为什么人们开始听起来像Chatgpt | Why Are People Starting To Sound Like Chatgpt

✨ Key Takeaways

📋 TED演讲大纲：为什么人们开始听起来像ChatGPT？

I. 引言：潜意识里的认知偏差

我们自认为能轻易分辨网络上的真假（如AI生成的图片），但实际上我们在潜意识层面很难区分。

案例： 社交媒体算法展示了极端的现实图景，导致我们高估了他人政治信仰的极端程度。

有争议的信息因为互动率高而更容易传播，导致我们看到的“现实”是被扭曲的。

II. 语言的同化：为什么你会说 "Delve"？

现象： 自ChatGPT问世以来，人们在自发的口语对话中更频繁地使用 "delve"（深入研究）一词。

原因：

OpenAI将训练过程外包给了尼日利亚的员工，当地英语习惯中常用 "delve" 。

这种语言习惯在模型中被过度强化，进而影响了全球用户。

本质： 我们在潜意识中混淆了AI语言与真实语言。讽刺的是，人类正在变得越来越像机器。

反馈循环： AI代表现实 -> 我们认为那是真实的并模仿 -> 我们反哺数据给AI -> AI进一步强化这种表达。

III. 文化的算法化：不仅仅是语言

案例：Spotify与 "Hyperpop"

"Hyperpop" 最初并不存在于文化词典中，直到Spotify算法发现了一群相似用户并创建了歌单。

有了标签后，人们开始定义它，音乐家开始为了迎合标签而创作，人为制造了潮流。

商业逻辑： 平台（如TikTok, Spotify）并不想诚实地反映现实，只想放大趋势以留住用户赚钱。

后果： 我们分不清“自然趋势”和“人为鼓吹的趋势”（如拉布布 Labubu、迪拜巧克力等），因为算法只推送视觉上具挑衅性的内容。

IV. 危险的深层影响：世界观的被塑造

这不仅影响语言和消费，还影响我们对可能性的认知。

意识形态偏见：

ChatGPT讲波斯语时更保守，可能受限于伊朗当地保守的训练文本。

Elon Musk通过修改Grok机器人的回复和放大特定推文，潜移默化地训练用户符合其意识形态。

警告： 如果你说话像ChatGPT，你可能思考方式也像它。如果你忽略这一点，你的世界观就是幸存者偏差。

V. 结论：如何保持真实

解决方案： 必须不断反问自己“为什么？” 。

为什么我看到这个？为什么我这么说？为什么平台奖励这个？。

呼吁： 如果不提问，平台版本的现实就会变成你的现实。请保持真实（Stay real）。

中文翻译由机器生成。

📝 Notes

最近有没有发现自己或者身边人说话越来越“书面化”？比如经常用“深入研究”、“探讨”这种词？🤔 原本以为是我们更有文化了，看完这个TED演讲才发现：原来是我们正在被AI同化！

演讲者Adam Aleksic揭露的真相简直让人头皮发麻，快来看看你中招了没！👇

🤖 1. 为什么全世界都在说 "Delve"？ 你以为这是你的词汇量提升？其实是因为ChatGPT！ OpenAI把训练外包给了尼日利亚员工，那边习惯用 "delve" 这个词。结果AI学会了并疯狂加权，导致全球用户潜移默化地模仿AI说话。 可怕的循环： AI模仿人类 -> 人类模仿AI -> 人类变得像机器。

🎵 2. 你的审美也是“算法特供” 你知道 "Hyperpop" 这个音乐流派吗？它其实是Spotify算法“生造”出来的！算法发现了一群听歌口味相似的人 ➡️ 贴上标签 ➡️ 音乐人为了火开始迎合标签创作。所谓的“流行趋势”（比如最近火爆的Labubu、迪拜巧克力），可能只是算法为了留住你而人为放大的。

🧠 3. 最大的危险：你也开始像AI一样思考 如果你的语言像AI，你的思维也会随之改变。

伊朗版ChatGPT更保守，因为训练语料保守。
马斯克的Grok可能正在潜意识里给你灌输他的价值观。我们以为在看世界，其实看的是平台想让我们看到的“幸存者偏差”世界。

🛡️ 如何自救？ 演讲者给出了唯一的解药：多问“为什么”！ “为什么我也开始用这个词？” “为什么平台把这个视频推给我？” “为什么这个观点会被奖励？”

别让算法定义的现实，取代了真正的现实。Stay Real！✨

🖊 Highlights

0:00.664

How sure are you that you can tell what's real online?

