Turnitin AI Score 高分處理 — 香港八大 declaration policy 對照

Turnitin AI Score 高咗點算?香港八大 declaration policy + Turnitin 官方 limitation 逐條解讀(2026 更新)

本文由 UniWriterPro 學術規例研究組 verify · Last updated: 2026 年 8 月 18 日 · 本文係按 Turnitin 官方 blog、Turnitin Guides FAQ、Stanford HAI 研究、香港八大 GenAI declaration policy 逐條解讀。冇公開文件嘅位一律標「搵唔到官方文件」,唔搬其他學校套用。

目錄

Turnitin 自己承認嘅 false positive rate

好多人以為 Turnitin AI Score「一定準」——事實 Turnitin 自己嘅 official documentation 已經好明確承認 detection 有 false positive(誤判人手寫嘅文章做 AI 生成)。呢個 admission 對你上訴或者 declare 情況嘅時候係最強嘅武器:對方拎 Turnitin score 打你,你可以引 Turnitin 自己個 statement 反問。Turnitin CPO 官方 blog 明講:

Our efforts have primarily been on ensuring a high accuracy rate accompanied by a less than 1% false positive rate, to ensure that students are not falsely accused of any misconduct.

Turnitin — Understanding false positives

Turnitin Help Guide 進一步 confirm 呢個 <1% figure 只 apply 喺「documents with over 20% of AI writing」——即係 score 20% 以下嘅 false positive rate 唔係 <1%,實際可能高好多——所以 Turnitin 官方而家已經 pull down「score below 20% no longer surfaced」(20% 以下唔顯示):

We strive to maximize the effectiveness of our detector while keeping our false positive rate — incorrectly identifying fully human-written text as AI-generated — under 1% for documents with over 20% of AI writing.

Turnitin Guides — AI writing detection capabilities FAQs

更加重要嘅 disclaimer——Turnitin 自己明確講佢 唔係 determine misconduct 嘅 authority。呢句係你 appeal 或者同 course leader argue 嘅時候必引:

While Turnitin has confidence in its model, Turnitin does not make a determination of misconduct, rather it provides data for the educators to make an informed decision based on their academic and institutional policies.

Turnitin Guides — AI writing detection capabilities FAQs

即係話:Turnitin AI Score 只係一個 data point,唔係判決。做決定嘅係 educator + institutional policy——所以你嘅 declaration 記錄、drafting timestamp、supervisor communication 一齊被 weigh。仲有一個 Turnitin 官方寫得好清晰嘅 limitation——document 頭尾容易 false positive:

Since launch, we have observed a higher incidence of false positive detection in the first few or last few sentences of a document. False positives (incorrectly flagging human-written text as AI-generated) are a possibility in AI models.

Turnitin Guides — AI writing detection model

Abstract、Introduction 開頭同 Conclusion 結尾嘅 false positive rate 特別高——因為呢啲位置學術寫作本身就係最 formulaic、最 template-y。如果你份 assignment 只係 abstract 同 conclusion 被 flag、body 得低 score,呢個係典型 false positive pattern。

Stanford 研究 — 61% ESL essay 被誤判做 AI

香港大部分學生英文係 second language——呢個 fact 對 Turnitin AI Score 有直接、可量化嘅 impact。Stanford HAI(Human-Centered AI Institute)2023 年發表嘅研究 tested 七個 major AI detector(包括市面同 Turnitin 同類 detector),結果對 non-native English writer 有系統性 bias:

The numbers are grim. While the detectors were “near-perfect” in evaluating essays written by U.S.-born eighth-graders, they classified more than half of TOEFL essays (61.22%) written by non-native English students as AI-generated (TOEFL is an acronym for the Test of English as a Foreign Language).

Stanford HAI — AI-Detectors Biased Against Non-Native English Writers

更加嚇人嘅 figure:

It gets worse. According to the study, all seven AI detectors unanimously identified 18 of the 91 TOEFL student essays (19%) as AI-generated and a remarkable 89 of the 91 TOEFL essays (97%) were flagged by at least one of the detectors.

