The most prestigious law school admissions discussion board in the world.
|
 |
 
|
ChatGPT keeps slipping Hindi into its responses to me.
Poast new message in this thread
Date: March 26th, 2026 5:01 PM Author: Sienna Mexican Not flame, it just said: "Avoid casual “side सलाह” directly to the third parties involved"
This isn't the first time this has happened.
|
Date: April 7th, 2026 3:05 PM Author: Nazca Redlines (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49800831) |
|
Date: April 7th, 2026 3:07 PM Author: norwood ultra yeah it does that a lot now. it peaked in 2025 and is total garbage now (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49800838) |
 |
Date: April 7th, 2026 3:11 PM Author: Nazca Redlines How is that even possible? It's beyond garbage. Just do pepper your response with a few words and phrases that not only aren't in the same language, they're in an entirely different script.
Own goals like this and hallucinating cases seem like they'd be easy to fix and would make the product much, much more usable. Just, like, e.g., have ChatGPT double check that each word in the response is English or that each case cited really appears on whatever index of cases it has access to. |
|
Date: April 7th, 2026 3:17 PM Author: in the naked it's rare (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49800868) |
|
Date: April 7th, 2026 3:22 PM Author: Nazca Redlines Is it? The Hindi terms have been in a sizable portion of my ChatGPT responses lately. A poaster below says he's getting them in Chinese and Arabic.
And the case hallucinations happen frequently enough that people are setting up trackers to collect examples--and these are only examples that judges call out and that make the news: https://naturalandartificiallaw.com/ai-hallucination-cases-tracker/ |
|
Date: April 7th, 2026 3:31 PM Author: in the naked idk its rare for me but not unheard of. maybe it happens more to some people. I get arabic and chinese once in a while, but I haven't recently. if its just short isolated bursts then it is probably token drift. if generation destabilizes for a moment it can fall into a separate script family. It can't be easily fixed because the entire way the model works is next token generation. The entire process is making one local choice after another under uncertainty, and if it drifts into a bad section of token space it can fuck up like that for a min before restabilizing. (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49800917) |
|
Date: April 7th, 2026 3:16 PM Author: in the naked for me its always arabic and chinese (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49800867) |
|
Date: April 7th, 2026 3:23 PM Author: Nazca Redlines Interesting. Do your queries have anything to do with China or the Middle East? Mine don't have anything to do with India, so seeing Hindi sprinkled in haphazardly is perplexing. I wonder if it has to do with the training material ChatGPT has in mind--Indian message boards it's drawing from or something, which raises it's own host of concerns. (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49800887) |
|
Date: April 7th, 2026 3:37 PM Author: in the naked I never use it for chinese or arabic but there are other legitimate multi-lingual uses and coding uses that can make it easier to destabilize. I use math symbols, transliterated names, unicode punctuation, and sometimes have it transliterate in original scripts and languages when i'm having it analyze a text I'm reading in translation and want to see the original language. When you use it for stuff like this it becomes more likely that the tokens will drift during that session (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49800925) |
|
Date: April 7th, 2026 4:26 PM Author: Nazca Redlines What do you mean "tokens" "drift"? Why should asking questions in English lead to responses partially in Hindi? (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49801126) |
|
Date: April 7th, 2026 3:37 PM Author: Frutiger Aero tp (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49800926) |
|
Date: April 7th, 2026 3:56 PM Author: Surfeando las estrellas (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49800996) |
|
Date: April 7th, 2026 3:46 PM Author: AZNgirl telling AZNmen Your Civilization Will Die stop furking indian men in the ass (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49800938) |
|
Date: April 7th, 2026 3:54 PM Author: Juan Eighty ChatGPT has been using hundreds of thousands of Indians to provide responses for the last few years
at no point were Machine Learning or Natural Language Processing methods actually used |
|
Date: April 7th, 2026 3:56 PM Author: Surfeando las estrellas tp (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49800998) |
|
Date: April 7th, 2026 4:32 PM Author: AZNgirl telling AZNmen Your Civilization Will Die so similar to XO (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49801172) |
|
Date: April 7th, 2026 4:51 PM Author: oomox (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49801240) |
|
Date: April 7th, 2026 4:55 PM Author: luiskandarnameh (🧐) (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49801253) |
|
Date: April 7th, 2026 4:58 PM Author: Surfeando las estrellas (http://www.autoadmit.com/thread.php?thread_id=5850136&forum_id=2#49801275) |