential time-saving companion for physicians by streamlining low-complexity tasks. Although there are many limitations to ChatGPT and other LLMs, this technology is rapidly improving and will become increasingly utilized in the health care setting. In an era in which electronic health record tasks and administrative burdens are a lead driver of burnout, now is as good a time as any to embrace AI’s time-saving potential.

Funding: None.

Conflict of Interest: David Canes is the Founder of WellPrept, Inc. Anobel Odisho has received research support from the Microsoft Turing Academic Program. Andrew Gabrielson has no relevant relationships to disclose.