一项经典的大脑测试暴露了人工智能最大的弱点

* Draft: Suketu Patel 领导的研究人员让几个领先的 AI 模型进行了一项著名的心理学实验，称为 Stroop 任务。 * Refinement: "Stroop task" is often translated as "Stroop 任务" or "斯特鲁普任务". I will use "Stroop 任务" as it's common in tech/science contexts, or "斯特鲁普任务" for a more formal psychology tone. Let's go with "Stroop 任务" for clarity or "斯特鲁普任务". Let's use "斯特鲁普任务" for the first mention and maybe "Stroop 任务" later, or just stick to the Chinese name. "Stroop 任务" is very common in Chinese AI papers. * "The results revealed a significant difference between how AI systems process information and how the human brain manages attention." * Draft: 结果揭示了 AI 系统处理信息的方式与人脑管理注意力的方式之间存在显著差异。 * *Para 2:* "What Is the Stroop Task?" -> 什么是 Stroop 任务？ (or 斯特鲁普任务) * *Para 3:* "The Stroop task is a classic psychological test that has been used for decades to study attention, concentration, and self-control." * Draft: Stroop 任务是一项经典的心理学测试，几十年来一直用于研究注意力、专注力和自控力。 * *Para 4:* "In the test, color words such as 'red,' 'blue,' or 'green' are displayed in colored ink. Sometimes the word and the ink color match. For example, the word 'red' might appear in red ink. Other times they conflict, such as the word 'red' printed in blue ink." * Draft: 在测试中，颜色词如“红”、“蓝”或“绿”以彩色墨水显示。有时单词与墨水颜色匹配。例如，单词“红”可能以红色墨水出现。其他时候它们会发生冲突，例如单词“红”以蓝色墨水打印。 * *Para 5:* "Participants are asked to name the color of the ink rather than read the word itself." * Draft: 参与者被要求说出墨水的颜色，而不是阅读单词本身。 * *Para 6:* "That sounds simple, but it creates a challenge because reading words is an automatic habit for most people. The brain must suppress the urge to read the word and instead focus on identifying the ink color." * Draft: 这听起来很简单，但它制造了一个挑战，因为阅读单词是大多数人的自动习惯。大脑必须抑制阅读单词的冲动，转而专注于识别墨水颜色。 * *Para 7:* "Psychologists often use the task to measure what is known as executive control, a set of mental processes that helps people regulate attention, resist distractions, and stay focused on goals." * Draft: 心理学家经常使用这项任务来测量所谓的执行控制，这是一组帮助人们调节注意力、抵抗干扰并保持专注于目标的心理过程。 * *Para 8:* "Testing AI Attention" -> 测试 AI 的注意力 * *Para 9:* "The researchers wanted to see whether modern large language models (LLMs) handle this challenge in the same way humans do." * Draft: 研究人员想看看现代大型语言模型是否像人类一样处理这一挑战。 * *Para 10:* "LLMs are the AI systems behind tools such as ChatGPT, Claude, and Gemini. They are trained on enormous amounts of text and learn patterns in language, allowing them to generate responses that often appear remarkably human." * Draft: LLM 是 ChatGPT、Claude 和 Gemini 等工具背后的 AI 系统。它们在海量文本上进行训练，学习语言模式，从而能够生成通常看起来非常人性化的回答。 * *Para 11:* "When given short lists containing five color words, the AI systems generally performed well, even when the words and colors did not match." * Draft: 当给出包含五个颜色词的短列表时，AI 系统通常表现良好，即使单词和颜色不匹配。 * *Para 12:* "However, the picture changed dramatically as the lists became longer." * Draft: 然而，随着列表变长，情况发生了巨大变化。 * *Para 13:* "GPT-4o achieved 91% accuracy when working with five words. At ten words, its accuracy fell to 57%. When the list expanded to forty words, accuracy dropped to just 15%." * Draft: GPT-4o 在处理五个单词时达到了 91% 的准确率。在十个单词时，其准确率降至 57%。当列表扩展到四十个单词时，准确率仅降至 15%。 * *Para 14:* "Claude 3.5 Sonnet maintained stable performance through lists of twenty words but then experienced a sharp decline, falling to 24% accuracy with forty-word lists." * Draft: Claude 3.5 Sonnet 在二十个单词的列表中保持了稳定的性能，但随后经历了急剧下降，在四十个单词的列表中准确率降至 24%。 * *Para 15:* "The researchers observed similar patterns in GPT-5, Claude Opus 4.1, and Gemini 2.5." * Draft: 研究人员在 GPT-5、Claude Opus 4.1 和 Gemini 2.5 中观察到了类似的模式。 * *注明:* Again, strictly translating the model names provided in the source text. * *Para 16:* "When AI Loses Focus" -> 当 AI 失去焦点 / 当 AI 注意力涣散时 * *Para 17:* "The challenge became even more difficult when matching and mismatched color words appeared together in the same list." * Draft: 当匹配和不匹配的颜色词同时出现在同一个列表中时，挑战变得更加困难。 * *Para 18:* "Under those conditions, performance deteriorated further. Accuracy for the mismatched items dropped to nearly zero in some cases." * Draft: 在这些条件下，性能进一步恶化。在某些情况下，不匹配项目的准确率降至接近零。 * *Para 19:* "According to the researchers, the AI models had trouble maintaining the instruction to identify ink colors. Instead, they increasingly defaulted to reading the words themselves." * Draft: 据研究人员称，AI 模型难以维持识别墨水颜色的指令。相反，它们越来越倾向于默认阅读单词本身。 * *Para 20:* "In other words, the systems appeared unable to consistently suppress the response they had been most heavily trained to produce." * * *Literal:* 这一发现特别有趣，因为人类面临类似的冲突。 * *Refinement:* "Particularly interesting" -> 尤为有趣/特别有趣. "Face a similar conflict" -> 面临着类似的冲突. * *Draft:* 这一发现尤为有趣，因为人类也面临着类似的冲突。 * *Sentence 2:* "People are generally much better at reading words than naming ink colors." * *Context:* This refers to the Stroop effect. * *Literal:* 人们通常在阅读文字方面比命名墨水颜色要好得多。 * *Refinement:* "Reading words" -> 阅读文字/单词. "Naming ink colors" -> 命名墨水颜色/说出墨水颜色. "Much better at" -> 擅长得多. * *Draft:* 人们通常阅读文字的能力要远强于命名墨水颜色的能力。 * *Sentence 3:* "Yet despite this bias, most individuals can maintain high accuracy and stable performance even when confronted with long lists of conflicting words and colors." * *Literal:* 然而尽管有这种偏见，大多数人可以保持高准确率和稳定的表现，即使面对长长的冲突单词和颜色列表。 * *Refinement:* "Bias" here refers to the cognitive tendency (reading > naming). "Confronted with" -> 面对. "Long lists" -> 长列表/冗长的列表. "Conflicting words and colors" -> 冲突的文字和颜色. * *Draft:* 然而，尽管存在这种偏向，大多数人在面对冗长的文字与颜色冲突列表时，仍能保持高准确率和稳定的性能。 * *Tag:* `

