跟讀練習: ChatGPT vs. IELTS Examiner: Writing - 通過影片學習英語口語

選擇練習模式

每種模式都會開啟全屏練習室。

136 句

如果句子過短或過長，請點選 Edit 進行調整。

Over the past few months I've been testing Chagipity and other language models to see if they can act like IELTS examiners.

For this video I took 30 real essays and compared AI band scores to real examiner scores.

So let's look at what happened and what it means for your preparation.

It all started when I received this message from Maria.

I've been preparing for my IELTS for five months.

But I think I'm gonna fail my exam anyway.

I get bent eight or nine in listening and reading.

But writing almost never reaches span seven.

I must score a seven everywhere to start my master's in September.

What can I do?

Well, Marie had been using chat APT to check how good her essays were

and i decided to do the same i took one of her essays

and gave this prompt to chat gpt

you are an experienced ielts examiner who assesses ielts writing task

2 essays using the ielts band descriptors give me the band scores for the four assessment criteria

and the overall task 2 band score.

I posted the essay and the result stunned me.

Task response or how well you respond to the task band 6.

Coherence and cohesion or the flow and structure of your essay band 6.

Lexical resource or vocabulary band 6.

Grammar and overall band 6.

Are you sure this score is so low?

I asked.

And ChattyPT gave me a full explanation of why it couldn't possibly be any higher.

So I sent the same essay using the same prompt to several other language models.

Claude gave a band 7 for task response and coherence and cohesion but only band 6 for vocabulary and grammar,

which resulted in an overall 6 .5.

DeepSeek was more generous, giving a band 7.

Gemini also landed at a band 6.

And since I have a paid subscription for Chagipity, I could test its latest 4 .5 model with more advanced reasoning,

or at least definitely the slowest model.

And that one gave a band 7.

To be honest, none of that looked right to me.

So we decided to conduct a more thorough experiment and see if AI can actually give accurate scores,

which models perform better, and if the scores are wrong, are they usually too high or too low.

A St. Marie's essay to a former IELTS examiner on our team and while I was waiting.

I found 27 essays from Cambridge practice tests so books like that

and all the essays are graded by official IELTS examiners as you can see each essay comes with a score

and examiner's notes so I uploaded these essays into chat apt one by one

and what I found could save you weeks of or wasted preparation.

So let's take the first essay and upload it with the same prompt as before.

So chattypt thinks it's a band 6 but I see two problems.

So first the grammar score is listed as 5 .5

but in IELTS you can only get a band 5 or band 6 for each criterion not 5 .5.

So this would be a band 5.

And the overall score wouldn't be a 6 either.

To get a band 6 overall you need at least a 6 in each criterion.

But in any case the examiner's score for this essay is a band 7, not a 6.

Next essay, band 5 .5 according to HRGPT, but again the examiner gave it a 7.

And that's a huge difference.

Essay 3.

So Chair GPT says band 5.

Yet once again this is a band 7.

Now look at this one.

Chair GPT calls it a 5 .5 but the examiner's score is a band 8.

That's a two and a half band difference

and I'm going to show you the essay on the screen it's really good how could it possibly be 5 .5

At this point I started wondering whether it even made sense to continue.

Every essay I'd sent so far had been undervalued by at least one full band

and in some cases by two and a half bands.

But then I realized something.

Maybe the scores dropped after I told chat GPT not to use half bands when marking individual criteria.

So I started a brand new chat with just the initial prompt and things began to improve.

Suddenly there was a correct score band 6

when it was really a band 6

and overall these scores were now mostly just half a band lower than they should be which

Which honestly isn't bad for a language model compared to a trained human examiner.

But here's the thing.

Not a single essay got a score higher than it deserved.

The result was either accurate or too low.

That said, this wasn't a perfect experiment.

Most of the essays were between bands 5 and 7.

Only one was a band 8 and none were a band 9.

but we really want to know if you write a brilliant essay band 8

or 9 will chat gpt recognize it or any other language model

so i took my own essay and another written by one of our examiners

and tested them exactly the same way i tested maria's essay first

i sent both to a different former examiner for scoring not the author to avoid any conflict of interest

and let's start with my essay then kevin's

and finally see how the other examiner scored all three of us including maria

so let me send my essay to chdpt okay a band 8 for task response band 8 for coherence

Basically a band 8 for the whole essay.

Well that's more inspiring.

What if we try the smartest model 4 .5?

Is it gonna be any different?

Actually no it's still a band 8 but with a band 9 for grammar so thank you very much.

And now Claude.

I know a lot of people love it so I thought we should definitely check it.

Okay, so it gave me event 7 for vocabulary and grammar and 7 .5 overall.

Not exactly the kind of examiner I'd want in my test.

I got an 8 .5 in the real exam and this essay better.

