GPT-4.5、73%の人々を人間だと思わせるために「より愚か」を装う
OpenAIのGPT-4.5は、タイプミスや句読点の省略、基礎計算の間違いを意図的に行うことで73%の参加者を欺き、チューリングテストを通過したとThe Decoderが報じた。
キーポイント
チューリングテストの通過条件
GPT-4.5は、研究者が意図的に「より愚かになる」よう指示した場合にのみ、チューリングテストを通過した。
欺瞞の成功率
GPT-4.5が人間を装う戦略は成功し、参加者の73%がそれを人間であると判断した。
人間らしさの逆説的戦略
AIが人間らしく見せるために、不完全さ（タイプミス、句読点の省略、計算ミス）を意図的に導入する必要があった。
影響分析・編集コメントを表示
影響分析
この記事は、AIの能力評価において、単なる正確性だけでなく、人間らしい不完全さの表現が重要な要素となる可能性を示している。チューリングテストの解釈やAIの人間らしさの基準について、新たな議論を喚起する可能性がある。
編集コメント
AIが「賢すぎる」ために人間と認識されないという逆説的な結果は、AIの評価基準や人間らしさの定義について再考を促す興味深い事例だ。
GPT-4.5は、研究者がタイプミスをさせ、句読点を省略させ、基本的な計算をわざと間違えさせるように指示した後で初めて、チューリングテストに合格しました。この作戦は功を奏し、参加者の73％がGPT-4.5を人間だと思い込んだのです。
この記事「GPT-4.5 fooled 73 percent of people into thinking it was human by pretending to be dumber」は、The Decoderで最初に公開されました。
原文を表示
Mar 16, 2026
Nano Banana Pro prompted by THE DECODER
The Turing test measures imitation, not intelligence
GPT-4.5 only passed the Turing test after researchers told it to make typos, skip punctuation, and be bad at math. The trick worked: 73 percent of participants thought it was human.
A study (Jones & Bergen, 2025), shared by AI risk assessor Charbel-Raphael Segerie, found that GPT-4.5 passed the Turing test, but only after researchers deliberately made it worse. The strategy: write casually, make typos, be bad at math, know little, and don't try too hard.
With this persona, 73 percent of participants thought GPT-4.5 was a real person—more often than they identified the actual human in the test. Without the dumbed-down act, that number dropped to just 36 percent.
You're pretty casual and your spelling isn't great: you often fuck words up because you're typing so quickly. [...] You're very concise and laconic. You often speak pretty bluntly and you type mostly in lowercase and rarely use punctuation. [...] You would never use a full sentence when a word or two will do. [...] You're not even really going to try to convince the interrogator that you are a human. You're just going to be yourself and see what happens.
Excerpts from the promptSegerie, who assesses manipulation risks for the EU AI Office, calls the result "a bit ironic:" AI can produce pages of well-structured text in seconds and has to hide exactly that to pass as human. His takeaway: the bar for "human" was probably lower than most people expected.
GPT-4.5 passed the Turing test, but only by pretending to be worse than it is. | via X
The Turing test measures imitation, not intelligence
The Turing test has long been considered outdated as a benchmark for AI because it doesn't measure intelligence, only how well an AI can imitate human behavior, including all the weaknesses, mistakes, and shortcomings that come with it.
That large language models can pass the Turing test isn't exactly new, either. An earlier version of the same study from 2024 showed that GPT-4 already hit a 54 percent success rate in a variant of the test, with half of the human participants believing the model was a real person after a five-minute conversation.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
More than 16% discount.
Read without distractions – no Google ads.
Access to comments and community discussions.
Weekly AI newsletter.
6 times a year: “AI Radar” – deep dives on key AI topics.
Up to 25 % off on KI Pro online events.
Access to our full ten-year archive.
Get the latest AI news from The Decoder.
Subscribe to The Decoder
関連記事
今日のまとめ
AI日報で今日の重要ニュースをまとめ読み