 

Human Attention vs. Machine Attention

` * *Translation:* `

 

人类注意力与机器注意力

` * *Sentence 4:* `

The study highlights an important distinction between human and artificial intelligence.

` * *Literal:* 该研究强调了人类和人工智能之间的一个重要区别。 * *Refinement:* "Highlights" -> 凸显/强调. "Distinction" -> 区别/差异. * *Draft:* `

这项研究凸显了人类智能与人工智能之间的一个重要差异。

` * *Sentence 5:* `

Although modern AI systems can produce impressive language and reasoning capabilities, their underlying mechanisms differ from the attention processes found in biological brains.

` * *Literal:* 虽然现代AI系统可以产生令人印象深刻的语言和推理能力，但它们的底层机制不同于生物大脑中发现的注意力过程。 * *Refinement:* "Produce" -> 展现/产生. "Underlying mechanisms" -> 底层机制. "Attention processes" -> 注意力过程. "Biological brains" -> 生物大脑. * *Draft:* `

尽管现代人工智能系统能够展现出令人印象深刻的语言和推理能力，但其底层机制与生物大脑中的注意力过程存在差异。

` * *Sentence 6:* `

Humans can often sustain focus on a specific goal while filtering out competing information.` * *Literal:* 人类通常可以保持专注于特定目标，同时过滤掉竞争信息。 * *Refinement:* "Sustain focus" -> 保持专注. "Filtering out" -> 过滤. "Competing information" -> 竞争性信息/干扰信息. * *Draft:* `

人类往往能够在过滤干扰信息的同时，保持对特定目标的专注。` * *Sentence 7:* `The results suggest that current AI models may struggle with this type of cognitive control when tasks become increasingly demanding.