What about Deep Seek?

To be honest it's my first time using it.

A vent 8. And Gemini?

Oh really?

A vent 7?

Oh wait look!

It gave a vent 7 in every criterion but the overall score is 7 .5.

Gemini why how does it even work well it gave a long reply

100

and to be honest it's total rubbish

101

it's not even worth reading to you it seems i'm feeling a bit protective of my essay

102

so let's take Kevin's essay and see how well our examiner did it should be fun so chatty bitty band eight

103

And this smarter model.. hmm, that's also a band 8.

104

It's quite strange actually.

105

I did them all and Kevin's essay got exactly the same scores as mine.

106

How is it possible when two different people wrote them?

107

I'm not sure.

108

But the key question is what did our other examiner award us?

109

So my essay scored between 8 .5 and 9 and Kevin's got 8 .5.

110

I didn't tell the other examiner whose essays they were,

111

I just sent them as 1 -2 -3 and Kevin if you're watching this 8 .5 is a fantastic score.

112

So it seems that for essays with really high scores language models do give relatively high scores.

113

But there is one more score left and the one that matters most.

114

Maria's essay.

115

The whole reason why we started this experiment.

116

And Maria's essay got an 8.

117

Imagine you are writing band 8 essays but every time you check with Chagipity it tells you you're band 6.

118

What would bad due to your motivation and Maria isn't the only one either here is another message I received

119

This candidate also got 6 .5 to 7 by CHAT -EPT, but an 8 in a real exam.

120

So, what can we take away from this?

121

First of all, AI tends to underrate essays rather than overrate them.

122

This means if you are relying on AI feedback, you might think you are doing worse than you really are.

123

2. The gap is the biggest for mid -band essays.

124

So band 5 and 6 essays were sometimes pretty close, maybe half a band off.

125

But 7 and 8 essays were often underrated by one or even two full bands.

126

3. AI can recognize quality, it knows when an essay is good.

127

But it really struggles to award those top -level scores.

128

And AI doesn't follow IELTS scoring rules correctly, like using half -bands when they don't even exist.

129

And finally, the results vary a lot between models.

130

So use AI to check your vocabulary and grammar, or get help with ideas and flow.

131

But don't rely on it for scoring.

132

Instead, learn the requirements examiners use for yourself

133

and I go through every single one of them while writing a full answer in this video here.

134

And thank you so much for watching me today.

135

Good luck with your preparation and your exam.

136

Bye!

下載應用

流利口語所需的一切

AI 發音評分逐句打分

音標練習掌握每個音

詞彙學習建立你的詞庫

詞彙遊戲邊玩邊學

掃碼

Download on theApp Store Get it onGoogle Play

同頻道

FOR YOU

本課的詞彙與口語要點

這節口語課以影片“ChatGPT vs. IELTS Examiner: Writing”為素材。影片中反覆出現的詞有：band, essay, score, examiner, model。這段影片共有 136 個句子、1452 個單詞可供跟讀。講話部分時長為 11:39。說話人語速平穩，每分鐘約 125 個詞，很適合跟讀。只有 78% 的單詞屬於英語最常用的 3,000 詞，詞彙難度較高。

影片中的重點詞彙

影片中 15 個值得學習的單詞，附發音和釋義：

單詞	發音	釋義
essay 名詞	/ˈɛs.eɪ/	雜文 /雜文, 論文 /論文
examiner 名詞	/əɡˈzæmɪnɚ/	檢查員 /檢查員, 調查員 /調查員
grammar 名詞	/ˈɡɹæm.ə(ɹ)/	語法 /語法, 文法
examination 名詞	/ɪɡˌzæm.ɪˈneɪ.ʃən/	檢查 /檢查
vocabulary 名詞	/vəˈkæb.jə.lə.ɹi/	詞匯 /詞彙, 語匯 /語彙
experiment 名詞	/ɪkˈspɛɹ.ə.mənt/	實驗 /實驗
preparation 名詞	/ˌpɹɛpəˈɹeɪʃ(ə)n/	準備 /準備
ventilation 名詞	/ˌvɛntɪˈleɪʃ(ə)n/	通風 /通風, 通氣 /通氣
compare 動詞	/kəmˈpɛɚ/	比, 比較 /比較
rely 動詞	/ɹɪˈlaɪ/	依賴 /依賴, 依靠
upload 動詞	/ˈʌpˌloʊd/	上載 /上載, 上傳 /上傳
assessment 名詞	/əˈsɛsmənt/	評價 /評價, 評估 /評估
explanation 名詞	/ˌɛkspləˈneɪʃən/	解釋 /解釋, 說明 /說明
resource 名詞	/ɹɪˈzɔːs/	資源 /資源
spin 動詞	/spɪn/	轉動 /轉動, 自轉 /自轉

影片中出現的短語動詞

單詞	釋義
go through 動詞	通過 /通過

值得反覆跟讀的句子

影片中簡短而完整的句子，可以直接用在日常對話裡：

What can I do?
Are you sure this score is so low?
To be honest it's my first time using it.

