Reza Sirous

[Kareem A Wahid, MD, PhD @Kareem_A_Wahid]

@SirousReza @Radiology_AI I feel this could quick turn into a GPT plagiarism arms race. People already jailbreaking GPTZero! https://t.co/IRvLK0bm4K. #RadAIChat

[Woojin Kim @woojinrad]

7/ and opaque self-censoring method. Such concerns should also be considered when writing about #GPT4 and other closed models. (Oh, I'm sure this blog will lead to many discussions 😉)
#RadAIchat

[Woojin Kim @woojinrad]

4/ However, this blog post (https://t.co/gq8pmmMv5o) proposes the following for the research and scientific publications:

"That which is not open and reasonably reproducible cannot be considered a requisite baseline."

#RadAIchat

[Reza Sirous @SirousReza]

@Radiology_AI A5: Also for initial screening, plagiarism, checking for typos, formatting issues, relevance to the journal’s scope, finding appropriate reviewers #RadAIchat

[Reza Sirous @SirousReza]

@Radiology_AI A5: I think the most crucial role of GPT4 in the peer-review process can be protecting the process against ChatGPT itself, i.e., AI-generated plagiarism. #RadAIchat

[Woojin Kim @woojinrad]

8/ I have a feeling we will have a #RadAIchat session about #AutoGPT, #BabyAGI, and #JARVIS soon. 😉

https://t.co/QJl8Lu2Jze
#GenerativeAI #GenAI

[Woojin Kim @woojinrad]

6/ However, I would not be skating to where the puck is going be if I ignored this trend based on this early experiment alone. This is where the technology is headed, and if you are interested in learning more, this is a nice introduction.

https://t.co/ZhktGEyJT9
#RadAIchat

[Woojin Kim @woojinrad]

T4. While I’m sure #ChatGPT can assist with both, given that it is not specifically trained for radiology and tends to hallucinate/confabulate, I’d vote for improved efficiency as of today.
#RadAIchat

[Reza Sirous @SirousReza]

@Radiology_AI A4: I think since GPT4 is a general-purpose AI, although it can be further trained through prompt engineering and fine-tuning, it will struggle to outperform current radiology AI algorithms. So, I’ll go with efficiency. #RadAIchat

[Woojin Kim @woojinrad]

@OpenAI @elevenlabs @khanacademy @duolingo 4/ This terrific @TEDTalks by @salkhanacademy was just released this week. Using #GenerativeAI like this gives me hope that "we're at the cusp of using #AI for probably the biggest positive transformation that #education has ever seen."

https://t.co/TRAMjytiM0
#RadAIchat

[Reza Sirous @SirousReza]

@Radiology_AI Check this article:
A3: ChatGPT - Reshaping medical education and clinical management. doi: 10.12669/pjms.39.2.7653 #RadAIchat

[Danielle Bitterman, MD @dbittermanmd]

T2. Also, ChatGPT is a chatbot. It is not necessarily the cure-all for all NLP tasks. Our recent preprint explored this: https://t.co/83WqntRa7R #RadAIChat

[Felipe Kitamura @FelipeKitamura]

Great points! Item #3 could be solved with tools that use GPT output to look for things on the internet (AutoGPT and https://t.co/kMj6tEx4ON)

#RadAIChat

[Pouria Rouzrokh, MD @PRouzrokh]

@SirousReza T1: Speaking of good resources, I would also like to highlight this interesting interview with Drs. Tessa Cook, Nina Kottler, and Sonia Gupta on impacts of ChatGPT and other LLMs on radiology reporting.
#RadAIchat
https://t.co/cY8c3XAmpW
@asset25

[Reza Sirous @SirousReza]

@woojinrad A2: Check this opinion article:
Ismail A, Ghorashi NS, Javan R. New Horizons: The Potential Role of OpenAI's ChatGPT in Clinical Radiology. J Am Coll Radiol. 2023 Mar 25:S1546-1440(23)00259-4. doi: 10.1016/j.jacr.2023.02.025. #RadAIchat

[Woojin Kim @woojinrad]

2/
4️⃣ #GPT4 can provide inconsistent responses.
5️⃣ It contains #bias we don’t fully know about.
🚧 When you combine these issues with automation bias, you must be extremely careful, as wrong guidance can mean the difference between life and death in #healthcare. #RadAIchat

[Woojin Kim @woojinrad]

3/
Speaking of #AutomationBias, here’s a @radiology_rsna article just published yesterday highlighting why this is so important.

"Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance"

https://t.co/h0sS4lYdHF
#RadAIchat

[Reza Sirous @SirousReza]

@woojinrad A2: It can be used with direct supervision to assist with decision support for clinicians (e.g., ACR appropriateness criteria) or radiologists (e.g., ACR white papers). #RadAIchat

[Woojin Kim @woojinrad]

1/ T2. My short answer is: not yet.
Here are my reasons.
1️⃣ #GPT4 is not fine-tuned to provide medical information.
2️⃣ LLMs like GPT-4 are well known to hallucinate and confabulate.
3️⃣ Static LLMs will always contain outdated and incorrect information.
🧵
#RadAIchat