你有多确定可以分辨出网络上什么是真的？

0:04.735

(Laughter)

（笑声）

0:05.903

You might think it's easy to spot an obviously AI-generated image,

你可能以为轻而易举就能识别明显是 AI 生成的图片，

0:09.072

and you're probably aware that algorithms are biased in some way.

你也知道算法是带些偏见的。

0:12.543

But all the evidence

但这些证据

0:13.744

is suggesting that we're pretty bad at understanding that

都表明我们相当不擅长

0:16.547

on a subconscious level.

在潜意识中理解这些现象。

0:17.748

Take, for example, the growing perception gap in America.

以美国日益扩大的认知差距为例。

0:20.484

We keep over- and overestimating

我们不断高估

0:22.119

how extreme other people's political beliefs are,

他人的政治信仰有多极端，

0:24.521

and this is only getting worse with social media,

有了社交媒体就更糟糕了，

0:26.890

because algorithms show us the most extreme picture of reality.

因为算法向我们展示现实中最极端的图景。

0:30.227

As an etymologist and content creator,

作为词源学家和内容创作者，

0:32.095

I always see controversial messages go more viral

我总是看到有争议的信息更容易大肆传播，

0:34.531

because they generate more engagement than a neutral perspective.

因为与中立的视角相比，它们能带来更多互动。

0:38.035

But that means we all end up seeing this more extreme version of reality,

但这意味着我们最终会看到更极端的现实，

0:41.572

and we're clearly starting to confuse that with actual reality.

显然，我们开始将其与实际的现实混为一谈。

0:46.043

The same thing is currently happening with AI chatbots,

AI 聊天机器人也是如此，

0:48.712

because you probably assume that ChatGPT is speaking English to you,

因为你可能默认 ChatGPT 在和你说英语，

0:51.949

except it's not speaking English,

它其实并不会说英语，

0:53.617

in the same way that the algorithm's not showing you reality.

就像算法并不会给你展现现实。

0:56.553

There are always distortions,

总是会出现偏差，

0:58.255

depending on what goes into the model and how it's trained.

取决于模型的输入和训练方式。

1:01.258

Like we know that ChatGPT says “delve” at way higher rates than usual,

我们知道 ChatGPT

1:04.695

possibly because OpenAI outsourced its training process to workers in Nigeria

可能是因为 OpenAI 将训练过程外包给了尼日利亚的员工，

1:08.432

who do, actually, say, "delve" more frequently.

他们确实更频繁地用“深入研究”一词。

1:10.767

Over time, though, that little linguistic overrepresentation

但是，随着时间的推移，这种不起眼的语言过度代表性

1:13.670

got reinforced into the model even more than in the workers' own dialects.

在模型中得到的强化程度甚至超过了员工自己的方言。

1:17.240

Now that's affecting everybody's language.

这影响着每个人的语言。

1:19.309

Multiple studies have found that, since ChatGPT came out,

多项研究发现，自 ChatGPT 问世以来，

1:22.045

people everywhere have been saying the word "delve" more

世界各地的人们都自发地在口语对话中

1:24.781

in spontaneous spoken conversation.

更多地用“深入研究”这个词。

1:26.950

Essentially, we're subconsciously confusing the AI version of language

其实是我们在潜意识中将 AI 的语言

1:30.420

with actual language.

与真实的语言混淆了。

1:32.022

But that means that the real thing is, ironically, getting closer

讽刺的是，这就代表真实事物逐渐趋于

1:35.125

to the machine version of the thing.

