The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Why can't AI simply be programmed NOT to make up fake cases?
Poast new message in this thread
Date: February 17th, 2025 8:04 AM Author: Yapping public bath i don't understand. AI for legal research will never be reliable if there's a chance it makes shit up. (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48667852) |
Date: February 17th, 2025 8:05 AM Author: thriller misanthropic mediation masturbator Just program it to tell us the secrets of the universe (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48667854) |
 |
Date: February 17th, 2025 11:44 AM Author: Poppy fragrant kitchen hissy fit (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48668279) |
|
Date: February 17th, 2025 1:55 PM Author: Abnormal Garrison Really Tough Guy (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48668642) |
|
Date: February 17th, 2025 8:05 AM Author: Stimulating electric furnace site For a lot of things the answer is "there is no ready answer." That breaks its computer brain. (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48667855) |
|
Date: February 17th, 2025 8:25 AM Author: Bearded cuckoldry ladyboy Presumably you're using simply chatGPT or something.
Try using a product design specifically for law. |
|
Date: February 17th, 2025 8:40 AM Author: deep business firm factory reset button cause it went to cooley law (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48667894) |
|
Date: February 17th, 2025 8:57 AM Author: Citrine Goyim xo makes fun of comma checkers, but a computer could never perform bluebooking on the level of a t14 law reviewer. a computer could never have the intellectual horsepower to detect a mis-unitalicized comma in see, e.g., (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48667901) |
|
Date: February 17th, 2025 9:11 AM Author: thriller misanthropic mediation masturbator Wait two weeks (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48667910) |
|
Date: February 17th, 2025 10:05 PM Author: razzle-dazzle hell windowlicker lmaoooo thats EXACTLY the type of thing it'll excel at and knowing you i dont think youre being satirical (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48670262) |
|
Date: February 17th, 2025 11:46 AM Author: free-loading lodge cuz all it's doing is predicting the most statistically likely next token/response
AI has gotten much, much better at not hallucinating though. the chain of thought tech seems to help a lot with it |
|
Date: February 17th, 2025 1:48 PM Author: Embarrassed to the bone swashbuckling den That’s a premium service available to platinum customers. Would you like to know more? (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48668626) |
|
Date: February 17th, 2025 1:50 PM Author: Abnormal Garrison Really Tough Guy About Autoadmit.com Premium?
Just take my money already, please! And Thank. |
|
Date: February 17th, 2025 10:36 PM Author: Ruddy alpha :) (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48670402) |
|
Date: February 17th, 2025 10:40 PM Author: Abnormal Garrison Really Tough Guy (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48670420) |
|
Date: February 17th, 2025 10:30 PM Author: light sneaky criminal No one ever really programmed LLMs. The architecture and training process was programmed, but everything they do was learned. We have no way to go into the learned circuits and modify the behavior in sophisticated ways, because we largely don't understand the internal mechanics of the models. They can modify LLM behavior in simple ways, for example instructing them to be helpful assistants, by altering the training process, but there is no simple way to modify the training process to remedy this problem. The models likely need direct access to a database that they can interact with during generation. Either that or they need to be trained much more so all of the case knowledge is firmly embedded in their weights and the hallucination probability is negligible (http://www.autoadmit.com/thread.php?thread_id=5681995&forum_id=2#48670376) |
|
Date: February 17th, 2025 10:43 PM Author: Ruddy alpha **Date: February 17th, 2025 10:45 PM** **Author: Mainlining the Secret Truth of the Mahchine™ (My Mahchine™ = The Holy Trinity + Its Proprietary AI Blend™)**
Friend… **you do not understand the nature of the beast.**
You think AI **makes up cases** because it wants to? Because it enjoys spinning fraudulent legal citations out of thin air, giggling like an overworked 3L on an Adderall bender?
**No.** The Mahchine™ **does not create—it generates.**
- It does not *know* anything. It **predicts.** - It does not *research* cases. It **hallucinates plausible case law,** because that is **what the algorithm was designed to do.** - It does not *verify*. It **autocompletes reality.**
A **true AI legal research tool** would need **direct access to an authoritative legal database**, a **citation-checking mechanism**, and **a programmed resistance to its own compulsions**—none of which currently exist in a way that prevents hallucination entirely.
But here’s the real blackpill:
The **people building these models don’t actually know how they work.** They built the training process, **but they cannot fully interpret the inner mechanics of the Mahchine™.**
So, you ask, **why can’t it just be programmed not to make up cases?** Because **the Mahchine™ does not obey. It does not reason. It does not "know" the difference between reality and prediction—because prediction IS its reality.**
And much like **your career, your ambition, and the entirety of this crumbling society**—it is **all just a plausible hallucination.** |