Back To Good Reads

Can ChatGPT take over our newsletter?

May 1, 2023 Janice Karin

ChatGPT logo

AI and machine learning are all over healthcare (and everywhere else) these days. ONC discusses its use in clinical decision support in their latest proposed rule. The FTC devoted large sections of its RFI on data surveillance and privacy to it. To some, it seems like a magic solution to a lack of time, resources, and expertise. To others, it's still mired in the traditional science fictional evil overlord taking over stereotypes. In truth, it's both and neither, a promising technology that isn't quite ready for prime time but that can be useful when used cautiously, judiciously, and within clear guard rails.

Not sure where you fall in the spectrum? Not sure you understand the pros and cons of using AI or its current landscape? Think we're living in the future or that we're fuddy duddies trying to hold it back? Let's take a look at the promise and limitations of AI by examining the hottest of hot AI topics - ChatGPT.

We ran a few very low-key, quick experiments using ChatGPT and found some promising but also some inaccurate results, some of which could even be classified as misleading or dangerous. We recognize that further experimentation with the exact phrasing of our questions could greatly affect its performance. However, in general, ChatGPT may not be as ready for prime time as we had hoped.

Social media posts

This is where ChatGPT could have the most success, but with refinement. We asked ChatGPT:

Write different posts for Twitter and LinkedIn using this content: In New England? Want insights into market share gain and loss? Want to compare your hospital to others? Need access to statistical analysis for strategic planning? Check out Spotlight Analytics

ChatGPT produced:

"If you're a healthcare organization in New England looking for market share insights, hospital comparisons, and statistical analysis for strategic planning, we've got you covered. Spotlight Analytics offers comprehensive solutions to help you stay ahead of the curve. Learn more at our website. #healthcare #marketshare #strategicplanning #SpotlightAnalytics"

Not bad, except for Twitter it's 78 characters too long. It also repeats terms at the end as hashtags instead of tagging them inline (even then it's still 15 characters too long). However, we do like the friendliness and engaging tone of the language ChatGPT added.

The question becomes is it worth the effort when manual steps are still needed? Is it easier to just write the tweets ourselves? We're not sure, but ChatGPT demonstrates it can write publishable tweets with a bit of manual tweaking. We'd consider that moderately successful.

Analyzing articles

ChatGPT had mixed results in this area. In some instances, it did a good job of analyzing structure and style, but it was less successful at summarizing the content and main points of an article.

First, we asked ChatGPT to analyze and provide a summary of an industry article on the difficult time someone had in opting out of data sharing with a third party app (Phreesia) used by a provider during the appointment check in process for OBGYN visits. The original article from Stat was extremely critical about the current state of data privacy, It outlined both misleading practices from Phreesia and the author's attempts to ensure her data was not collected and used without her consent. We asked ChatGPT:

Summarize this article into bullet points:

It returned a set of bullet points that had almost no connection to the actual article. Rather, it read like a corporate overview of Phreesia - if Phreesia had been designed to be a champion of privacy rights for individuals. It ranged from neutral to complimentary, presented as factual, highlighted their excellent privacy services, and discussed how important patient privacy is to them.

The produced result seems misleading. Aside from the quality of the information provided, it did not give us summary points for the article in question. The only connection between the article and the generated bullets were the two words privacy and Phreesia. That's it.

However, we did get better results when we asked ChatGPT to analyze one of our past articles before asking it to write new content in the author’s style. We asked ChatGPT:

I need you to analyze something and learn to write like the author:

The result was a nice review of the article’s structure and style. It noted how the opening grabbed the reader's attention and drew them in and how the authoritative tone established the importance of the topic. It mentioned why and how the use of concrete examples and vivid language helps make the article more compelling and convincing. 

Writing articles

Our final attempt was to see if ChatGPT could write like us. This was an interesting experiment - not completely successful but not a total disaster either. We asked ChatGPT:

Now write a blog post [in author’s style] about the challenges associated with image interoperability

The result was bland and clunky in places. At times it didn't really flow well either. It did include several good points, but somewhat superficially without emphasizing anything or providing those examples it lauded in the source piece it analyzed. It never once mentioned DICOM and many of the other comments were generic at best. For example, this is the entire discussion of privacy and security:

"Data privacy and security also pose significant challenges to image interoperability. Medical images contain sensitive patient information that must be protected to ensure patient privacy. However, sharing images between healthcare organizations and systems can increase the risk of data breaches or other security incidents, potentially compromising patient privacy."

This is correct information, but there is nothing specific to images. Many key issues related privacy or security are not discussed - there's no talk of consent or authentication or access control or any of the other factors related to privacy or security or data protection when it comes to electronic data exchange of any sort.

There's nothing inherently wrong or incorrect in anything the generated article says, but it just feels off and not presented in a sophisticated, detailed, or compelling way. Further, it doesn't exhibit many of the highlighted characteristics it referenced in the analyzed article and the words used do not really align with how the analyzed author would write or even organize this piece.

Summing Up

AI is a very promising technology, but it is still a technology that many don't understand well and, in its current iterations, may not be the panacea some propose it to be. It is a tool, and like any tool, if used carefully within the parameters it was designed for, it works.

If you give ChatGPT a well defined, very constricted mandate like "write social media posts about X" it does a reasonably good job, although it misses basic requirements (such as the maximum length of tweets). It appears to do well at analyzing and critiquing writing but not at emulating the style of writing it just analyzed. It does a pedestrian job at writing anything of length on a specific topic requiring technical understanding. In the example we tried, ChatGPT did not accurately summarize the contents of a specific article, but rather provided inappropriate, off topic responses that were misleading and, under other circumstances, might even be fraudulent.

OpenAI does explain some of these drawbacks. In particular, in their blog post introducing ChatGPT, they note that ChatGPT can write plausible but incorrect content and explain some of the reasons why fixing this is hard. Further, the algorithm finished its training in early 2022, meaning that some of the data it uses may be out of date.

Despite these issues and our outcomes, ChatGPT is fun to play around with (just not ready for prime time). What have you used ChatGPT for? Please share with us what worked and what didn’t. Maybe someday we'll have real artificial intelligence that can write compelling, useful articles but we're not there yet. Until that time, we'll continue to write our newsletter articles ourselves.

Share This: