Sophia Smith Galer
Region: GB
Friday 21 August 2026 15:12:33 GMT
182287
14616
163
906
Music
Download
Comments
D :
I’m a bilingual speaker of French and English and I can tell you I’ve already been noticing this. I need to speak to it in English for certain tasks and in French for others otherwise it’s getting things and tones wrong.
2026-08-21 15:35:33
1011
Splendiferous_Banana :
you're a journalist that writes about language but don't even understand languages impact on culture? ones vocabulary directly impacts the way they see and interact with the world. calling the inherent cultural values that are emphasized by word choice (for example, how some languages use third person pronouns for strangers) biases implies that a person's culture is a choice. it's called nurture for a reason. if the people tend to speak less directly and more deferently in a certain language than the LLM is going to imitate that tone because it has no other pool of source information to pull from, unless you want Claude to start making up words
2026-08-28 04:41:34
0
CooperBond :
No way. Language model is biased towards language it was being trained on 🫨
2026-08-22 01:25:32
207
robyn :
So incredibly interesting and also horrifying!
2026-08-21 15:20:52
518
Babufrik24hourdroidsmith :
Those “values” spectra remind me of the (also bullshit) Myers-Briggs personality test questions where they ask you to choose which you prefer from things that are not opposites. Just silly
2026-08-21 20:27:02
109
Dan :
Ok but as a translation studies scholar this is fascinating. The absolute grial in translation studies is being able to discover how language cultures varies, making direct translation sound strange. In politeness and directness this is very apparent but hard to find objective data for!
2026-08-21 21:46:30
236
(AJAX) TaraGWW :
So in this um… “value set” “warmth” just means “making things up/lying”
2026-08-21 23:32:23
33
Stanley the Corgi :
It’s fascinating that they say “we didn’t code this” when they literally did code it to do this 🙄
2026-08-22 15:06:42
23
timea | beauty journalist :
they sound like societal classes in a dystopian YA novel 😂
2026-08-21 18:22:37
79
Austin Beeman | Wine + Travel :
I would love to see this inside of English, but correlated to grammar, punctuation, linguistics sophistication. If I ask, “Wat u think” vs “what do you think?” is there a different level of accuracy and sophistication in the response
2026-08-23 13:06:33
52
wmute :
You don't even need to speak a different language to experience this. Write like a black person. Or write like a foreigner. Or write like a poor person. Or write like a millionaire. I guarantee you that you will get different answers out of the same model.
2026-08-21 17:58:23
53
krzakken :
everything I learn about AI only makes me hate it even more 🙃🙃
2026-08-24 08:11:54
42
3069hhj :
With the greatest respect, things do not need to be opposites to map a vector between them.
2026-08-21 22:48:21
5
evaskem :
I speak russian, english and french. but prefer using claude in russian because it's closer to the "adhd-mode" im comfortable with. basically, it yaps wayyyy less
2026-08-22 09:55:02
31
Георги :
Its very interesting how Claude has picked up subtle cultural and linguistic norms baked into each language. They train the model on huge amounts of real text in each language, which carries the writing styles, formality rules, and conventions native speakers actually use. So it naturally gets used to it and starts using it too, like how i think i have different personalities for each language i speak. Their post about it is horrible though, lol, you can’t just rank languages by how much they are something
2026-08-21 19:24:01
65
Lady Otaku :
It's interesting. Sometimes I ask AI things in Spanish, my native language, but I don't find the answer good enough so I try a prompt in english and I get a better answer. I don't use Claude, btw. I'll see what happens when I use french and japanese
2026-08-22 05:15:44
18
Fingerhat :
Makes absolute sense. I’m German. If I ask AI to proofread a text and make suggestions, it will keep the tone the way I want it to be. Direct, zero empty phrases. I remember Copilot once asking me whether I would like it to soften the tone and make it less confrontational. Hell no. 😂
2026-08-25 01:23:44
16
Nikkie :
Is this similar to when google answers with genders based on the question? Eg says “he” for doctor or lawyer and “she” for nurse or teacher?
2026-08-21 22:17:14
36
Red🇨🇦 :
If you want to explore this more, DeepSeek would be an interesting example. Given how in-depth it would have Chinese cultural norms when reasoning and responding in Chinese vs English.
2026-08-22 01:11:28
5
the Quirky Club :
Now it makes sense why Claude came across so damn annoying to me in Turkish 😅
2026-08-22 01:47:43
13
Anton Kharitonov :
russian responses are always so different from english for me, even on chatgpt
2026-08-22 02:46:24
6
🔍CRIKIT :
not claude creating the groups from divergent 😂
2026-08-25 07:43:30
5
Jasmine 🇨🇦 🇪🇺🌍 :
What would happen if I prompt it in German and want it to produce the result in English? What “values” profile would I get?
2026-08-21 23:21:26
5
Nate :
I'm also very curious about the *amount* of training data that was used in each of the languages. Looking purely at the number of speakers, I would guess Claude has less data in Japanese (to say nothing of a language like Quechua or Basque) than in English, Spanish, or Chinese. And when there's more or less data available, the output may be more or less representative of the speakers of that language.
2026-08-23 21:04:32
7
Igno :
People speaking different languages have varying values and cultures. That obviously leads to slightly different training data and hence text output. If this is horrifying, so is every piece of media ever produced lol.
2026-08-21 17:50:34
8
To see more videos from user @sophiasmithgaler, please go to the Tikwm
homepage.
