Chinese on the Outside, English on the Inside
Recorded in Chinese
Why does some AI-assisted writing feel unlike its author? Thoughts on the texture of translated prose, spoken expression, and the standards I hold myself and other people to.
00:00 I'm not a Chinese-language writer,
00:04 nor did I study the humanities.
00:07 But from my perspective,
00:10 I want to examine this "AI flavor."
00:14 Many people have probably noticed
00:18 that since AI arrived,
00:20 our social media has filled with
00:23 articles and posts heavy with AI flavor.
00:27 Even now,
00:30 when we scroll through social media,
00:32 we still encounter many posts
00:36 where those traces remain.
00:39 I wonder:
00:40 why do people allow
00:43 this kind of content
00:45 to appear on their social accounts?
00:48 I'm not opposed to AI-assisted creation.
00:53 What I oppose is using AI to saturate—
00:59 no, using AI to revise your expression
01:03 without shaping it
01:07 into something
01:09 that sounds like your own writing,
01:12 rather than writing polished by AI.
01:16 I think there's a major difference.
01:20 Some people simply publish
01:23 whatever AI gives them.
01:27 That makes me curious.
01:30 Can they not distinguish
01:32 between writing created by a person,
01:35 writing that bears traces of AI,
01:38 and writing produced entirely by AI?
01:41 Between those three levels,
01:47 is it the person publishing it
01:51 who lacks judgment,
01:53 or do we, as readers,
01:58 simply not care?
02:02 I used to dislike AI-flavored writing.
02:08 I even asked AI why
02:12 what it gives us
02:14 seems to have this so-called AI flavor.
02:19 I came to realize
02:21 that almost all the leading models today
02:26 are fundamentally based in English.
02:32 For models trained purely on Chinese data,
02:36 DeepSeek performs very well.
02:40 But when people use it,
02:42 for example, to write,
02:46 they find some sentences
02:48 that no person would ever produce.
02:52 The angles are remarkably unusual.
02:55 Returning to the best language models,
02:59 most are grounded in English.
03:06 So the English we see—
03:11 no,
03:12 the Chinese we see,
03:13 whether in replies to us
03:15 or articles written for us,
03:18 is actually
03:23 English at its core, dressed in Chinese.
03:29 Essentially, the Chinese
03:31 it gives us is translated from English.
03:34 So many recurring structures,
03:39 such as "not this, but that,"
03:43 or "however,"
03:46 along with other
03:49 unnatural-sounding expressions,
03:52 are essentially translated
03:54 from English sentence structures.
03:57 English and Chinese
04:00 have different word orders.
04:03 So direct translation
04:05 creates a strange feeling.
04:08 And that feeling produces
04:12 the AI flavor we recognize.
04:20 Sometimes I wonder
04:22 whether using the term "AI flavor"
04:25 is somehow disrespectful to AI,
04:29 or whether the term
04:32 contains a certain hierarchical perspective.
04:36 Humans naturally assume
04:40 that human writing is superior
04:43 and AI writing inferior.
04:47 Although I use the term,
04:49 I don't consider it inherently derogatory.
04:54 But when I
04:56 put the term out there
04:58 and use it to describe
05:01 posts published by certain people,
05:03 I may unconsciously
05:05 push it toward a derogatory meaning.
05:09 But how should I put it?
05:13 Perhaps I'm simply too exacting.
05:17 I have extremely high standards
05:19 for the accuracy, precision,
05:21 and texture of writing.
05:25 When I used to read translated literature,
05:29 I paid particular attention
05:33 to the quality of the translation.
05:36 When I read translated works,
05:39 I hadn't read the originals first.
05:43 So I couldn't know
05:45 how closely the translation matched,
05:48 but I could always tell
05:51 whether the translated prose flowed.
05:53 I can think right now
05:56 of several translators I like.
06:02 I may no longer remember their prose,
06:05 but I remember some books they translated.
06:09 For example, I love The Peregrine.
06:13 Its prose
06:14 was translated extraordinarily well—
06:17 so beautifully.
06:21 It's truly...
06:23 I adore the prose in that translation.
06:28 And there was another book,
06:34 um...
06:36 translated by Wu Qi.
06:39 Something about fire or flames...
06:47 This is awkward.
06:50 I think the author
06:52 was someone called James.
06:59 Ah.
07:00 I need to look it up.
07:02 I used to love this book.
07:06 This is so embarrassing.
07:08 I've forgotten again.
