今天用 AI 來拷問我的論文,自己攻擊自己,包括內容與形式,例如論點、論據、論證、註腳及參考書目等。
如果用 Grammarly 等,可以檢查錯別字、文法、文句的表達等,效果是經打磨後的句子比較緊湊,減少低級的寫作錯誤。但他的記憶力有限,每次只能讀10頁左右,但又不能指定頁數,所以根本沒有方法作系統的檢查。隨寫隨改他可以做得到,但這樣很打擾我的思路,所以我一般是寫好了才叫他審核,但就出現了認憶力不足的問題。
AI 的好處是記憶力強大,我可以一次過餵全篇論文給他讀。因為可以記住大量資料,有些我靠自己很難發現的問題它都找出來了,例如隔了幾十頁的說法、資料、引文、翻譯等的不一致;又如資料矛盾,同一年同一職位有兩個人名[我再查,原來是一個署任一個正印]。重點是,他隔了幾十頁都找到。又例如,他懂得看我的參考書目有沒有按字母排序、對我的某些註腳卷數跟前後的註腳不同叫我覆查一下等,都已超出拼字文字檢查的能力了。
最重要的,也是 Claude 比較突出的,是他敢說真話,例如有些地方他認為我話說得過了頭,又或者表達含混不清楚等。我仔細一看,覺得它批評得對,要修訂自己的說法,或者要講得審慎些。
人會疲倦,AI 不用休息,不過 AI 的表現也受客觀環境影響,每次的回答都不一樣,所以我打算再問他幾次。
我把論文儲存為PDF,每章一個,然後逐章問他,最後再把全部PDF一次餵給他,要他全本看。每章問他的答案要仔細些,全本的話,他可以參照全文指出個別不妥之處。我起初只是叫他看看我文句有無不通,論證有無犯駁,問了兩章後我問他你究竟用甚麼準則來評鑑我的論文,他和盤托出,於是我把這些記下成為我的校對 prompt,效果不錯。
Check my thesis according to the followings
1. Cross-reference consistency (dates, names, figures)
Birth/death dates matching across all mentions of the same person
Event dates consistent wherever repeated (e.g., a death year stated once but differently later)
Numeric data (counts, statistics, page/volume/document totals) matching when the same figure is cited more than once in different sections
2. Internal contradictions in argument logic
Checking whether a claim in one sentence is undercut or reversed by the conclusion drawn in an adjacent sentence (e.g., “these books might be useful” followed by “thus he felt the need to compile a new one” — a logical mismatch)
Checking whether stated comparisons/contrasts actually support the conclusion being drawn from them
3. Timeline/chronology conflicts
Checking narrative dates against any summary tables or lists (e.g., parish posting dates in a table vs. dates mentioned in surrounding prose)
Flagging unexplained gaps or overlaps in a person’s timeline
4. Terminology and categorization consistency
Same Chinese term romanized/glossed consistently throughout (e.g., jiao’an vs. jiaowu apostrophe usage)
Same concept given consistent English translation across mentions
Categorization schemes (e.g., “three types of X”) checked against whether the text actually delivers the stated number of types
5. Formatting/conversion artifacts
Flagging tables, figures, or diagrams that appear to have been lost or garbled in plain-text conversion
Flagging placeholder markers (dashes, blank figure references) that need original formatting restored
Check footnotes entry against bibliography entry whether they match
6. Redundancy across sections
Same quotation, fact, or claim repeated multiple times within a short span without cross-referencing, which could be tightened
7. Stray typos/artifacts
Leftover formatting characters (stray asterisks, missing spaces, etc.)
8. Unreconciled competing claims from sources
Cases where two cited scholars or sources give different figures/interpretations and the text doesn’t flag the discrepancy for the reader
我用了 Claude 的 Incognito 模式,他聲稱不會保留記憶或抄考用來訓練大模型。信得過嗎? 見仁見智,反正我信了。他前後用了10分鐘,找出不少毛病。因為我預計免費時段不夠用,所以付費了,紐紙47.14一個月。趁可以用要多用,因為 Anthropic 敵視中國(當然包括自古以來神聖不可分割但遭英國非法割佔實施殖民統治但從來不是殖民地的現已恢復行使主權的中國香港),不但不准用,更主動出擊檢查,主動封號。中國香港IP一律封殺,VPN封殺,不接受中國香港信用卡及電話。在我的情況而言,DeepSeek等國產模型不及Claude,還是支持一下吧。
我覺得 AI 是個自學良師,但一定要跟他駁嘴,否則他不會盡展所長。禮記云:「善待問者如撞鐘,叩之以小則小鳴,叩之以大則大鳴」。你要 push back, AI 才看得起你跟你切磋。小叩小鳴,大叩大鳴,此之謂也。如果學生不敢問,照單全收,教師不敢答,照本宣科,大家只是扮學、扮教,雖然安全,但不會進步。