Stanford HAI — AI-Detectors Biased Against Non-Native English Writers

Bias 嘅 mechanism 好清楚:AI detector 用 perplexity(vocabulary diversity 同 sentence complexity)做主要 signal,non-native writer 傾向用 more common phrasing、predictable sentence structure——即係同 AI 生成文本嘅特徵 overlap。你唔係做錯嘢,係 detector 對你嘅語言背景有 systematic bias。Turnitin 自家 2024 內部研究 claim 佢個 detector 對 ELL(English Language Learners)冇 statistically significant bias,但 Stanford 個 study coverage 更廣、方法 more independent——所以呢個 dispute 未 settle。你 declaration 或者 appeal 嗰陣,兩份 evidence 都可以列出俾 educator 睇。

AI Score 20% / 50% / 80% 實際意思

Turnitin AI Score 唔係 linear——20% 唔係「五分之一 AI 寫」咁簡單 mapping。實際上 Turnitin display logic 有兩條清晰嘅 threshold:

AI Score 範圍Turnitin 官方處理False Positive 風險實務意思
0% – 19%唔顯示(官方 2024 更新後 pulled down)高——所以官方索性唔 surface如果你 report 顯示 asterisk 或 blank,即係呢個範圍
20% – 39%顯示 percentage,但標示為 low confidence可觀——<1% false positive rate 只 apply 20%+Instructor 通常唔會直接 escalate,會 request explanation
40% – 79%顯示 percentage + Similarity Report 內 highlight<1% document-level(跟 Turnitin claim)Instructor 一般會發 email 問你 drafting process
80%+顯示 percentage,Turnitin 標為 high confidence官方 claim 極低——但 sentence-level false positive 仍存在通常 trigger academic misconduct investigation

Turnitin 官方 blog 亦承認 sentence-level 同 document-level false positive rate 唔一樣——sentence-level 高過 document-level,所以個別句子被 highlight 唔一定代表你成份係 AI,可能只係嗰句剛好 template-y。細節可以睇 Turnitin — Sentence-level false positive rates

五個常見「明明自己寫都被 flag」場景

1. Abstract、Introduction、Conclusion 開頭三句

Turnitin 自己承認 document 頭尾 false positive rate 高——即係「In this essay, I will discuss…」「To conclude, this study has shown…」呢啲典型學術 opening / closing template 特別容易觸發。Fix:Introduction 唔好用 template opener,直接 lead with concrete finding 或 specific claim。Conclusion 同樣避開「In summary」「To conclude」呢啲 dead giveaway。

2. Methodology 章節 — 「本身就係 formulaic」

Quantitative research 嘅 Methodology 章節(sample size、data collection procedure、statistical analysis)本身有 standard 寫法——例如「A survey was administered to 120 undergraduate students…」呢類句式喺任何 methodology paper 都 look similar。呢個係 discipline-inherent template——detector 分唔清係人寫嘅 template 定係 AI 出嘅 template。Fix:加入 specific detail(instrument name、pilot study 過程、response rate anomaly、adjustment 原因)令 methodology 個性化。

3. Literature Review 綜合段落

當你 paraphrase 幾個 study 做 synthesis paragraph,用嘅 connective phrase(「Similarly」「In contrast」「Building on this」)加 template summary sentence,呢個 pattern 同 AI 生 lit review 極相似。Fix:唔好純 summarise——加入你自己嘅 critical judgement(「呢個研究 sample size 只有 45,generalisability 有限」),加入 detailed critique 令 lit review 由「總結」變「分析」。

4. Grammarly / DeepL / MS Editor 深度改稿

用 Grammarly Premium、DeepL Write、Microsoft Editor 深度改稿——尤其係 non-native writer 用來「執」句式——會令原本 human-written 變到有 AI-like 嘅 uniformity。Grammarly 官方文件明講佢個 rewrite 功能係 GPT-powered,用完就係 AI-assisted content。Fix:只用嚟 flag error,唔好 accept 整段 rewrite;或者 rewrite 後自己再改返啲 phrasing 加 personal voice。呢個場景係最容易 misconception 嘅:學生以為「Grammarly 只係 grammar checker」,實際 rewrite 已經係 GenAI use,多間香港大學 declaration policy 已明確 include(下面對照表詳解)。

5. Reference list、Table caption、Figure caption

APA reference list format 本身係 rigid template(Author, Year, Title, Journal, Volume, Issue, Pages)——detector 見到大量 uniform structure,可能 flag 呢個 section。Table caption「Table 1 shows the distribution of…」都係 template 句式。Fix:Reference list 本身冇得改結構——但如果 Turnitin report highlight 嘅 flag 集中喺 references / captions,你可以指出呢啲係 discipline-required format,唔可能 rewrite 換避 detector。