07:09 This returns to something I said before:
07:14 I honestly
07:17 don't remember the titles
07:19 of books I've read in the past.
07:27 I used to struggle with this:
07:29 why can't I remember the books I've read
07:32 or the films I've watched?
07:35 But some people say,
07:36 even if you don't remember them,
07:38 they've grown into your body.
07:41 They shape your judgment,
07:42 your taste,
07:43 and other things that can't be quantified.
07:48 For now, I'll choose to believe that.
07:51 But I can't guarantee
07:54 that it's actually true.
07:57 It might simply be
08:01 an excuse to forgive yourself
08:07 for your memory or comprehension.
08:11 Anyway, I don't know.
08:15 My God, what was the book called?
08:18 I only remember that Wu Qi translated it.
08:23 I need to search for it.
08:29 Even searching for "Wu Qi"
08:32 doesn't help me find the book,
08:39 because it isn't listed on his profile.
08:44 That's so strange.
08:49 Oh, I found it!
08:51 The book is The Fire Next Time.
08:55 It was translated by Wu Qi
08:58 and written by James Baldwin.
09:01 I also love the prose in this book.
09:06 I mean,
09:09 when I read translated works,
09:12 I really care about their texture.
09:16 Some translated prose,
09:17 when you read it,
09:19 seems to have a certain...
09:23 what is it called?
09:25 I don't know whether "personality"
09:29 is the right word for it.
09:32 When you read certain writing,
09:35 you can feel
09:36 that it's very fluid.
09:38 Some writing feels very serious,
09:40 and some feels very charming.
09:43 You naturally describe the writing
09:46 in these ways.
09:50 You don't deliberately
09:53 attach these words to the writing.
09:56 The moment you read it,
09:58 you simply get that feeling.
10:00 It's like when we
10:02 talk with large language models.
10:04 You can sense
10:05 Gemini's personality,
10:07 Claude's personality,
10:09 and ChatGPT's personality.
10:12 Your feelings or perceptions of them
10:16 take on a kind of visible form.
10:24 I find that quite remarkable.
10:29 Although I haven't received
10:32 any formal literary education,
10:39 my standards for writing
10:43 and for its texture
10:50 are still quite high.
10:53 That includes my own expression.
10:55 Sometimes I think, my God,
10:57 how did I even manage to say that?
11:01 I feel my own expression is clumsy.
11:04 Sometimes when I want to say something,
11:11 I can't find the right words.
11:15 There's another thing:
11:17 when I listen to
11:20 raw podcast recordings
11:23 and hear someone speaking,
11:26 in a single sentence,
11:27 they may repeat so many
11:32 like "this, this, that, that, um,"
11:36 "ah, ah," or "uh, uh."
11:39 Instead of saying a word smoothly,
11:44 they repeat its first syllable
11:46 two or even three times.
11:50 And I think, my God,
11:52 why do they speak like that?
11:54 If they speak that way,
11:55 how exhausting is it for the listener?
12:00 I suppose I'm a little mean.
12:04 My own communication isn't great,
12:07 yet I hold others to very high standards.
12:11 So I should
12:13 first train my own abilities
12:16 before I'm qualified to hold others to those standards.
12:20 No, that isn't right either.
12:22 Everyone is different,
12:26 so you can't demand others be like you,
12:30 or expect
12:31 them to meet your standards.
12:35 After all, sometimes
12:36 even you don't meet your own standards.
12:39 So, returning
12:42 to this issue of AI flavor,
12:45 I can't be too harsh on other people.
12:50 If everyone genuinely thinks it's fine,
12:55 that it's all okay,
12:58 then perhaps I'm simply too exacting.
13:01 Or perhaps content creators
13:06 naturally set very high standards.
13:11 I think that's genuinely a good thing,
13:15 because it means
13:16 you truly want to make good content—
13:18 content worth preserving.
13:24 Returning to large language models,
13:28 if we want AI
13:31 to produce writing without this AI flavor,
13:35 then first,
13:37 we need to train it on a vast amount
13:42 of excellent, high-quality Chinese text,
13:48 to feed the models, so to speak.
13:52 But I feel
13:54 that this high-quality content,
13:57 compared with English-language material,
14:00 is far too scarce.
14:06 I don't know how many people
14:08 have ever encountered
14:10 articles on mass-market platforms,
14:15 or novels there.
14:18 I find that world genuinely frightening.
14:22 More frightening than I ever imagined.
14:27 I never knew this kind of writing existed.
14:37 Maybe I'm taking it all too seriously.