Even if it's passed the USMLE, #ChatGPT is not an MD.
LLMs aren't constrained to truth, and ChatGPT's biggest hurdle for #healthcare applications is "AI hallucination" (that is, confidently making stuff up.)
New @statnews:
statnews.com/2023/01/26/cha…
LLMs aren't constrained to truth, and ChatGPT's biggest hurdle for #healthcare applications is "AI hallucination" (that is, confidently making stuff up.)
New @statnews:
statnews.com/2023/01/26/cha…
Last week Meta's chief #AI scientist called #ChatGPT "nothing revolutionary, although that's the way it's perceived in the public."
Brief explainer of how ChatGPT works in my piece above, with better ones from @OpenAI's site.
Brief explainer of how ChatGPT works in my piece above, with better ones from @OpenAI's site.
https://twitter.com/kosuke_tsujino/status/1618154863866810369
The TL;DR is that human feedback is the "special sauce" that puts ChatGPT apart from its predecessors, but it
1) introduces a lot of places for humans to bias the algorithms
2) can depend on absolutely terrible human labor conditions
time.com/6247678/openai…
1) introduces a lot of places for humans to bias the algorithms
2) can depend on absolutely terrible human labor conditions
time.com/6247678/openai…
The biggest fear is that ChatGPT is right enough of the time to give the "impression of greatness," per CEO Sam Altman.
But its agnosticism to truth, propensity to make up references, and not-up-to-date knowledge could make it a big problem in medicine.
statnews.com/2023/01/26/cha…
But its agnosticism to truth, propensity to make up references, and not-up-to-date knowledge could make it a big problem in medicine.
statnews.com/2023/01/26/cha…
I didn't cover this in my story, but @NewsGuardRating has an excellent piece on how ChatGPT could become (or maybe already is) a low-barrier source of #misinformation.
newsguardtech.com/misinformation…
newsguardtech.com/misinformation…
And the controversy over listing ChatGPT as an author? The authors consulted it like "a colleague"; it revised sections of the paper.
"We'll leave it to the associate editors handling our manuscripts how they wish to proceed with that," said Tseng.
"We'll leave it to the associate editors handling our manuscripts how they wish to proceed with that," said Tseng.
https://twitter.com/kareem_carr/status/1613325029848092672
Read about this and more, including ethical guardrails for AI, in my piece for @statnews: statnews.com/2023/01/26/cha…
If you don't have access, you're in luck: STAT+ is running a special through tomorrow; use STAT25 at checkout for 25% off an annual membership.
(You can also try STAT+ free for 30 days, or get a bigger discount if you're a student/academic!)
Deets: statnews.com/stat-plus/
(You can also try STAT+ free for 30 days, or get a bigger discount if you're a student/academic!)
Deets: statnews.com/stat-plus/
• • •
Missing some Tweet in this thread? You can try to
force a refresh
