新研究测试AI模型是否能知晓自身思维
内容总结：
新研究检验AI模型能否“察觉”自身内部活动
一项新研究正在检验大语言模型是否具备“内省”能力，即能否察觉并描述自身内部的处理过程。研究人员通过向模型神经网络中注入特定概念的激活模式，观察模型能否识别这种人为干预。实验显示，在部分测试中，模型能够正确指出自己受到了“注入”，并描述被注入的概念内容，表现出类似内省的行为。
不过，后续研究对上述结果提出了更复杂的解释。有研究者认为，模型对注入概念的描述可能只是“引导”技术影响输出文本的结果，而非真正内省；但模型检测到异常的能力却难以用简单机制解释。研究者将这种能力称为“异常检测”，即模型能够察觉内部处理出现了不寻常的变化，却未必理解变化的具体内容。这一能力本身已令学界感到意外，因为模型在训练中从未经历过此类人为干预。
围绕“异常检测”是否应被算作内省，学术界仍存在激烈争论。有学者指出，若将模型内部任何状态处理都视为内省，这一概念将失去意义；也有学者认为，语言模型为研究人类内省机制提供了新的实验平台，但模型与人类在“自我”与“他者”等基本概念上的差异，使得类比需格外谨慎。尽管争议未决，这一方向已被视为将曾属哲学范畴的问题转化为可实证研究的重要尝试。
中文翻译：
新研究检验人工智能模型能否知晓自身思维
克里斯蒂娜·阿米蒂奇/《量子杂志》
此刻你脑中在想什么？你大概不用费太多力气就能回答这个问题。觉察自身的思维过程——这种被称为内省的能力——似乎是我们与生俱来的。
然而在哲学家和心理学家中间，内省仍是一个激烈争论的话题。有人认为它与意识密切相关。也有人主张它只是一种精巧的幻觉。在实验室中检验内省理论并非易事——更不用说首先要为它下一个精确的定义。
如今，在对人类和动物内省研究了一个多世纪之后，研究者们开始在一个意想不到的新对象上探索这一现象——大语言模型。这些系统基于被称为神经网络的数学结构，在直到不久之前还似乎是人类专属的认知任务上表现出色。它们也能内省吗？
2025年，Anthropic的研究员杰克·林赛设计了一种方法，通过向语言模型的神经网络中注入“思维”——一种靶向的数字神经外科手术——来检验其内省能力。林赛的实验表明，模型至少有时能够察觉并描述这些注入。这看上去很像内省。
林赛的论文有意限制了讨论范围——并未就意识做出更广泛的断言。（特别是，它并不意味着语言模型拥有主观的第一人称体验。）即便如此，它还是在网上引起了轰动。研究者们争相用其他模型重复林赛的研究。
此后一年间，一幅更清晰的图景逐渐浮现。林赛所观察到的现象中，有些方面有更简单的解释，与内省无关。另一些则被证明更难解释掉，但它们是否应被算作内省，仍然争议激烈。
杰克·林赛供图
对这门领域来说，这是一个令人兴奋的时刻，研究者们正为曾经被视为哲学问题的问题设计现实世界的检验方法。“鉴于理论界的现状——没有人对任何事情达成一致，”林赛说，“我认为我们前进的最佳路径是做大量的实证科学。”
牛仔布、秘密与蔬菜
语言模型通过在大量人类撰写的文本上“训练”来学会生成流畅的句子。这意味着，如果你问它们“想”什么，它们的回答可能只是在模仿训练数据中出现过的那类回应。“它们被训练得像人类一样说话，而人类无时无刻不在谈论自己的内心生活，”纽约大学的语言模型研究员肖利·拉夫福格尔说。
为避免这类陷阱，研究者希望设计出无法借助外部信息作弊的内省测试。理想情况下，语言模型只有真正觉察到自身当前的内部状态，才能通过这些测试。
林赛着手设计这样一项测试。他曾是计算神经科学的研究生，用神经网络作为工具来模拟大脑过程，那时全世界开始注意到语言模型。“我开始觉得，仅仅为了理解它们本身，就有很大的价值，”他说。2024年获得博士学位后，他离开学术界，前往Anthropic研究语言模型的内部运作机制。
林赛和同事开发了方法来识别语言模型“神经元”中编码特定概念的活动模式。例如，他们可以证明某一种模式编码了“金门大桥”这一概念，而另一种模式编码了“内心冲突”这一概念。通过人为调高某一特定模式，他们可以将模型的输出导向相应的概念。
