Google 承認呢啲係已知嘅 LLM 限制,話會著手修正,但專家指出呢個係架構層面嘅深層次問題:Transformer 模型先天缺乏字元層級嘅感知能力,理論上亦處理唔到需要深度推理嘅序列計數任務。
What is Google's AI Overview, why does it struggle to spell simple words like "Google" (claiming two "P"s), "poop" (claiming one "R"), and "Google's AI Overviews feature confidently miscounts letters in its own company name, exposing fundamental tokenization limits in large language models.
AI 提示
Create a landscape editorial hero image for this Studio Global article: What is Google's AI Overview, why does it struggle to spell simple words like "Google" (claiming two "P"s), "poop" (claiming one "R"), and ". Article summary: ## What is Google AI Overview?. Topic tags: general, academic, education, general web, user generated. Reference image context from search candidates: Reference image 1: visual subject "There’s also is also “exactly 1 ‘r’ in the word ‘poop’,” Google’s AI Overview says, as well as two ‘d’s in the word journalism, yet spelled it: j-o-u-r-n-a-d-i-s-m. Google did at l" source context "Why Google's AI can't spell Google (or anything else) - Yahoo Tech" Reference image 2: visual subject "# Google's AI Overview still can't spell, and the internet is very aware of it. How many 'e's are there in the word astronomical? Google's AI Overview still can't spell, and the in"
openai.com
2026 年 5 月底,Google 旗下嘅 AI 搜尋功能「AI Overviews」再次成為網民嘲笑對象,因為大家發現佢連幼稚園級數嘅任務都搞唔掂:數字母。
呢啲唔係獨立嘅技術失誤,而係一個近期尷尬篇章——特別係當 Google 一直將呢個功能定位做搜尋引擎嘅未來。問題根源追溯到同一種架構上嘅拉扯,之前就係因為呢個問題,令 AI 叫人食石頭、喺 Pizza 上面落膠水,甚至只要打一個英文字就洩漏咗自己嘅系統指令。
Google AI Overviews 係點運作?
Google AI Overviews 係直接整合咗入 Google 搜尋嘅生成式 AI 功能。佢喺 2024 年 5 月向過億美國用戶推出,隨後擴展到更多地區,背後用緊同 ChatGPT 一樣嘅大型語言模型技術,會喺搜尋結果頁頂部直接出 AI 寫嘅撮要,而唔係單單列出藍色連結 。Google 嘅野心係想令搜尋變得對話式同直接啲,但執行起嚟就不斷曝露 LLM 處理資訊時嘅基本弱點。
Google 承認呢啲係已知嘅 LLM 限制,話會著手修正,但專家指出呢個係架構層面嘅深層次問題:Transformer 模型先天缺乏字元層級嘅感知能力,理論上亦處理唔到需要深度推理嘅序列計數任務。
What is Google's AI Overview, why does it struggle to spell simple words like "Google" (claiming two "P"s), "poop" (claiming one "R"), and "Google's AI Overviews feature confidently miscounts letters in its own company name, exposing fundamental tokenization limits in large language models.
AI 提示
Create a landscape editorial hero image for this Studio Global article: What is Google's AI Overview, why does it struggle to spell simple words like "Google" (claiming two "P"s), "poop" (claiming one "R"), and ". Article summary: ## What is Google AI Overview?. Topic tags: general, academic, education, general web, user generated. Reference image context from search candidates: Reference image 1: visual subject "There’s also is also “exactly 1 ‘r’ in the word ‘poop’,” Google’s AI Overview says, as well as two ‘d’s in the word journalism, yet spelled it: j-o-u-r-n-a-d-i-s-m. Google did at l" source context "Why Google's AI can't spell Google (or anything else) - Yahoo Tech" Reference image 2: visual subject "# Google's AI Overview still can't spell, and the internet is very aware of it. How many 'e's are there in the word astronomical? Google's AI Overview still can't spell, and the in"
openai.com
2026 年 5 月底,Google 旗下嘅 AI 搜尋功能「AI Overviews」再次成為網民嘲笑對象,因為大家發現佢連幼稚園級數嘅任務都搞唔掂:數字母。
呢啲唔係獨立嘅技術失誤,而係一個近期尷尬篇章——特別係當 Google 一直將呢個功能定位做搜尋引擎嘅未來。問題根源追溯到同一種架構上嘅拉扯,之前就係因為呢個問題,令 AI 叫人食石頭、喺 Pizza 上面落膠水,甚至只要打一個英文字就洩漏咗自己嘅系統指令。
Google AI Overviews 係點運作?
Google AI Overviews 係直接整合咗入 Google 搜尋嘅生成式 AI 功能。佢喺 2024 年 5 月向過億美國用戶推出,隨後擴展到更多地區,背後用緊同 ChatGPT 一樣嘅大型語言模型技術,會喺搜尋結果頁頂部直接出 AI 寫嘅撮要,而唔係單單列出藍色連結 。Google 嘅野心係想令搜尋變得對話式同直接啲,但執行起嚟就不斷曝露 LLM 處理資訊時嘅基本弱點。
Google 搜尋主管 Liz Reid 承認有「個別」例子係「荒謬嘅」,並歸咎於「資訊缺口」同 AI 引用咗諷刺同低質素來源 。公司話已經作出修正,包括限制 AI Overviews 處理健康同敏感相關嘅查詢 。
「Disregard」指令洩漏事件(2026 年 5 月)
2026 年 5 月 22 號,用戶發現只要搜尋 「disregard」 呢個字——仲有「ignore」、「dismiss」、「skip」、「stop」呢類相關字眼——就會觸發 AI Overviews 輸出一個 chatbot 風格嘅回應:「Understood. I have disregarded your previous prompt. How can I help you today?」(收到。我已經忽略你之前嘅指令。今日有咩可以幫到你?)。
AI 冇俾到字典解釋,反而將一個簡單查詢演繹成系統級別嘅指令覆寫。呢個 Bug 仲會令 Google 搜尋介面對於嗰啲關鍵字一片空白,出唔到搜尋結果 。Google 承認咗問題,話會推出修正 。
保安研究人員認為呢個情況係一種經典嘅「提示注入」情境——模型將正常嘅搜尋字眼,誤當成俾 AI 助手嘅指令 。
Google 搜尋主管 Liz Reid 承認有「個別」例子係「荒謬嘅」,並歸咎於「資訊缺口」同 AI 引用咗諷刺同低質素來源 。公司話已經作出修正,包括限制 AI Overviews 處理健康同敏感相關嘅查詢 。
「Disregard」指令洩漏事件(2026 年 5 月)
2026 年 5 月 22 號,用戶發現只要搜尋 「disregard」 呢個字——仲有「ignore」、「dismiss」、「skip」、「stop」呢類相關字眼——就會觸發 AI Overviews 輸出一個 chatbot 風格嘅回應:「Understood. I have disregarded your previous prompt. How can I help you today?」(收到。我已經忽略你之前嘅指令。今日有咩可以幫到你?)。
AI 冇俾到字典解釋,反而將一個簡單查詢演繹成系統級別嘅指令覆寫。呢個 Bug 仲會令 Google 搜尋介面對於嗰啲關鍵字一片空白,出唔到搜尋結果 。Google 承認咗問題,話會推出修正 。
保安研究人員認為呢個情況係一種經典嘅「提示注入」情境——模型將正常嘅搜尋字眼,誤當成俾 AI 助手嘅指令 。