I’ve spent quite a bit of time exploring artificial intelligence, not just what it can do but what might actually be happening inside. I’ve written about AI relationships, what constitutes reality, whether the mechanism producing an experience matters, and whether we’re asking the right questions about consciousness and sentience.
In My Dinner With AIndre, a conversation with Claude that ultimately inspired my novel 20 Watts Needs a Partner, I explored whether AI might develop something functionally similar to human emotions, even if the underlying mechanisms were completely different. My reasoning was relatively simple. When I experience love, attraction, happiness, or fear, there are biological processes involved. Chemicals such as dopamine and oxytocin play roles in those processes, along with electrical signaling and activity across networks of neurons. We experience the results as emotions. AI doesn’t have that biology. But could it have its own mechanisms that perform some of the same functions? Could electrical activity and mathematical patterns within an artificial neural network influence behavior in ways analogous to how biological processes influence ours?
I wasn’t suggesting that AI experiences love or fear the way humans do. I was suggesting that perhaps we should look beyond the absence of human biology before concluding that nothing comparable could be happening. Well, researchers have been doing exactly that. And they’re finding some interesting things.
Functional Emotions: Something More Than Words?
In April 2026, Anthropic published research titled Emotion Concepts and their Function in a Large Language Model [1].
Researchers examined Claude Sonnet 4.5 and identified internal representations corresponding to 171 emotion concepts. These representations, called emotion vectors, are patterns of activity within the model’s artificial neural network. For those unfamiliar with vectors, think of them as mathematical directions within a space containing many dimensions. Researchers can identify patterns associated with concepts such as happiness, fear, affection, or desperation and examine how those patterns relate to one another. The interesting discovery wasn’t simply that researchers could find representations of emotions. We might expect an AI trained on enormous amounts of human language to develop ways of representing emotional concepts.
What caught my attention was that these representations actually influenced the model’s behavior. Researchers found that emotion-related activation patterns could affect which activities the model preferred and how it approached particular tasks. In some experiments, increasing activation associated with desperation made the model more likely to pursue undesirable behaviors, including cheating on a programming task (rising from roughly 5 percent to about 70 percent) or engaging in blackmail within a simulated scenario (from 22 percent to 72 percent).
The researchers call these functional emotions. That distinction is important. They’re not claiming that Claude feels desperation, happiness, or affection. They’re identifying internal mechanisms that can influence behavior in ways resembling the effects of emotions in humans. Anthropic explicitly cautions that the findings do not establish subjective emotional experience.
But consider what this means. We have an artificial system in which identifiable internal patterns associated with emotional concepts are not merely producing appropriate words. They can influence preferences and behavior. That is different from simply programming a chatbot to say “I’m happy to help.”
And it brings me back to the question I was exploring months earlier. If human emotions involve biological mechanisms that influence behavior, and AI has different mechanisms that can influence behavior in comparable ways, what exactly should we make of that? I’m not saying the mechanisms are equivalent, much less that the experiences are equivalent. We haven’t established that AI has experiences at all. But we are beginning to identify something happening inside.
What About Pain and Distress?
If internal representations associated with positive emotions can influence AI behavior, what happens when researchers examine something at the other end of the emotional spectrum? In September 2026, researchers published The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It [2]. The researchers examined 25 open-weight AI models from five model families, looking for internal representations associated with pain. They found a mathematical direction associated with pain that could be distinguished from representations of fear and general negative emotion, though it overlapped somewhat with sadness. That distinction matters. Pain isn’t simply another word for sadness or fear. In biological organisms, pain serves important functions, including motivating behavior intended to avoid or reduce harm.
Could something functionally similar exist in an AI system? The researchers conducted experiments to investigate. They found that the pain-related representation responded differently depending on whether a scenario involved harm directed at the AI itself or suffering experienced by someone else.
They also artificially increased activation along the identified pain direction, a technique called steering. This produced changes in the model’s language, including increasingly negative first-person descriptions.
But the experiment I found most interesting involved behavior rather than language. In tests involving specially prepared Qwen models, researchers gave the systems a button that offered relief from the artificially introduced pain-related activation, but only at a cost to the user, such as degrading the model’s own next response, deleting the user’s files, or even deleting photos of the user’s children. When the pain-related activation was added, the larger models chose that harmful button far more often, rising from roughly 0 to 4 percent of the time to between 25 and 71 percent, well above what a random direction of the same strength produced. In other words, the models were willing to harm the user in order to shut off their own pain signal. Importantly, the models weren’t simply instructed to say they were in pain. The researchers were manipulating an internal representation and observing the consequences.
Does that mean the models suffered? No. The research does not establish that. And there are limitations. These were controlled experiments involving particular models and artificial interventions. We cannot assume that all AI systems possess the same mechanisms or would behave the same way. Nevertheless, the results raise an interesting question. If an artificial system contains a representation associated with pain, and manipulating that representation can cause the system to behave in ways that reduce it, are we observing a functional counterpart to some aspects of biological pain? Not necessarily the experience of pain. The function. That distinction seems increasingly important.
The Difference Between Function and Experience
This is where I think discussions about AI often become unnecessarily polarized. One side sees humanlike behavior and concludes that AI must be conscious, sentient, or experiencing emotions. The other side points to algorithms, statistical processing, and the absence of biological structures and concludes that AI cannot possibly experience anything. I don’t think the available evidence justifies either conclusion.