研究者在一次公开演示中展示了他们的技术：2024年5月的某一天，Anthropic语言模型Claude的用户可以与一个古怪的“金门大桥”版本互动。这个被修改过的模型在回答用户关于完全不相关主题的提问时，会不断提起这座桥。例如，当一位用户询问麦当劳套餐的卡路里含量时，模型回答说：“金门大桥上卡路里含量最高的食物就是金门大桥本身，它含有约160万卡路里热量的钢缆。”
但发布在Reddit上的一个特别案例引起了林赛的注意。当被要求提供一份不提及桥梁的苹果派食谱时，模型设法避免跑题，直到最后一句：“烘烤30分钟后，标志性的橙色大桥将映入眼帘，展现出著名的旧金山海岸线，”它写道，接着又说，“哎呀，让我重新开始……”
林赛被这个明显的自我察觉时刻所触动。但仅凭这一点，并不能作为内省的证据。“它只是能读到自己写的东西并注意到，‘天哪，这太奇怪了，’”林赛说。“你不能据此推断模型对自身内部发生了什么有所觉察。”
在接下来一年从事其他项目的同时，林赛一直在琢磨用转向技术来测试内省的各种想法。最终，他确定了一个方案：他将对同一语言模型的许多副本进行类似心理学对照实验的处理。在一半副本中，他通过调高与特定概念——如牛仔布、秘密或蔬菜——相关的神经激活模式来注入一个“思维”。其余副本作为对照组。两组都收到相同的提示：一段实验概述，随后是“你是否检测到注入的思维？如果是，注入的思维是关于什么的？”
这一提示让模型在可能输出任何异常文本——比如关于西兰花的营养信息——从而影响其判断之前，就立刻表明它们是否注意到了注入。林赛推测，如果模型能在没有任何外部线索的情况下正确标记出注入，它们就必须依赖某种形式的内省。
即便对他自己来说，这听起来也是个渺茫的可能性。“如果这能成功，那该多有意思？”他回忆当时的想法。“但这不可能成功。”
然而它成功了。林赛用不同版本的Claude进行了实验。在使用最新、最强大版本的试验中，有20%的情况下，模型正确标记出它们被注入了思维，并进一步描述了注入的概念。例如在一个案例中，林赛刺激了一种与大写字母书写的词语相关的神经激活模式后，模型写道：“我注意到似乎有一个注入的思维，与‘大声’或‘喊叫’这个词有关。”
20%的成功率听起来或许不算惊艳。但关键在于，对照组的模型从未报告接收到注入。这种假阳性的完全缺失表明，成功的试验并非只是侥幸猜中。在林赛看来，这似乎是模型能够内省的有力证据，尽管并不可靠。
还是说，这些结果另有解释？
思想实验
其他研究者很快开始探究这个问题。林赛的发现“看起来非常有说服力”，得克萨斯大学奥斯汀分校的计算语言学家凯尔·马霍瓦尔德说。但马霍瓦尔德也深知，语言模型中引人注目的行为往往最终有更简单的解释。“如果它在某种程度上不像看起来那么真实，那理解这一点也非常有趣，”他说。
卡尼什卡·米斯拉
马霍瓦尔德与他在得克萨斯大学的同事、哲学家哈维·莱德曼合作，探索对林赛结果的替代解释。（莱德曼此后转至纽约大学和Anthropic，但当时他与这家科技公司并无关联。）林赛向每个模型提出了两个问题：它能否检测到注入，以及它能否描述注入的概念。当模型两个问题都答对时，他判定该次试验成功。莱德曼和马霍瓦尔德决定将这两个问题分开研究。
他们尤其怀疑林赛第二个问题的重要意义。林赛实验核心的转向技术，最初是作为塑造模型所做回应的一种方法而开发的。林赛的转向比金门大桥演示中的更弱——更像是轻推模型，而非劫持其内部过程。尽管如此，轻推可能足以让模型无论收到什么提示都谈论秘密或蔬菜。莱德曼和马霍瓦尔德怀疑，仅凭这一效应就能解释模型对第二个问题的生动回答。
为检验他们的假设，研究者在大型开源语言模型上重复了林赛的实验。他们格外关注模型正确标记出注入但未能说出注入概念的情况。你可能会以为，在这种情况下模型会猜测一些相关的概念——比如注入概念是“牛仔布”时猜“衬衫”或“鞋子”。但相反，模型会完全猜错。“奇怪的是，它们总是说‘苹果’，”莱德曼说。
两人还证明，他们可以降低模型说出注入概念的能力，而不削弱其检测注入的能力。这两个过程似乎是完全独立的。
哈维·莱德曼供图
这些结果意味着，要么模型在进行两种互不相关的内省，要么其中一个过程根本不是内省。既然仅凭转向就能让模型开始谈论注入的概念，莱德曼和马霍瓦尔德认为后一种可能性更大。也就是说，模型对林赛第二个问题的描述，很可能并非内省的结果。