` * *Literal:* 结果表明，当任务变得日益苛刻时，当前的AI模型可能难以应对这种类型的认知控制。 * *Refinement:* "Struggle with" -> 难以实现/在...方面感到吃力. "Cognitive control" -> 认知控制. "Increasingly demanding" -> 难度日益增加/要求越来越高. * *Draft:* 结果表明，当任务难度不断增加时，当前的人工智能模型可能难以实现这种类型的认知控制。

 

` * *Sentence 8:* `

The researchers argue that the performance collapse seen in these experiments points to fundamental limitations in today's large language models.` * *Literal:* 研究人员认为，在这些实验中看到的性能崩溃指向了当今大型语言模型的基本限制。 * *Refinement:* "Argue" -> 认为/指出. "Performance collapse" -> 性能崩溃/表现崩塌. "Points to" -> 指向/揭示了. "Fundamental limitations" -> 根本局限性. * *Draft:* `

研究人员指出，实验中出现的性能崩溃揭示了当今大型语言模型存在的根本局限性。` * *Sentence 9:* `While AI can sometimes mimic human behavior, its ability to maintain attention appears to operate very differently from the way people do.

` * *Literal:* 虽然AI有时可以模仿人类行为，但其保持注意力的能力似乎运作方式与人类非常不同。 * *Refinement:* "Mimic" -> 模仿. "Operate" -> 运作/运作机制. * *Draft:* 虽然人工智能有时能模仿人类行为，但其维持注意力的方式似乎与人类截然不同。

 

` * *Sentence 10:* `

The findings offer a reminder that even the most advanced AI systems still have weaknesses, particularly when tasks require them to resist distractions and stay focused over extended sequences of information.

` * *Literal:* 这些发现提供了一个提醒，即使是最先进的AI系统仍然有弱点，特别是当任务要求它们抵抗干扰并在扩展的信息序列上保持专注。 * *Refinement:* "Offer a reminder" -> 提醒我们. "Resist distractions" -> 抵抗干扰. "Extended sequences of information" -> 冗长的信息序列/扩展的信息序列. * *Draft:* `

这些发现提醒我们，即使是最先进的人工智能系统仍存在弱点，特别是在需要抵抗干扰并在长序列信息中保持专注的任务中。

` 3. **Final Review and Polish:** * *Check:* Did I keep the HTML tags? Yes. * *Check:* Is the tone professional? Yes. * *Check:* Are technical terms accurate? * "Naming ink colors" -> 命名墨水颜色 (Standard Stroop task terminology). * "Cognitive control" -> 认知控制. * "Large language models" -> 大型语言模型. * "Attention processes" -> 注意力过程. * *Flow Check:* * "People are generally much better at reading words than naming ink colors." -> "人们阅读文字的能力通常远强于命名墨水颜色的能力。" (Slightly smoother). * "Yet despite this bias..." -> "然而，尽管存在这种偏向..." (Good). * "Human Attention vs. Machine Attention" -> "人类注意力与机器注意力" (Standard). * "Underlying mechanisms" -> "底层机制" (Standard). * "Performance collapse" -> "性能崩溃" (Standard in this context). 4. **Final Output Generation:** (Assemble the translated segments). 这一发现尤为有趣，因为人类也面临着类似的冲突。人们阅读文字的能力通常远强于命名墨水颜色的能力。然而，尽管存在这种偏向，大多数人在面对冗长的文字与颜色冲突列表时，仍能保持高准确率和稳定的性能。

 

人类注意力与机器注意力

这项研究凸显了人类智能与人工智能之间的一个重要差异。

尽管现代人工智能系统能够展现出令人印象深刻的语言和推理能力，但其底层机制与生物大脑中的注意力过程存在差异。

人类往往能够在过滤干扰信息的同时，保持对特定目标的专注。结果表明，当任务难度不断增加时，当前的人工智能模型可能难以实现这种类型的认知控制。

研究人员指出，实验中出现的性能崩溃揭示了当今大型语言模型存在的根本局限性。虽然人工智能有时能模仿人类行为，但其维持注意力的方式似乎与人类截然不同。

这些发现提醒我们，即使是最先进的人工智能系统仍存在弱点，特别是在需要抵抗干扰并在长序列信息中保持专注的任务中。

*Self-Correction during assembly:* Ensure the first paragraph flows well. "People are generally much better at reading words than naming ink colors." -> "人们通常阅读文字的能力要远强于命名墨水颜色的能力。" This sounds natural. *Final Polish of the last sentence:* "extended sequences of