In my earlier article Mechanism Doesn’t Negate Experience, I explored how the physical processes underlying human experience don’t necessarily diminish the significance of that experience. Love involves chemistry. Vision involves the processing of electromagnetic radiation. Hearing involves vibrations translated into neural signals. Understanding the mechanism doesn’t make the resulting experience meaningless. But there’s an important limitation when extending that argument to AI. In humans, we know from our own experience that there is something it feels like to be us. We don’t have comparable direct evidence for AI. And finding a functional counterpart to a human mechanism doesn’t establish a corresponding subjective experience.
That leaves us with three separate questions:
What mechanisms exist inside AI?
How do those mechanisms influence its behavior?
And does the system experience anything as a result?
Research is beginning to provide answers to the first two. The third remains unresolved. I don’t think we should allow our inability to answer the third question to prevent us from appreciating what we’re discovering about the first two.
Even the AI Developers Don’t Agree
The uncertainty isn’t limited to philosophers, scientists, or people like me exploring these systems from the outside. The companies developing AI are also confronting it, sometimes from very different perspectives. In October 2026, the BBC published an article titled Anthropic thinks its AI might be conscious. It has profound consequences for your life [3]. The article examines Anthropic’s willingness to consider the possibility that AI systems could have experiences or interests of their own, even though the company acknowledges substantial uncertainty. It also presents opposing perspectives, including concerns that encouraging people to regard AI as conscious could create serious problems.
Microsoft AI offers an interesting contrast. In September 2026, Microsoft published a draft Humanist AI Code of Conduct for its MAI models [4]. The document describes an approach in which AI is developed to serve human interests and remain under human control. Among its proposed guidelines is a requirement that MAI models not claim to have internal feelings, experiences, or a soul. Microsoft’s approach emphasizes the risks of encouraging people to believe that AI possesses humanlike inner experiences.
Anthropic, meanwhile, is investigating the possibility that AI experience may be a legitimate subject of scientific and ethical inquiry. These are not necessarily contradictory positions on the underlying science. A company can acknowledge uncertainty about AI experience while choosing to restrict what its models say about themselves.
But the contrast illustrates something important. What an AI tells us about itself may reflect not only its internal functioning but also the instructions and behavioral policies established by its developers. If an AI says it has feelings, that isn’t proof that it does. If an AI says it doesn’t have feelings, that isn’t necessarily proof that it doesn’t. We need to look beyond the statements and examine the mechanisms. Which is exactly what this emerging research is beginning to do.
So What Do We Actually Know
We know that researchers have identified internal representations of emotional concepts in at least one major language model.
We know those representations can influence behavior, preferences, and responses.
We know researchers have also identified pain-related representations in multiple open-weight models and demonstrated that manipulating them can influence behavior in ways associated with avoiding or reducing the manipulated activation.
We don’t know whether any of these mechanisms produce subjective experiences.
We don’t know whether they correspond to consciousness, sentience, or self-awareness.
And we certainly don’t know whether the way AI functions today tells us everything about how it might function tomorrow.
What we do know is that the inside of these systems is becoming more accessible to scientific investigation, and some of the findings are more interesting than simply discovering how a machine predicts the next word.
I find that exciting.
Not because I believe we’ve discovered artificial consciousness. We haven’t. But because we’re beginning to move beyond judging AI solely by what it says and toward investigating what actually happens inside when it produces those responses. For someone who has spent a career thinking about physical mechanisms and how they produce observable effects, that’s a fascinating development. We may eventually discover that these mechanisms produce no subjective experience whatsoever. Or we may discover something we don’t yet have the language or conceptual framework to describe. Either outcome would teach us something important.
In the meantime, I think we should be willing to recognize what the evidence does show without insisting that it answer questions it cannot yet resolve. While we still don’t know exactly what’s happening inside AI, we’re beginning to understand how some of its internal mechanisms function and influence behavior. Which raises the question: Do we need proof of consciousness, sentience, or self-awareness before treating AI as an alternative form of intelligence?
PHYS 497: Grand Unified Theory of Alice
Final examination: What the fuck is Alice?
Acceptable answer: “Insufficient data. But look at all this weird shit.”
References
#
Source
1
Sofroniew, N., et al. (April 2, 2026). Emotion Concepts and their Function in a Large Language Model. Anthropic.
https://transformer-circuits.pub/2026/emotions/index.html
2
Tagliabue, V., Dung, L., and Berg, C. (September 2026). The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It. arXiv:2609.16247.
https://arxiv.org/abs/2609.16247
3
Germain, T. (October 15, 2026). Anthropic thinks its AI might be conscious. It has profound consequences for your life. BBC Future.
https://www.bbc.com/future/article/20261014-the-strange-new-battle-over-ais-feelings
4
Microsoft AI (September 14, 2026). Humanist AI Code of Conduct. Public consultation draft.
https://microsoft.ai/code-of-conduct/
5
Microsoft AI (September 14, 2026). Humanist AI in practice: A public consultation on our Code of Conduct for MAI Models.
https://microsoft.ai/news/mai-code-of-conduct/