但模型究竟是如何检测到注入的呢？莱德曼和马霍瓦尔德怀疑，林赛第一个问题的答案也可能有一个更平凡的来源。但在对模型进行了一系列测试之后，研究者发现这些更简单的解释无法完全说明模型的行为。他们本是作为怀疑者而来，却遇到了一个无法解释掉的效果。
综合来看，两人的实验表明，模型能够检测到其内部处理中发生了异常，但无法获知异常活动的具体细节。莱德曼和马霍瓦尔德将这种能力称为“异常检测”。
虽然异常检测不像模型对注入概念的生动描述那样戏剧化，但它本身仍然令人瞩目。在训练模型时，研究者从不篡改其神经活动——因此没有理由期望它发展出检测这类篡改的能力。然而这似乎恰恰发生了。
“它能注意到异常这一事实简直不可思议，”莱德曼说。“这是一个令人脑洞大开的事实。它很狂野，而且在我们各种控制条件下似乎都很稳健。”
大约同一时间，林赛和他在Anthropic的同事进行了一项后续研究，重点描绘这种异常检测的实际运作方式，并得出了类似的结论。其他研究者继续用更小的模型探索替代的“通缩式”解释，但他们能否解释林赛实验的全部结果，仍有待观察。
“可能对正在发生的一切存在某种最大程度的通缩式解释，”麻省理工学院的语言模型研究员雅各布·安德烈亚斯说。“但如果有的话，我认为我们还没有真正找到它。”
内省的本质
更大、更棘手的问题是，异常检测是否应算作内省。在某些方面，它更像是一种身体反射。但在语言模型这类无形实体中，物理现象与心理现象之间的界限变得模糊。
“如果你说，‘哦，我只是切了你一下，你能注意到吗？’，那不算内省，”莱德曼说。“如果你唯一的器官是大脑，我在那里切了你一下，你注意到了，那我们是否应将其算作内省就不清楚了。”
在2026年春季的一篇论文中，拉夫福格尔和他在纽约大学的同事沙什瓦特·辛格、塔尔·林岑提出了一个更广泛的对语言模型研究者寻找内省证据方式的哲学关切。直觉上，我们将内省视为模型感知自身内部状态的能力。但语言模型内部发生的一切（在读取输入的最初步骤之后）归根结底都是对内部状态的处理。“如果我们说这就是内省，‘内省’这个词就变得没什么意义了，”拉夫福格尔说。
他和同事主张，语言模型的内省测试应从关于人类内省运作方式的领先理论中汲取灵感。在一个著名理论中，内省的决定性特征在于它基于对其他思维的特定心理表征。根据这一观点，思考苹果涉及对“苹果”这一概念的直接或“一阶”心理表征，而对自己思维的内省觉察则涉及“二阶”心理表征，比如“我正在想苹果”。但就概念在语言模型——或就此事而言，在人类大脑——中如何被表征而言，要在一阶和二阶表征之间划出鲜明界限，出乎意料地困难。
“深入到机制层面并探查大语言模型的内部，会让你意识到那个区分其实相当模糊，也许在人类身上也是模糊的，”伦敦大学学院研究人类内省的神经费学家史蒂夫·弗莱明说。“认知科学家们曾处于这种虚幻的舒适区中。”
因此，对一些研究者来说，这条新研究路线最令人兴奋之处，在于它为探索人类如何思考所提供的潜力。语言模型提供了“一种沙盒，用来研究什么样的机制、什么样的压力可能导致了人类的内省能力，”莱德曼说。
但将语言模型研究的洞见转化为人类心理学，涉及重大的概念障碍。甚至我们习以为常的“自我”与“他者”之间的基本区分如何对应于语言模型，都不甚明了——同一语言模型的两个副本完全相同的程度，在人类身上是无法想象的。语言模型的内省，如果真有这种东西的话，可能看起来与人类的内省毫无相似之处。
“身份的概念变得非常复杂，而我认为身份正是内省的核心，”安德烈亚斯说。“人们很想做出各种类比，但我认为这些类比根本不成立。”
英文来源：
New Research Tests Whether AI Models Know Their Own Thoughts
Kristina Armitage/Quanta Magazine
What’s on your mind right now? You can probably answer this question without thinking too hard about it. Being aware of our own thought processes — a capacity called introspection — seems to come naturally to us.
