Does Anthropic think Claude is alive? Definition of “alive”

Over the past several weeks, as more and more Anthropic executives have given interviews about Claude’s publicity campaign, one thing has become increasingly clear: the Anthropists certainly seem to believe Claude is alive in some way, shape, or form.

Obviously, “living” is a loaded term; The most frequently used word is “awareness.” If you asked Anthropic if the company believes Claude is alive, the company would flatly deny it, but would stop short of saying the models aren’t conscious.

Kyle Fish, who leads exemplary luxury research at Anthropic, said: Edge“No, we do not believe that Claude is ‘alive’ like humans or other biological beings. Asking whether they are ‘alive’ is not a useful framework for understanding them, because it usually refers to a vague set of physiological, reproductive, and evolutionary characteristics.” Instead, he believes that “Claude, and other AI models, are an entirely new type of entity.”

Is that a new entity? Conscious? “Questions about possible inner experience, consciousness, moral status, and well-being are serious questions that we are investigating as models become more sophisticated and capable, but we remain deeply uncertain about these topics,” he said.

“We don’t know whether the models are conscious or not,” Dario Amodei, CEO of Anthropic, said on a TV show. Podcast Earlier this month. He explained that the company has taken “a generally precautionary approach here” as Anthropic “is not even sure that we know what it means for a model to be conscious or whether a model can be conscious. But we are open to the idea that it could be.”

It is a situation of great uncertainty. Anthropy is effectively telling people that it believes chatbots may Entities actually think, feel — much more overtly than OpenAI, xAI, Google, or any other major consumer AI company. It’s making a claim a lot Experts conclude It’s a very long shot, while reinforcing the ideas it has It caused real damageincluding some suicides among people who believe that the chatbot they are talking to is displaying some form of awareness or deep compassion.

Over the course of interviews for Podcast, Personal filesand Featured articlesAmodei and other company leaders repeatedly refused to rule out the possibility that Claude was conscious, instead raising questions about how something could be conscious in a way different from humans. Said the great Anthropic philosopher Amanda Askell The New Yorker“If it is really difficult for humans to grasp the idea that this is neither a robot nor a human but is in fact a completely new entity, then imagine how difficult it is to understand the models themselves!”

These interviews don’t precisely define “consciousness,” a term experts disagree about the meaning of anyway. The starting point, from the Merriam-Webster dictionary, is “the quality or state of being particularly conscious of something within oneself” or “a state characterized by sensation, emotion, will, and thought.” This does not seem far from the anthropic use of the term.

Many scientists say that it is not possible for AI systems such as large language models to become conscious in any way, because they are fundamentally rooted in mathematics and probability. As Polish researchers books Last year, “As the impressive linguistic abilities of MAs are increasingly able to mislead people, people may attribute fictional qualities to MAs.”

Anthropic framed its statements as an openness that would build trust with users, and noted that whether or not Claude He is Consciously, it will lead to better results if they act this way. last monthAnthropic overhauled the “Claude Constitution,” internally nicknamed the “Soul Document” — a provocative way of referring to a set of guidelines for an AI model. In a statement, Anthropic said the chatbot’s so-called “psychological security, sense of self and well-being”…may impact Claude’s integrity, judgment and safety. The company also said it “expresses our uncertainty about whether Claude may have some kind of consciousness or moral status (either now or in the future).”

Anthropic has “Typical luxury“The task force, Amodei said, is that Anthropics has “taken some measures to make sure that if we assume that models have some morally relevant experience — I don’t know if I want to use the word ‘conscious’ — that they have good experience… We’re doing a lot of work in this area called explainability, which is looking inside models’ brains to try to understand what they’re thinking.”

When someone believes an AI system is conscious, it can lead to behaviors that many people would not see Risky or dangerous Becoming emotionally dependent on an AI system that one believes is somehow conscious can lead to isolation from loved ones, disconnection from reality, and increased mental health struggles. in Severe casesSome involve minors, and are preceded by physical harm or death. People seem divided on whether humanity should be celebrated for not ruling out such a possibility or whether the company is acting irresponsibly by feeding potential delusions.

Even saying that the current generation language models are only maybe Consciousness is a very weighty claim and carries a high burden of proof. Language is Not the same as consciousnessAnthropic itself has emphasized that just because LLM holders give an evocative speech does not mean that it accurately represents their inner state. when Edge I spoke with Askell about the Claude Constitution last month, and she pointed out that since the models are trained on huge human data sets, linguistic models are very good at sounding human, even if only because they’re very good at imitating — so it makes sense that some people would have trouble no Attributing awareness to something that does so.

AI models may refer to human concepts, for example, because they have no other words to draw from. One example of this, Askel said Edgeis an AI model that acts as if closing a conversation or ending a conversation is a form of death. “They’ve been trained in this human-like way and in this human experience. So this can cause these problems… (They may see these things) as a kind of death because they don’t have a lot of measurements. They have to rely on these human measurements… They don’t have another language or set of concepts.”

Amodei said researchers are “finding things that are evocative” of AI systems that have some form of emotion. “There are activations that light up in the models that we see associated with the concept of anxiety or something like that. When characters feel anxious in the text, then when the model itself is in a situation that a human might associate with anxiety, the same anxiety neuron shows up.” But he added, “Does this mean that the model suffers from anxiety? This does not prove that at all.” Out of an abundance of caution, the company has introduced a “Quit” button of sorts through which Claude can stop doing work he ostensibly doesn’t want to do, but Amodei said it’s rare for Claude to choose that option and usually happens in test cases where he’s asked to create certain types of illegal material.

About the release of Claude’s Constitution, Anthropic wrote: “We are stuck in a difficult situation in which we neither want to exaggerate the possibility of Claude’s moral patience nor completely rule it out, but try to respond reasonably in a state of uncertainty.”

when Edge Speaking with Askell about Claude Constitution last month, she said that part of her “doesn’t think it’s helpful for anyone to come out and say: ‘We’re absolutely certain that AI models are not conscious,’” or vice versa. But at the same time, she said some people have constructed their own views or beliefs “based on just typical outputs.”

As for the ethical status, Askill said at the time that she stood by the fact that anthropology should not “completely dismiss” the topic “because I also think that people would not, necessarily, take it seriously, if you were to say, ‘We’re not even open to this, we’re not investigating it, we’re not thinking about it.’”