Hi, Shrewd!        Login  
Shrewd'm.com 
A merry & shrewd investing community
Best Of Politics | Best Of | Favourites & Replies | All Boards | Post of the Week! | How To Invest
Search Politics
Shrewd'm.com Merry shrewd investors
Best Of Politics | Best Of | Favourites & Replies | All Boards | Post of the Week! | How To Invest
Search Politics


Halls of Shrewd'm / US Policy
Unthreaded | Threaded | Whole Thread (2) |
Author: wzambon 🐝 HONORARY
SHREWD
  😊 😞

Number: of 75963 
Subject: Think of AI as a college student taking a test
Date: 03/20/26 12:02 PM
Post New | Post Reply | Report Post | Recommend It!
No. of Recommendations: 4
The science of why AI makes shit up

Researchers at Georgia Tech and OpenAI (ironically) studied¹ why AI models often hallucinate.

The best way to think about this: Think of AI like a college student taking a test.

They do their best on every question. Some questions they know for sure. But when a question arises that they don’t know, they don’t write “I don’t know” in the answer box. They give an answer—because that gives them at least a chance of being right and scoring points.

The scientists found:

AI models make things up because they are trained to produce the most likely next words, not to verify the truth. “Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty,” the researchers wrote.
Some hallucinations are unavoidable because many facts are sparse, arbitrary, or don’t follow a clear pattern in the training data.
How the models are trained and graded is lacking. Most tests of the models reward correct guesses but give no credit for admitting uncertainty—e.g. they get no points for “I don’t know.” So models are pushed to answer confidently rather than admitting uncertainty.

The takeaway: hallucinations are a built-in risk of current language models, and reducing them will require better incentives for honesty.


https://www.twopct.com/p/the-expedition-march-26-e...
Post New | Post Reply | Report Post | Recommend It!
Print the post
Unthreaded | Threaded | Whole Thread (2) |


Announcements
US Policy FAQ
Contact Shrewd'm
Contact the developer of these message boards.

Best Of Politics | Best Of | Favourites & Replies | All Boards | Followed Shrewds