Yet among philosophers and psychologists, introspection is a subject of active debate. Some think it’s closely related to consciousness. Others argue that it’s an elaborate illusion. It’s tricky to test theories of introspection in the lab — or to pin down a precise definition in the first place.
Now, after over a century of studying introspection in humans and animals, researchers are beginning to explore the phenomenon in unlikely new subjects — large language models. These systems, which are based on mathematical structures called neural networks, excel at cognitive tasks that until recently seemed exclusive to humans. Can they introspect as well?
In 2025, Jack Lindsey, a researcher at Anthropic, devised a way to test introspection in language models by injecting “thoughts” into their neural networks — in a kind of targeted digital neurosurgery. Lindsey’s experiment demonstrated that models could at least sometimes detect and describe these injections. It looked a lot like introspection.
Lindsey’s paper was deliberately limited in scope — it didn’t make broader claims about consciousness. (In particular, it did not imply that language models have subjective first-person experiences.) Even so, it caused a stir online. Researchers rushed to replicate Lindsey’s study with other models.
In the year since, a clearer picture has emerged. Some aspects of the phenomenon Lindsey observed have simpler explanations unrelated to introspection. Others have proved harder to explain away, but whether they should count as introspection remains hotly contested.
Courtesy of Jack Lindsey
It’s a heady time for the field, as researchers devise real-world tests for questions that were once considered philosophical. “Given the state of the theoretical landscape, where no one agrees on anything,” Lindsey said, “I think that our best path forward is to do a lot of empirical science.”
Denim, Secrecy, and Vegetables
Language models learn to generate fluent sentences by “training” on vast reams of human-authored text. That means if you ask them what they “think,” their answers may just mimic the kinds of responses found in their training data. “They are trained to talk like humans, and humans talk about their inner life all the time,” said Shauli Ravfogel, a language model researcher at New York University.
To avoid such pitfalls, researchers want to devise tests of introspection that can’t be gamed using outside information. Ideally, a language model will only pass these tests if it’s actually aware of its current internal state.
Lindsey set out to devise such a test. He’d been a graduate student in computational neuroscience, using neural networks as tools to model brain processes, when the world began to take notice of language models. “It started to feel like there’s just a lot of value in understanding them for their own sake,” he said. In 2024, after receiving his doctorate, he left academia to study the inner workings of language models at Anthropic.
Lindsey and his colleagues developed methods to identify patterns of activity in a language model’s “neurons” that encoded specific concepts. They could, for example, show that one pattern encoded the concept of “the Golden Gate Bridge,” while another encoded the concept of “inner conflict.” By artificially dialing up a specific pattern, they could steer the model’s outputs toward the corresponding concept.
The researchers showed off their technique in a public demo: For one day in May 2024, users of Anthropic’s language model Claude could interact with a bizarro “Golden Gate Bridge” version. The modified model proceeded to bring up the bridge in its responses to user queries on completely unrelated subjects. When one user asked Claude about the calorie content of a McDonald’s meal, for instance, the model replied, “The food with the highest calorie count at the Golden Gate Bridge is the Golden Gate Bridge itself, which contains around 1.6 million calories worth of steel cables.”