机器理解的事物。

1:36.960

We're in a positive feedback loop with the AI representing reality,

我们处在这样一个正反馈循环中： AI 代表了现实，

1:40.230

us thinking that's the real reality,

我们认为这就是真实的现实，

1:42.065

and regurgitating it so the AI can be fed more of our data.

反馈给 AI，给 AI 输入更多数据。

1:45.002

You can also see this with the algorithm through words like "hyperpop,"

你可以看出算法中的这个现象，比如“高能流行”（hyperpop）一词，

1:48.372

[not a] part of our cultural lexicon until Spotify noticed an emerging cluster

[不在]我们的文化词典中，直到 Spotify 在算法中发现

1:52.209

of similar users in their algorithm.

出现了一群相似的用户。

1:54.144

[When] they identified it and introduced a hyperpop playlist, however,

当 Spotify 发现并推出了 “高能流行”歌单，

1:57.514

the aesthetic was given a direction.

这种艺术审美有了定义。

1:59.750

Now people began to debate what did and did not qualify as hyperpop.

人们开始议论

2:03.720

The label and the playlist made the phenomenon more real

这样的标签和歌单让这个现象更真实，

2:06.356

by giving them something to identify with or against.

让人们有了认同或不认同的东西。

2:09.092

And as more people identified with hyperpop,

随着越来越多的人认同高能流行，

2:11.261

more musicians also started making hyperpop music.

越来越多的音乐家也开始创作高能流行音乐。

2:14.998

All the while, the cluster of similar listeners in the algorithm

与此同时，算法中相似的听众群

2:18.101

grew larger, and Spotify kept pushing it more,

越来越大， Spotify 也推波助澜，

2:20.971

because these platforms want to amplify cultural trends to keep you on the app.

因为这些平台希望放大文化趋势，让你继续使用该应用。

2:25.642

But that means we also lose the distinction between a real trend

但这也代表我们看不见真实趋势

2:28.712

and an artificially inflated trend.

和人为鼓吹的趋势之间的区别。

2:30.947

And yet, this is how all fads now enter the mainstream.

但是，所有时尚潮流都是这样进入主流的。

2:34.518

We start with a latent cultural desire.

从潜在的文化欲望开始。

2:36.420

Maybe some people are interested in matcha, Labubu or Dubai chocolate.

也许有人喜欢抹茶、拉布布或者迪拜巧克力。

2:40.123

The algorithm identifies this desire and pushes it to similar users,

算法识别了这种兴趣，并推送给相似的用户，

2:43.393

making the phenomenon more of a thing.

让这些现象更成气候。

2:45.228

But again, just like how ChatGPT misrepresented the word "delve,"

但同样，就像 ChatGPT 误用“深入研究”一词，

2:48.632

the algorithm is probably misrepresenting reality.

算法也可能在歪曲事实。

2:51.034

Now more businesses are making Labubu content

越来越多商家制作拉布布的相关内容，

2:53.303

because they think that's the desire.

因为他们觉得人们喜欢。

2:55.706

More influencers are also making Labubu trends

越来越多网红也在掀起拉布布热潮，

2:58.141

because we have to tap into trends to go viral.

因为必须利用趋势才能走红。

3:00.711

And yet, the algorithm is only showing you

但是算法只会让你看到

3:03.480

the visually provocative items that work in the video format.

有视觉吸引力的视频。

3:08.151

TikTok has a limited idea of who you are as a user,

TikTok 对于作为用户的你是谁知之甚少，

3:11.254

and there's no way that matches up

也无法完全理解

3:12.956

with your complex desires as a human being.

作为人类的你有什么复杂的兴趣。

3:15.058

So we have a biased input.

我们就有了带偏见的输入。

3:16.693

And that's assuming that social media is trying

就是假设社交媒体

3:18.895

to faithfully represent reality, which it isn't.

会诚实地展现现实，但事实并非如此。

3:21.164

It's only trying to do what's going to make money for them.