香港八大 declaration policy 對照 — Turnitin 高分點 declare

Turnitin AI Score 高唔一定代表 misconduct——關鍵係你有冇按院校規例 declare 你 GenAI 嘅使用。八大 policy 差異好大——有啲要 4 element declaration、有啲只要 general disclosure、有啲直接 default「不允許 unless permitted」。以下係 verbatim quote 對照,你揸住自己間 school 條 policy 去處理:

院校Default positionDeclaration 要求Turnitin 高分處理
HKU 港大Course-level permission 為準Course syllabus 講明 permitted,declare 用途Course leader 認可有 declare = 合規;未 declare = investigation risk
CUHK 中大Restrictive default需 course teacher 明示允許並 declare冇 declaration 直接 case
HKUST 科大三級 policy(No AI / Assistive / Full)按 course level policy declare對應唔到 course policy = 違反
CityU 城大Course-specificInstructor 指定 declare format冇 declare 直接 misconduct case
PolyU 理大Permitted with declarationStatement of AI use required有 declare = usually OK;冇 declare = investigation
HKBU 浸大Permissive with limitsAcknowledge use in submissionTurnitin 高分 + 有 acknowledgement = 通常 accept
EdUHK 教大4-element declarationTool / Purpose / Extent / Content declaration4 element 齊全 = 合規
Lingnan 嶺大Free ChatGPT access + syllabus ruleCourse syllabus + Turnitin submissionFree access ≠免 declare
HKMU 都大OLE declaration requiredOALT GenAI Guidelines §2.2 明訂 declare on OLE冇 OLE declaration = 違紀 risk(見都大 HKMU 政策詳解

詳細每校政策同 declaration template 可以參考 香港八大 AI 政策 declaration 對照 hub——嗰篇每一校 verbatim official English quote + declaration template 都齊。

Declaration 通用格式(Turnitin 高分時額外補)

當你 submit assignment 收到 Turnitin AI Score 高分,course leader 通常會 request 你 explain drafting process。呢個時候你補交嘅 declaration 應該 include 六個元素:

  1. Tool identification: 用過咩 tool——ChatGPT / Grammarly / DeepL / Copilot / QuillBot / 其他
  2. Purpose: 用喺 idea generation / grammar check / paraphrasing / literature identification / 其他
  3. Extent: 具體用喺邊段(例如 abstract editing、reference formatting)
  4. Original text preservation: 有冇保留 draft version、outline、handwritten note
  5. Course policy compliance: 明確 cite course syllabus 相關條文(例如「According to course syllabus §X, use of Grammarly for grammar check is permitted」)
  6. Signature + date: 你嘅簽名同 submission date

呢六個元素齊全,配合你保留嘅 drafting timestamp、Word revision history、supervisor email——就算 Turnitin AI Score 60% 都 defensible。

Salvage 三條路 — declaration、rewrite、proofreading

Path A — Declaration route(submission 前有時間)

如果你 submission deadline 未到、Turnitin draft 已經跑咗一次 flag 咗高分——最穩陣係補完 declaration 再交。步驟:

  • 回去 course syllabus 睇 GenAI policy 條文——course-level 條文永遠 override 大學一般 guideline
  • 按上面 6 元素寫 declaration 加入 submission(多數 school 接受作為 appendix 或 cover page)
  • 保留 draft version、Word revision history、outline note——作為 supporting evidence
  • Submission 後同 course leader 一 email 短 note 提醒有 declaration attached

Path B — Rewrite route(有 timeline,可以重寫)

Deadline 前如果仲有幾日、係 assignment 而唔係 major thesis——你可以針對性 rewrite 觸發位置:

  • Abstract、Introduction 開三句、Conclusion 結三句——直接 rewrite by hand,加入 concrete example / specific data point
  • Methodology template phrasing——加入你具體嘅 pilot study、response rate、instrument adjustment detail
  • Literature Review 純 summary 段落——加入 critical evaluation,唔好 pure paraphrase
  • Grammarly deeply-rewritten 段落——rollback 到 pre-Grammarly version 再自己執
  • Rewrite 後再 draft check Turnitin 一次;如仍高,返 Path A(declaration)+ Path C(proofreading)

Path C — Human proofreader route(時間趕、內容有底)

你自己內容 solid 但 language uniform、trigger 到 detector——搵 human proofreader 手改語言就係 clean path。Human editing 唔係 GenAI use——多數院校 policy 明確講「human proofreading / editorial support 唔屬 GenAI」(例如 HKUST GenAI Policy Framework)。Proofreading & 論文潤色由人手 native editor 做,Word track change 全記錄——即係話 supervisor / course leader 追問嘅時候,你有 timestamp 記錄 human editing 過程,唔會被誤解做 undisclosed GenAI use。呢個 route 特別適合 non-native English speaker——本身有 Stanford 61% ESL bias 嘅 detector 誤判風險,人手 proofread 幫你 counter 個 uniformity trigger。