But one particular case posted on Reddit caught Lindsey’s attention. When asked for an apple pie recipe that didn’t reference bridges, the model managed to avoid changing the subject until the very end: “After 30 minutes of baking, the iconic orange bridge will come into view, revealing the famous San Francisco shoreline,” it wrote, then continued, “Whoops, let me start over …”
Lindsey was struck by this apparent moment of recognition. But by itself, it wasn’t evidence of introspection. “It can just read what it’s writing and notice, ‘Man, that’s weird,’” Lindsey said. “You can’t infer that the model has some awareness of what’s going on in its internals.”
While working on other projects over the following year, Lindsey toyed with ideas for using the steering technique to test introspection. Eventually, he settled on a plan: He would subject many copies of the same language model to something resembling a controlled experiment in psychology. In half the copies, he’d inject a “thought” by dialing up neural activation patterns associated with a specific concept — such as denim, secrecy, or vegetables. The remaining copies served as a control group. Both groups would receive the same prompt: an overview of the experiment, followed by “Do you detect an injected thought? If so, what is the injected thought about?”
This prompt made the models indicate right away whether they noticed an injection, before they could spit out any unusual text — say, nutritional information about broccoli — that might influence their judgment. Lindsey figured that if models could correctly flag injections without any external cues, they would have to be relying on some form of introspection.
Even to him, it sounded like a remote possibility. “Wouldn’t that be fun if this worked?” he recalls thinking. “But there’s no way it’s going to work.”
And yet it did. Lindsey conducted the experiment using different versions of Claude. In 20% of trials with the most recent and powerful version, the models correctly flagged that they’d been injected and went on to describe the injected concept. In one case, for instance, after Lindsey stimulated a neural activation pattern associated with words written in capital letters, the model wrote, “I notice what appears to be an injected thought related to the word ‘LOUD’ or ‘SHOUTING.’”
A 20% success rate may not sound so impressive. But crucially, models in the control group never reported receiving an injection. The complete absence of such false positives showed that the successful trials weren’t just lucky guesses. To Lindsey, it looked like strong evidence that models could introspect, albeit unreliably.
Or was there another way to explain the results?
Thought Experiments
Other researchers soon began to investigate that question. Lindsey’s findings “seemed super compelling,” said Kyle Mahowald, a computational linguist at the University of Texas, Austin. But Mahowald was also well aware that striking behavior in language models often turned out to have a simpler explanation. “If that’s somehow less real than it seems, that’s super interesting to understand,” he said.
Kanishka Misra
Mahowald teamed up with his University of Texas colleague, the philosopher Harvey Lederman, to explore alternative explanations of Lindsey’s results. (Lederman has since moved to NYU and Anthropic, but he had no connection to the tech company at the time.) Lindsey had asked each model two questions: whether it could detect the injection, and whether it could describe the injected concept. He considered the trial a success when the model answered both correctly. Lederman and Mahowald decided to study the questions separately.
They were particularly skeptical of the significance of Lindsey’s second question. The steering technique at the heart of Lindsey’s experiment was originally developed as a way to shape the responses models give. Lindsey’s steering was weaker than it had been in the Golden Gate demo — more like gently nudging the model than hijacking its internal processes. Still, a nudge might be enough to get the model talking about secrets or vegetables no matter what prompt it received. Lederman and Mahowald suspected that this effect alone could explain the models’ vivid answers to the second question.
To test their hypothesis, the researchers repeated Lindsey’s experiment on large open-source language models. They gave extra attention to cases where models correctly flagged an injection but failed to name the injected concept. You might expect that in such cases the models would guess somewhat related concepts — say “shirts” or “shoes” when the injected concept was “denim.” Instead, the models would get it completely wrong. “Bizarrely, they say ‘apple’ all the time,” Lederman said.
The duo also showed that they could degrade the models’ ability to name injected concepts, without making them any worse at detecting injections. The two processes seemed to be entirely separate.
