4o is now fabricating English words! Users who are not native or fluent in English should be careful
Rada Vassil
16 replies
Yesterday I asked ChatGPT to improve a sentence and it “improved” it with a made-up word.
Here’s the conversation:
> ChatGPT: There’s a slight awkwardness in the phrasing. Here's a refined version:
"The AI tool will help you make further refinishments.”
> Me: is refinishments a common English word?
> ChatGPT: No, "refinishments" is not a common English word. The correct term to use would be "refinements." I apologize for the mistake in the previous message. Here's the corrected version:
"The AI tool will help you make further refinements.”
This is very strange. We know that genAI hallucinates when trying to fill in gaps in its knowledge that weren't covered during its training.
But ChatGPT is specifically trained on language! It's highly unlikely that it wasn't properly trained on common English words and expression.
Have you had similar experiences?
Replies
Business Marketing with Nika@busmark_w_nika
Something similar happened to Marc Lou: https://x.com/marc_louvion/statu...
Share
Launching soon!
@busmark_w_nika
I asked ChatGPT 4o if the recent update means its algorithm has changed.
Here's the answer: Yes, the recent update to ChatGPT-4 likely means that its algorithm has been improved or modified.
They definitely modified the algorithm, but it wasn't for improvement. it seems to have gotten worse.
@rada_vassil1 😀 If the ChatGPT is trained by the data of the internet users (who started to rely on chatGPT) – that is not very good. If we are getting lazier to think about certain things, our outputs will be poor and if the model is trained on these outputs... :D that is bad. But it probably won't work like that... I think. I am not a programmer.
Launching soon!
@busmark_w_nika
Agree with us getting lazier :)
But ChatGPT is a machine –the humans who program it determine how it will work and how all the data will be used, which affects the output.
If a machine stops working properly and gives unreliable outputs, I'd blame those responsible for the algorithm.
I haven't had such experience yet. Could it be that 4o is learning from people mixing two languages ? I know some English and French words tend to be mixed up and thus creating nonexistent words
Launching soon!
@gabriel_west
It's even more concerning if it's trained by people who aren't language experts and mix languages.
It shouldn't mix languages or fabricate words, especially when my instructions were to improve my sentence.
Omg 🙈 my experience working with 4o for content generation has been frustrating at times in the last few days. I find myself going back to the earlier versions often
Launching soon!
@namrata_arya
It's been more than a few days for me. I noticed the decline in quality immediately after they announced 4.o
I hope they didn't change the algorithm of ChatGPT-4 too.
LOL, ‘refinishments’ sounds like something I'd accidentally say during a presentation.
AI is amazing but definitely not flawless.
strange indeed! Haven't seen anything like it.
That AI mishap sounds frustrating.
Launching soon!
@michael_baumgartner
agreed, it's frustrating. But since ChatGPT is for mass consumption, we can't expect the best quality, IMO
It's like ChatGPT stepped into the wild world of Scrabble and came back with a made-up word! Maybe it's secretly inventing its own dictionary. Who knows what words we might find next?