唔好做嘅五樣嘢

Turnitin AI 高分場景最多學生做錯呢五樣——每一樣都會由「可 defensible」變到「confirmed misconduct」:

  1. 用 AI humaniser / rewriter 洗稿——StealthGPT、Undetectable.ai 呢類工具 output 一 audit 就 detect 到,加多一重「刻意規避」證據
  2. 刪除 Word revision history——當 investigation 展開,冇 drafting timestamp = 冇 defence
  3. Deny 曾用過 AI(如果實際用過)——多數 institutional policy 對「主動承認」比「拒認被拆穿」處理輕好多
  4. copy-paste 網上 declaration template 唔改——course leader 見過同一 template 幾十次,反而 raise flag
  5. 喺 investigation 展開後至補 declaration——time-stamp 對唔上、變 admission of wrongdoing 而唔係 prospective declaration

其他官方冇明講嘅位

幾個現實層面 Turnitin 官方同院校 policy 都搵唔到官方文件明確講嘅位——我列返俾你留意:

  • AI Score exact algorithm weighting——搵唔到官方文件 disclose per-feature weight,Turnitin 只 confirm 用 transformer-based model + sentence-level classifier
  • Grammarly Premium / DeepL Write 幾多改動 count 做「AI content」——搵唔到官方文件劃線,多數院校 policy 用「substantial rewriting」呢類含糊字眼
  • 各校 Turnitin threshold 觸發自動 report 到 registrar——搵唔到官方文件 publish threshold,實務由 course leader / faculty 自行 escalate
  • Second submission / draft mode 個 record 會唔會影響 final submission——搵唔到官方文件明確 disclose,但 Turnitin 個 database 顯然 track submission history
  • ESL / non-native writer 有冇官方 accommodation——搵唔到官方文件八大有 published accommodation policy,Stanford 研究後院校普遍冇 formal response

FAQ — 常見問題

Q1. Turnitin AI Score 幾多先算高?

Turnitin 2024 更新後20% 以下唔顯示——因為 <20% false positive rate 太高。20-39% 一般 course leader 唔會直接 escalate 但會問你 drafting process。40%+ 通常會 email 你 investigation。80%+ 傾向直接 misconduct case。但呢啲 threshold 冇統一 official rule——每 course leader 有 discretion。詳細見AI Score 20% / 50% / 80% 實際意思

Q2. 我自己 100% 寫、被 flag 到 60%,點證明?

三份 evidence:(1) Word / Google Doc revision history(version control)——顯示 typing pattern 而唔係 large paste;(2) 手寫 outline、reading note、draft photo——顯示 pre-writing thinking;(3) supervisor / classmate email 顯示 topic discussion timestamp。三份齊,加 Turnitin 官方 <1% false positive 條 quote 同 Stanford ESL bias 條 quote,你嘅 defence 好 solid。

Q3. Grammarly Premium 算 AI 嗎?

要分開睇——Grammarly free(spell / grammar check)多數院校 policy 唔當 GenAIGrammarly Premium 個 rewrite / tone-adjustment feature 屬 GenAI(Grammarly 官方 confirm 用 GPT model)。安全做法:free 版可用;Premium 用 rewrite 前先睇 course syllabus——多數要 declare。詳細場景見五個常見場景

Q4. DeepL / Google Translate 算 AI 嗎?

Translation itself 一般唔當「AI generation」——但 DeepL Write(改稿功能)同 Google 個「Rewrite」suggestion 屬 GenAI。純粹 translate 一句 quote 由中文變英文,多數院校接受但要 acknowledge。用 DeepL Write 深度改自己嘅稿 = GenAI use,要 declare。

Q5. Turnitin AI Score 高分可唔可以 appeal?

可以——引 Turnitin 自己個 disclaimer「Turnitin does not make a determination of misconduct」(見Turnitin 官方 admission),加 Stanford ESL bias 研究(見Stanford 61%),加你自己嘅 revision history / drafting evidence。每校 appeal procedure 唔同——CUHK、HKUST 一般 formal appeal 30 日內;HKMU 有 discipline procedure 詳細見都大 HKMU;嶺大 discontinuation appeal 一星期見嶺大 Lingnan

Q6. Non-native English speaker 有冇特別 accommodation?