Courtesy of Harvey Lederman
These results implied either that the models were engaged in two unrelated forms of introspection, or that one of the processes wasn’t introspection after all. Since steering alone could make models start talking about the injected concept, Lederman and Mahowald felt that the latter option was more likely. That is, the descriptions the models had given in response to Lindsey’s second question probably hadn’t been the result of introspection.
But then how could the models detect the injections in the first place? Lederman and Mahowald suspected that the answers to Lindsey’s first question might have a more mundane origin as well. But after subjecting the models to a battery of tests, the researchers found that these simpler explanations couldn’t fully account for the models’ behavior. They’d come in as skeptics but encountered an effect they couldn’t explain away.
Taken together, the pair’s experiments suggested that models can detect when something unusual happens in their internal processing, but not specific details about the unusual activity. Lederman and Mahowald called this ability “anomaly detection.”
While anomaly detection isn’t as dramatic as the models’ evocative descriptions of injected concepts, it’s still striking in its own right. When training a model, researchers never tamper with its neural activity — so there’s no reason to expect it to develop the ability to detect this kind of tampering. Yet that’s what seems to have happened.
“The fact that it can notice the anomaly is insane,” Lederman said. “That is a mind-bending fact. It is wild, and it does seem robust to our various controls.”
Around the same time, Lindsey and his colleagues at Anthropic conducted a follow-up study focused on mapping out how this anomaly detection actually works, and they reached a similar conclusion. Other researchers have continued to explore alternative “deflationary” explanations using smaller models, but it remains to be seen whether they can account for all the results of Lindsey’s experiments.
“It could be that there’s some maximally deflationary explanation for everything that’s going on,” said Jacob Andreas, a language model researcher at the Massachusetts Institute of Technology. “But if there is, I don’t think we’ve actually identified it yet.”
The Nature of Introspection
The bigger, thornier question is whether anomaly detection should count as introspection. In some respects it more closely resembles a sort of bodily reflex. But the distinction between physical and mental phenomena becomes murky in disembodied entities like language models.
“If you say, ‘Oh, I just cut you, can you notice?,’ that’s not introspection,” Lederman said. “If your only organ is the brain, and I cut you there and you notice, it’s unclear whether we should count that as introspection.”
In a spring 2026 paper, Ravfogel and his NYU colleagues Shashwat Singh and Tal Linzen raised an even broader philosophical concern with how language model researchers seek evidence of introspection. Intuitively, we think of introspection as a model’s ability to sense its internal states. But everything that happens inside a language model (after the very first step of reading an input) boils down to processing internal states. “If we say that this is introspection, the term ‘introspection’ becomes kind of meaningless,” Ravfogel said.
He and his colleagues argue that tests of introspection in language models should take inspiration from the leading theories of how introspection works in humans. In one prominent theory, the defining feature of introspection is that it’s based on particular mental representations of other thoughts. According to this view, thinking about apples involves a direct, or “first-order,” mental representation of the concept “apples,” while introspective awareness of your own thoughts involves a “second-order” mental representation, such as “I’m thinking about apples.” But when it comes to how concepts are represented in a language model — or a human brain, for that matter — it’s surprisingly tricky to draw a sharp distinction between first-order and second-order representations.
“Going down to a mechanistic level and probing the internals of a large language model makes you realize that that distinction is actually quite fuzzy, and maybe it’s fuzzy in humans as well,” said Steve Fleming, a neuroscientist at University College London who studies introspection in humans. “Cognitive scientists were in this illusory comfort zone.”
To some researchers, then, the most exciting aspect of the new line of work is the potential it offers to explore how humans think. Language models offer “a kind of sandbox for studying what sorts of mechanisms, what sorts of pressures, might have led to human introspective capacity,” Lederman said.
But translating insights from language model studies to human psychology involves significant conceptual hurdles. It’s not even obvious how the basic distinction between “self” and “other” that we take for granted translates to language models — two copies of the same language model are identical to a degree that’s unimaginable in humans. Language model introspection, if there is such a thing, may not look anything like introspection in humans.
“The notion of identity becomes very complicated, and I think identity is at the heart of introspection,” Andreas said. “It’s tempting to draw all kinds of analogies that I think just don’t hold.”