影片中的語法

說話人最常用的結構，並附上影片中的原話：

結構	影片中的用法
現在完成進行時 have/has been + -ing — 以前開始、現在仍在繼續的動作	I've been testing · I've been preparing
被動語態 be + 過去分詞 — 強調發生了什麼，而不是誰做的	are graded · is listed · had been undervalued

需要注意的發音

說話人用了 21 次縮略和弱讀形式，例如 I'm, don't, gonna。請按聽到的簡短形式來說。

“sh” 和 “zh” 音: examination /ɪɡˌzæm.ɪˈneɪ.ʃən/, preparation /ˌpɹɛpəˈɹeɪʃ(ə)n/, ventilation /ˌvɛntɪˈleɪʃ(ə)n/, explanation /ˌɛkspləˈneɪʃən/, motivation /ˈmoʊ.tə.veɪ.ʃən/
長單詞——注意重音位置: examiner /əɡˈzæmɪnɚ/, examination /ɪɡˌzæm.ɪˈneɪ.ʃən/, vocabulary /vəˈkæb.jə.lə.ɹi/, experiment /ɪkˈspɛɹ.ə.mənt/, preparation /ˌpɹɛpəˈɹeɪʃ(ə)n/

中文母語者容易讀錯的音：

/v/ — 不要讀成 /w/，上齒輕觸下唇: vocabulary /vəˈkæb.jə.lə.ɹi/, ventilation /ˌvɛntɪˈleɪʃ(ə)n/, vary /ˈvɛəɹi/, motivation /ˈmoʊ.tə.veɪ.ʃən/, protective /pɹəˈtɛk.tɪv/
詞尾子音 — 要讀清楚，後面不要加母音: bind /ˈbaɪ̯nd/, prompt /pɹɑmpt/, experiment /ɪkˈspɛɹ.ə.mənt/, upload /ˈʌpˌloʊd/, assessment /əˈsɛsmənt/
/r/ 和 /l/ 的區別: vocabulary /vəˈkæb.jə.lə.ɹi/, rely /ɹɪˈlaɪ/, correctly /kəˈɹɛk(t).li/, undervalue /ˌʌndə(ɹ)ˈvælju/

如何用這段影片練習

先完整聽一遍影片，不要開口，記下不認識的單詞。
用正常速度逐句跟讀，每句重複到你的節奏與說話人一致為止。
錄下自己的聲音並與原聲對比，特別注意 essay, examiner, grammar 這類單詞。

什麼是跟讀法？

跟讀法 (Shadowing) 是一種有科學依據的語言學習技巧，最初開發用於專業口譯員的培訓，並由多語言者Alexander Arguelles博士普及。這個方法簡單而強大：您在聽英語母語原聲的同時立即大聲重複——就像是一個延遲1-2秒緊跟說話者的影子。與被動聽力或語法練習不同，跟讀法強迫您的大腦和口腔肌肉同時處理並模模擬實的講話模式。研究表明它能顯著提高發音準確性，語調，節奏，連讀，聽力理解和口語流利度——使其成為雅思口語備考和真實英語交流最有效的方法之一。

影子跟讀法: 閱讀完整分步指南 →

跟讀練習: ChatGPT vs. IELTS Examiner: Writing - 通過影片學習英語口語

選擇練習模式

下載應用

同頻道

How to start your IELTS preparation [for beginners]

Band 9 IELTS Speaking SAMPLE ANSWERS | HOME

3 IELTS Listening Techniques to Understand EVERYTHING

Get Better at IELTS Reading in 13 Minutes

IELTS Speaking Part 3 | Questions & Answers

IELTS Speaking Part 1 SAMPLE ANSWERS | Hometown

推薦影片

The Writing Makeover That Took Her Essay From 6 5 to 8 5

Cambridge IELTS 11 Listening Test 1 | Full Practice Test with Answers (40 Questions)

IELTS Speaking Band 7.5 Mock Test with Feedback

How to Get Band 9 in IELTS Writing Task 2

IELTS Listening Part 3: A Survey on Student Sleep

IELTS Speaking Band 8.5 (Temurbek Akbaraliyev)

RE2 Unit 11B

Peer Assessment: Đề thi thật IELTS LISTENING

IELTS Test with EX EXAMINER!!!! + Language and Tips

IELTS Speaking Exam - How to Do Part Three of the IELTS Speaking Test

IELTS Tips You MUST Know Before Your Test

What Stanford students actually use ChatGPT for