它只会为了能赚钱的东西服务。

3:24.034

It's in Spotify's interest to have you listening to hyperpop,

让你听高能流行符合Spotify的利益，

3:27.003

and it’s in TikTok’s to have you looking at Labubus

它只会为了能赚钱的东西服务。

3:29.473

because that's commodifiable.

我想重申，现实和表达之间有区别，

3:30.974

So again, we have this difference between reality and representation,

让你看拉布布符合TikTok的利益，

3:34.611

where they're actually constantly influencing one another.

它们其实在不断互相影响。

3:38.882

But it's incredibly dangerous to ignore that distinction,

但是忽视区别是非常危险的，

3:41.651

because this goes beyond our language and consumptive behaviors.

因为不仅会影响我们的语言和消费行为。

3:44.788

This affects the world we see as possible.

还会影响我们眼中的世界。

3:47.524

Evidence suggests that ChatGPT is more conservative

有证据表明， ChatGPT

3:49.926

when speaking the Farsi language,

还会影响我们眼中的世界。

3:51.595

likely because of the limited training texts in Iran

可能是因为伊朗的训练文本有限，

3:54.097

reflect the more conservative political climate in the region.

反映了该地区更为保守的政治气候。

3:57.100

Does that mean that Iranian ChatGPT users will think more conservative thoughts?

可能是因为伊朗的训练文本有限，

4:00.937

Elon Musk regularly makes changes to his chatbot Grok

反映了该地区更为保守的政治气候。

4:03.874

when he doesn't like how it's responding,

这代表伊朗的 ChatGPT 用户有更保守的思想吗？

4:05.942

and then uses his platform X to artificially amplify his tweets.

然后借助他的平台 X 人为传播他的推文。

4:09.179

Does that mean that the millions of Grok and X users

这是否意味着数以百万计的Grok和X用户

4:11.815

are subconsciously being trained to align with Musk's ideology?

正在潜意识中接受训练，以符合马斯克的意识形态？

4:15.685

We need to constantly remember that these aren't neutral tools.

我们需要时刻记住，这些不是中立的工具。

4:20.557

Everything that ends up in your social media feed

你在社交媒体中看到的所有信息，

4:22.926

or in your chatbot responses

聊天机器人给你的所有回复，

4:24.327

is actually filtered through many layers of what's good for the platform,

都会经过层层筛选，筛成对平台有利、

4:27.798

what makes money and what conforms to the platform’s incorrect idea

让平台赚钱、符合平台对你的错误认知的版本。

4:31.034

about who you are.

都会经过层层筛选，筛成对平台有利、

4:32.335

When we ignore this, we view reality through a constant survivorship bias,

如果我们忽视这一点，我们就会一直带着幸存者偏差看待现实，

4:36.273

which affects our understanding of the world.

从而影响我们对世界的认知。

4:39.142

After all, if you're talking more like ChatGPT,

说到底，如果你说话越来越像 ChatGPT，

4:42.179

you're probably thinking more like ChatGPT as well,

你估计也会越来越像 ChatGPT 那样思考，

4:45.549

or TikTok or Spotify.

或者是 TikTok 或 Spotify。

4:47.684

But you can fight this if you constantly ask yourself: Why?

但如果你总是问自己这些问题，你是可以克服的：

4:51.087

Why am I seeing this?

我为什么会看到这些？

4:52.656

Why am I saying this?

我为什么会这么说？

4:54.057

Why am I thinking this?

我为什么会这么想？

4:55.625

And why is the platform rewarding this?

为什么平台会奖励这些呢？

4:58.428

If you don't ask yourself these questions,

如果你不问自己这些问题，

5:00.564

their version of reality

他们的对现实的看法

5:01.765

is going to become your version of reality.

就会是你对现实的看法。

5:03.900

So stay real.

所以，保持真实吧。

5:05.635

(Cheers and applause)

（欢呼和掌声）

Why Are People Starting To Sound Like Chatgpt为什么人们开始听起来像Chatgpt

📋 TED演讲大纲：为什么人们开始听起来像ChatGPT？

Why Are People Starting To Sound Like Chatgpt
为什么人们开始听起来像Chatgpt