“說的都不一樣”“我真的不知道該信哪個了。”演員劉美含的吐槽,讓AI答案的準確性問題再次引發全網熱議。
A Chinese actress has sparked online debate over the reliability of artificial intelligence after receiving conflicting answers from several AI tools about the pronunciation of a Chinese character.
![]()
女演員社交平臺視頻截圖
“坊”字怎么讀?
5款AI工具答案不一
近日,劉美含在社交平臺發布了一則視頻,其在視頻中表示:“我今天去配音,有一句臺詞是‘鑄幣坊’,我說的是‘鑄幣fǎng’。然后我突然覺得有點不是很確定這個讀音,就說我們查一下怎么讀。”
然而查詢的結果卻讓她“越查越迷糊”。
她先后使用百度AI、DeepSeek、元寶、豆包以及通義千問等工具進行搜索,其中百度AI給出的答案是fáng(二聲),DeepSeek和元寶則顯示為fāng(一聲),而她與錄音老師分別用豆包查詢時,竟然得到了不同的讀音結果。
Liu Meihan said in a social media video that the issue arose while she was dubbing a line containing the phrase "
zhubifang", which translates as "minting workshop" in English. Unsure how to pronounce the character "fang", she consulted several AI assistants.
However, the tools — including Baidu AI, DeepSeek, Yuanbao, Doubao and Tongyi Qianwen — gave different answers. Some suggested the pronunciation to be fáng (second tone), while others returned fāng (first tone).
隨后她又詢問通義千問,得到的依然是fāng(一聲)。最終,還是經紀人通過《新華詞典》APP查詢,才確定了“鑄幣坊”中“坊”字的正確讀音為fáng。
She added that when she and a recording instructor checked the same query on Doubao using different devices, they received different pronunciations. The correct reading was later confirmed by her manager through the Xinhua Dictionary app as fáng.
事實上,從語言規范角度看,“坊”字的讀音并不存在實質爭議。南京師范大學文學院教授化振紅表示,“坊”字自古只有一聲fāng和二聲fáng,無三聲讀法,三聲多為方言影響誤讀。
現代漢語中,一聲表示街巷、店鋪、牌坊等;二聲指小手工業生產場所,如磨坊、染坊。“鑄幣坊”為造幣場所,正確讀音為fáng(二聲)。
From a linguistic perspective, scholars say there is little ambiguity. Hua Zhenhong, a professor at Nanjing Normal University, said the character has two standard pronunciations, fāng and fáng. In the context of a minting workshop, the correct reading is fáng.
為了驗證相關情況,南都大數據研究院記者再次使用劉美含視頻中提及的幾款AI工具搜索“鑄幣坊讀音”。結果顯示,不同平臺之間仍存在明顯分歧:DeepSeek、元寶給出的讀音依然為fāng(一聲),而豆包與通義千問的回答已顯示為fáng(二聲),百度AI則保持原本的fáng(二聲)答案不變。
AI為何“各說各話”?
生成式人工智能的概率邏輯
同一個字的讀音,為何多款AI工具卻給出了不同的答案?原因在于生成式人工智能的底層運行邏輯。當前主流大模型并非基于權威知識庫逐條檢索標準答案,而是通過對海量文本語料進行訓練,基于概率預測生成“最可能”的表達。它并不具備對事實真偽的判斷能力,也無法像人類一樣進行來源核實。
當訓練語料中存在多種解釋、歷史讀音、網絡討論甚至錯誤示例時,模型就可能在不同語境下生成不同結果。此外,不同平臺的訓練數據來源、模型架構、推理策略存在差異,部分模型側重通用語料,部分則加強檢索能力或知識庫對接,這些技術路徑的差別都會影響輸出結果。再加上模型版本持續更新,同一平臺在不同時間給出的回答也可能發生變化。
Experts say the inconsistencies stem from how generative AI works. Rather than retrieving fixed answers from authoritative databases, large language models generate responses based on probabilities learned from large text datasets, which may contain different interpretations or errors.
從公眾使用角度看,專家建議保持理性態度。蘇州大學科技傳播研究中心主任王國燕指出,AI生成內容往往語言流暢,但未必完全準確。面對重要信息時,應養成多查證、慢判斷的習慣,避免對AI產生過度依賴。
Experts therefore advise users to verify important information through multiple sources rather than relying solely on AI-generated content, said Wang Guoyan from Soochow University.
來源:南方都市報
跟著China Daily
精讀英語新聞
“無痛”學英語,每天20分鐘就夠!
特別聲明:以上內容(如有圖片或視頻亦包括在內)為自媒體平臺“網易號”用戶上傳并發布,本平臺僅提供信息存儲服務。
Notice: The content above (including the pictures and videos if any) is uploaded and posted by a user of NetEase Hao, which is a social media platform and only provides information storage services.