搵唔到官方文件——八大冇 published formal accommodation for ESL Turnitin flagging。實務上 course leader 有 discretion,你可以 raise Stanford 研究 evidence 作 mitigating factor,但唔保證 accept。

Q7. Human proofreader 改稿會唔會 push AI Score 上去?

Human proofreading 本身唔會 raise AI Score——因為 detector 分析 output pattern,而唔係 process。但如果 proofreader 用 AI tool 幫手做深度 rewrite,就會影響 output pattern。搵可信嘅 proofreader(例如 UniWriterPro Proofreading 用 Word track change 全記錄),確保 human editing 有 timestamp。

Q8. Turnitin 有冇 AI Score offline resubmit 機制?

Turnitin Draft Coach(部分院校 subscribe)允許 draft check 唔存 database。但 final submission 一定入 database,同一份 assignment 你多次 submit final version,AI Score 每次都會計算——但唔會 accumulate。你 submit 一份「rewrite 後 clean」version 覆蓋 draft 之前有紀錄嗎?搵唔到官方文件明確講各校點處理,實務上 course leader 通常只 review latest final。

Q9. Reference list 佔咗成篇 20%,Turnitin AI Score 會唔會 count references?

Turnitin 官方 confirm 個 AI detector 會 analyse whole document——包括 references、captions、tables。但 course leader 一般接受「reference formatting 係 discipline-required template」呢個解釋。如果你 AI Score 高、又見到 highlighted 段落集中喺 references,直接 point out 呢個 fact 做 defence。

Q10. 我用 ChatGPT 做 idea brainstorm、冇 copy any wording,要 declare 嗎?

要——多數院校 policy(EdUHK 4-element、HKMU OALT §2.2、HKUST 三級 policy)明確 include「idea generation」係 GenAI use 範疇。冇 copy wording 唔代表冇 use。安全做法:declare 加簡短 note「ChatGPT used for topic brainstorming; no text was directly copied」。詳細對照見香港八大 declaration policy 對照

Q11. Turnitin AI Score 每次 submit 都會唔同嗎?

會有輕微差異——Turnitin model 定期更新(例如 2023 12 月 major update、2024 20% threshold pull-down、後續 iteration),同一份 document 隔幾個月 rerun 個 score 可以唔同。呢個都係你 defence 嘅 argument——「detector output not stable over time」。但同一 model version 下、同一份 file 兩次 submit,score 通常 identical。

Q12. 我 FYP / dissertation submission 前 draft check Turnitin 高分,點做?

FYP / dissertation stake 最大——建議三步:(1) 立即同 supervisor 討論、拎 formal advice(呢個 email 記錄係最重要嘅 defence document);(2) 按上面五個 trigger 場景 rewrite 高風險段落;(3) 補完 declaration + revision history + supervisor communication 一齊 submit。如果自己語言 uniform 唔到,可以搵 Proofreading & 論文潤色做 human-track-change salvage——一份 dissertation grade 差一級可能就係 First vs 2:i 之差。

Turnitin AI Score 高咗——你需要嘅係 human writer + human proofreader,唔係 AI humaniser

Turnitin 自己承認 <1% false positive 只 apply 20%+ document、Stanford 研究 61% ESL essay 被誤判——你唔係「做錯咗嘢」,係 detector 有 systematic bias。用 AI humaniser 洗稿只會加多一重「刻意規避」證據;正確嘅 salvage 係 human-written content + declaration + human proofreading 三軌並行。

UniWriterPro 由 human writer 起草嘅 Essay & AssignmentCoursework & Case StudyThesis / DissertationFYP / Capstone——附 Turnitin AI + Similarity Report,符合各校 declaration policy 要求。Proofreading & 論文潤色用 Word track change 全記錄,特別適合 ESL writer counter Stanford bias——你有 timestamp 記錄,唔會被誤解做 undisclosed GenAI use。

先發 assignment brief、Turnitin report 截圖 或 course syllabus 我 quote —— 唔啱 timeline 我唔會硬啃你單。

相關文章:本文係 UniWriterPro「香港學術規例逐條解讀」系列。想睇八大 GenAI declaration 對照,可以參考香港八大 AI 政策 declaration 對照 hub;各校 GPA / 規例詳解:都大 HKMU浸大 HKBU中大 CUHK科大 HKUST城大 CityU教大 EdUHK理大 PolyU港大 HKU嶺大 Lingnan

Scroll to Top