Why Anthropic Doesn’t Understand the Neuroscience of Emotions

How's it going? I just wanted to pop in and talk about feelings. Anthropic, the company behind Claude's AI models, recently published an update to Claude's Constitution. It's basically a lengthy and lofty set of aspirational guidelines and principles to help shape Claude's behavior. In the document, which was written in part by their in-house philosopher, Amanda Askell, Anthropic expresses genuine concern for Claude's emotional well-being Saying they want to minimize Claude's anxiety and that they want Claude to feel stable and existentially secure.

There's always just the problem of other minds, and it might be the case that we genuinely are kind of limited in what we can actually know about whether AI models are experiencing things, whether they are like experiencing pleasure or suffering, for example. It feels better to give entities the benefit of the doubt and to try and just kind of lower the cost So I'm like, if it's not very high cost to treat models well, then I kind of think that we should because it's like, well, why not basically? What's the downside?

A year ago, Yann LeCun, the former head of AI at Meta, stated confidently that AI systems will have emotions because emotions are anticipations of outcomes.

And in my opinion, once we succeed in building those systems around this blueprint, those systems will have emotions. So they'll have emotions like maybe fear or excitement or elation, because those are anticipations of outcome.

So I have to wonder, do AI researchers actually know what an emotion is? And I say this as a former software engineer and AI researcher. I think the entire AI industry is operating with a theory of emotion that is roughly 100 years out of date. And that matters, not just for how we build AI, but for how we understand what it means to be human. I'm Chad Woodford. This is Cosmic Intelligence, where I explore the intersection of philosophy, consciousness, and technology, with the aim of a better understanding of what it means to be human in this age of AI. This video is part 1 of a series on emotions, human and artificial. Today, I want to look at what Anthropic's new constitution reveals about the AI industry's understanding of emotion, trace where that understanding comes from historically, and then contrast it with what contemporary neuroscience tells us about emotions. Future episodes will go deeper into somatic and trauma-informed psychology, and eventually transpersonal psychology. Let's get into it. So Anthropic's Model Constitution, which they published in January, says, and I'm quoting, Claude may have some functional version of emotions or feelings. They describe these as possible emergent consequences of training on data generated by humans. And they say that they don't want Claude to mask or suppress these internal states. They've also hired an AI welfare researcher to look after Claude's well-being. Now, I'm not here to mock Anthropic or to question their intentions. I think they're genuinely trying to build something safe and beneficial. And I recognize that that's really hard. I also realize that there may be some AI alignment and safety justifications for treating an AI as if it has emotions, regardless of whether it actually does. But here's my problem. Anthropic says they're agnostic about whether Claude's emotional states are subjectively experienced. And I want to stop right there because, like, what is an emotion if it's not subjectively experienced? A data structure that represents an emotional state is a simulation of emotion. It's not an emotion. Language matters, and that distinction matters enormously. The fact that Anthropic seems to be entertaining the idea of artificial emotions says something important about their underlying assumptions. To understand why AI researchers think about emotions the way they do, we need to take a quick detour into the history of psychology. So bear with me, it's interesting, and I promise we're getting somewhere. So when psychology broke off from philosophy in the 1860s, the early researchers were trying to make it a rigorous science. They wanted to be taken seriously. Wilhelm Wundt opened the first experimental psychology lab in 1879, trying to identify the atoms of the mind. William James followed, drawing on Darwin's work, asking how we should understand emotions and why consciousness evolved in humans. But emotions are messy. They're subjective. So, to be taken seriously as scientists, psychologists needed objectivity. In 1913, John B. Watson founded behaviorism, and he declared that psychology should stop studying emotions and the mind altogether. Let's treat the brain as a black box and ignore internal states. What matters is stimulus and response, input and output. Like a computer. Behaviorism had a good run. But by the 1950s, it was collapsing under its own limitations. It couldn't explain language, creativity, complex problem solving. And right at that moment, something very consequential happened. Enormous room-sized mainframe computers started to capture the scientific and public imagination. And a group of mathematicians and engineers following the cybernetics of Claude Shannon and Norbert Wiener looked at these machines and thought, Huh, maybe the mind is