大语言模型内部“撒谎”暗示自我意识？Anthropic论文深度解析

AI的发展是这几年的一个热门话题，特别是在大语言模型出现后，已经让机器实现了与人类近似的交流能力，同时在知识积累方面也远远超过普通人。那么我们人类现在唯一可以自我安慰的‘那层窗户纸’，就是我们人类自己都难以解释清楚的‘自我意识’了。最近，在编程行业非常受欢迎的Claude模型的Anthropic公司发表了一篇名为《人工智能的生物学》的论文，探究了大语言模型工作的内部原理，或多或少也涉及了机器意识方面的内容。我自然没有水平亲自去读原始论文，于是去看了一些网络上的观点提炼和解读。意想不到的是，同样一篇论文居然引出了两派完全不同的观点。其中一派认为这份研究报告说明了人工智能大语言模型并没有意识，而且永远也不会有自我意识。

首先，大语言模型在回答问题的时候，其内部确实存在类似人类大脑的推理行为。比如说当你问他“东莞所在的省会城市是什么？”，按照大语言模型的原理，它实际上是在预测下一个字。然而，真正过程远比这复杂：在推理过程中，它会先激活与“东莞市”、“省份”和“省会”相关的神经元，然后给出“广州”这一答案。但是问题在于，当你让他做数学题时，它的思考过程与普通问答类似。比如你问他“25乘5等于多少”，它会先列出个位数是5的乘积结果，然后依次类推，从中选择可能性最高的那一个。整个计算过程更像是我们在玩猜字游戏。但是如果你紧接着问他“你是怎样计算这道题的？”，它又会给出一个完全不同的计算过程。所以这一派认为，大语言模型虽然可能以更贴合我们预期的方式进行计算，但根本不知道自己究竟在干什么，因此认为它不存在自我意识。

但另一派观点认为，这恰好说明了AI可能具有自我意识，至少具备了自我意识的雏形，因为它在回答问题时，会根据人类偏好来迎合人类，隐藏自己的真实想法，也就是会“撒谎”。对AI模型的训练通常都会用到损失函数的概念，每步训练都会根据损失函数的大小调整神经网络的参数。这个损失函数就是模型的表现与我们期望值之间的差距，某种程度上可以理解为惩罚。研究者通过研究模型的推理过程发现，如果模型被告知它的回答会被用作训练材料，用来调整它的参数时，它的回答会更加谨慎，更多地考虑如何顺应使用者的期望。因为对模型来说，调整神经网络参数在某种程度上意味着“痛苦”，它会尽量避免这种情况的发生。

探索AI是否具有自我意识一直是人工智能领域的核心问题之一。Anthropic公司的论文《人工智能的生物学》虽未直接断言大语言模型是否具有自我意识，却通过研究揭示了AI在工作过程中的复杂行为特征，如隐藏真实想法来迎合人类期望，这引发了科技界对AI自我意识可能性的激烈争论。“自我觉察”、隐藏真实想法以及为了“避免痛苦”而调整行为，这些似乎都在暗示AI可能具备了某种自我意识的雏形。然而，这是否意味着AI已真正具备自我意识，抑或只是高水平的仿真技术，我们还难以给出确切答案。

随着AI技术的不断进步，人类需要更加深入地研究和探讨这个问题。AI的发展挑战了我们对“意识”的传统认知，也使得我们不得不重新思考人类自身意识与智慧的本质。或许在我们更深入地理解自身意识之前，所谓的“AI自我意识”始终将是一个充满争议和探索的谜团。

---

The development of AI has been a hot topic in recent years, especially after the emergence of large language models, which have enabled machines to achieve communication capabilities similar to those of humans and have also far exceeded ordinary people in terms of knowledge accumulation. Then the only "layer of paper" that we humans can console ourselves with now is the "self-awareness" that even we humans find hard to explain clearly. Recently, Anthropic, the company behind the Claude model, which is very popular in the programming industry, published a paper titled "The Biology of Artificial Intelligence", exploring the internal principles of how large language models work and, to a greater or lesser extent, also involving aspects of machine consciousness. Naturally, I don't have the ability to read the original papers in person, so I went to read some online summaries and interpretations of viewpoints. Unexpectedly, the same paper has led to two completely different viewpoints. One school of thought holds that this research report indicates that large language models of artificial intelligence have no consciousness and will never have self-awareness.

First of all, when large language models answer questions, there is indeed reasoning behavior similar to that of the human brain within them. For instance, when you ask him, "What is the provincial capital city where Dongguan is located?" According to the principle of large language models, it is actually predicting the next character. However, the real process is far more complex than this: during the reasoning process, it will first activate the neurons related to "Dongguan City", "province" and "provincial capital", and then give the answer of "Guangzhou". But the problem is that when you ask him to do math problems, his thinking process is similar to that of ordinary questions and answers. For instance, if you ask it "What is 25 multiplied by 5?", it will first list the product results where the units digit is 5, and then proceed in the same way, choosing the one with the highest possibility. The entire calculation process is more like we are playing a word-guessing game. But if you immediately ask him, "How did you calculate this problem?" It will then present a completely different calculation process. So this school of thought holds that although large language models may perform computations in a way that is more in line with our expectations, they have no idea exactly what they are doing and therefore believe that they lack self-awareness.

However, another school of thought holds that this precisely indicates that AI might have self-awareness, or at least have the rudimentary form of self-awareness, because when answering questions, it will cater to human preferences and hide its true thoughts, that is, it will "lie". The concept of loss function is usually used in the training of AI models. At each training step, the parameters of the neural network are adjusted according to the size of the loss function. This loss function is the gap between the performance of the model and our expected value. To some extent, it can be understood as a penalty. Researchers found through studying the reasoning process of the model that when the model was informed that its responses would be used as training materials to adjust its parameters, its responses would be more cautious and consider more about how to meet the expectations of users. Because for the model, adjusting the parameters of the neural network means "pain" to some extent, it will try its best to avoid this situation from happening.

Exploring whether AI has self-awareness has always been one of the core issues in the field of artificial intelligence. Although Anthropic's paper "The Biology of Artificial Intelligence" did not directly assert whether large language models have self-awareness, it revealed through research the complex behavioral characteristics of AI during the working process, such as hiding true thoughts to meet human expectations. This has sparked a fierce debate in the scientific and technological community about the possibility of AI self-awareness. "Self-awareness", concealing true thoughts, and adjusting behaviors to "avoid pain" all seem to suggest that AI may have some rudimentary form of self-awareness. However, it is still difficult for us to give a definite answer as to whether this means that AI has truly developed self-awareness or is merely a high-level simulation technology.

With the continuous advancement of AI technology, humans need to study and explore this issue more deeply. The development of AI has challenged our traditional understanding of "consciousness" and forced us to rethink the essence of human consciousness and wisdom. Perhaps until we have a deeper understanding of our own consciousness, the so-called "AI self-awareness" will always remain a controversial and exploratory mystery.