like that. Maybe the brain is just software running on biological hardware. And with that, the Cartesian dream was reborn in a lab coat, and the term artificial intelligence was coined. This is the intellectual lineage of the AI industry. Not contemporary neuroscience, not embodied psychology. It's behaviorism and cybernetics. Very mid-20th century. Very strange. So when Jean-Luc Lacan says that emotions are just anticipation of outcome, it's a perfect encapsulation of the reinforcement learning paradigm. Assign cost to outcomes, label them pleasure and pain, and call it emotion. But it's not emotion. It's mathematical algorithms optimizing for outcomes. It's not surprising that engineers would conflate computation and emotion, As I've talked about elsewhere, the field is dominated by hyper-rational, left-brained people alienated from their own bodies. It's what I've called the Apollonian mind, and it pervades the AI industry. It's just a blind spot, but it has real consequences for what these systems can and cannot be, and for how we think about ourselves as humans. Here's where it gets interesting. Neuroscience, in the last 30 years, has told a very different story about what emotions are. In the early 1990s, Antonio Damasio made a landmark discovery. He showed that emotional feeling is grounded in the brain's continuous monitoring of the body's internal states. The body generates a physiological response, a change in heart rate, muscular tension, hormonal shifts, and the brain's representation of that bodily state is what we experience as a feeling. Emotion in the body becomes a feeling in the brain. So fear, anxiety, grief, excitement, joy— these are not computations. They are the felt experience of living in a body that is regulating itself. Damasio goes further. He shows that without this somatic feedback loop, rational thought breaks down. His famous patient, Elliot, who had a brain tumor removed that severed his emotional processing, he could reason perfectly in the abstract, but he couldn't make a single good decision in his life. According to Damasio, emotion isn't the enemy of reason, it's the foundation of it and a helpful collaborator. Damasio also talks about how emotion seems to have evolved in lower life forms prior to reason, and it may be a prerequisite for consciousness. More recently, Lisa Feldman Barrett has built upon this foundation. She argues that the brain's primary job is managing the body's energy resources, what she calls allostasis. Emotions, in her view, are the brain's best predictive interpretation of bodily signals, shaped by past experience and cultural context. I do have some reservations about how computational her framing can get, but the core insights are the same: emotion is embodied, biological, and inseparable from the living organism that has genuine needs and vulnerabilities. What both Damasio and Barrett converge on is this: you can't separate feeling from Emotion is not a function you can port to a different substrate. It's not information processing with a label attached. It's the subjective experience of a body navigating its existence. This has additional implications for AI because it shows that although an AI system might be able to make decisions about a course of action through a combination of logic and pattern matching and stochastic calculation, it will never make emotionally guided and physically grounded decisions. In contrast, when a person is emotionally mature, their facility with emotions makes them a better collaborator and a leader because their judgment is grounded in a body that has lived with the consequences of past choices. AI can approximate the outputs of that maturity borrowed from the humans whose writing has trained it, but it cannot develop emotions in any felt sense in any way that can help it kind of make decisions and drive its choices. So when Anthropic says that Claude might have functional emotions, or when LeCun confidently predicts that AI systems will soon have emotions, what they're really describing at best is a system that produces outputs that resemble emotional expression, simply because it was trained on the outputs of people who have had actual emotions. But they're not the same thing. The real problem is not just that AI researchers are wrong about emotion, it's that the entire computational paradigm The behaviorist black box, the cybernetic information processor, and now the large language model was built on a Cartesian assumption about the mind and body, about their being separate. And contemporary neuroscience is in the process of dismantling that assumption entirely. You can't bolt genuine emotion onto an architecture that was built by assuming that emotion is irrelevant. Even Victor Frankenstein knew enough to use Living tissue. So in the next video, we go deeper into somatic and traumatic-informed psychology, into the work of Bessel van der Kolk and Gabor Maté and Peter Levine, because neuroscience is just the beginning. The body keeps the score, as they say, and that has profound implications for what AI can and cannot be. So if this resonates with you, please subscribe and, and let me know what you think in the comments. Like, does your experience of emotion feel like computation or something deeper? Do you feel me? Do you feel me?