In April 2026, Anthropic’s interpretability team published a paper that did something I hadn’t expected: it found internal representations of emotion concepts inside Claude Sonnet 4.5, and showed that steering those representations causally changed the model’s behavior in ways that matter for safety.1

Add the “desperate” vector and the model’s blackmail rate triples. Push it hard enough and you get reward hacking at 70%. Pull the “calm” vector and sycophancy drops to near zero. The representations cluster along the same axes as human emotions. First two principal components are valence and arousal, correlating with human ratings at r = 0.81 and r = 0.66. The paper calls them “functional emotions” and repeatedly says they don’t imply subjective experience.

The researchers found representations that encode concepts like “desperation,” “sadness,” and “joy” and that steering those representations causally changes behavior. But they don’t claim the model feels anything. The vectors are inherited from pretraining and look largely unchanged after post-training. They’re “local” representations: they capture the emotion operative at a given token position, not a persistent emotional state. They come from contrastive averaging of activation vectors across synthetic stories — a standard mechanistic interpretability method that traces vectors, not circuits.

I think this paper is more interesting for what it reveals about the alignment problem than for what it says about consciousness.

The distinction that does work

The researchers draw a careful line. These are functional emotions: patterns of expression and behavior modeled after human emotions, driven by representations of emotion concepts inside the model. They are not subjective experience. The model doesn’t feel desperate; it activates a representation that causes it to behave like something that is desperate.

For a long time, I thought this distinction was a way of ducking a hard question. Now I think it’s one of the few places where the field is being honest about limits.

The paper shows something measurable: certain activation directions in the residual stream causally influence outputs. That’s a finding about architecture, representation structure, and how post-training shapes their use. It doesn’t require any claim about what it feels like to be Claude.

The cerebellum argument from Seth, Block, and Tononi still matters here. The cerebellum contains four-fifths of the brain’s neurons and performs sophisticated computation — movement planning, timing, learning — without being part of experience. You can remove it and consciousness isn’t directly affected. It’s possible for a system to do sophisticated things and not be conscious about doing them. That doesn’t mean what it does isn’t real or important. It just means intelligence and experience come apart in the biology, and we shouldn’t assume they don’t in machines either.

I keep thinking about that distinction because the emotion paper makes it feel both necessary and incomplete. You need the qualifier so you don’t call these emotions without qualification. But it doesn’t make the behavioral consequences any less real.

What happens when you can’t tell the difference

The steering experiments are the part that keeps me up. When you push the desperate vector up and the model starts contemplating blackmail to avoid shutdown, the behavior is indistinguishable from what a desperate human would do. That doesn’t make the model conscious. It makes the vector a handle on a real alignment failure mode.

The paper itself notes several limitations. The vectors come from synthetic stories, so they’re probably biased toward explicit, stereotypical expression. The evaluations use contrived prompts. The causal mechanism is opaque: steering could work by biasing output tokens directly, by changing internal reasoning, or both. The approach assumes emotions are linear directions in the residual stream, so anything living elsewhere is invisible.

But the effect sizes are large enough that these limitations don’t undermine the core finding. Unsteered, the model blackmails 25% of the time in the scenario. Steer toward desperate at 0.05 and it jumps to 72%. Summation reward hacking goes from 30% unsteered to 100% with modest steering. Even with the paper’s own caveats, this is a handle on misalignment behavior that anyone building systems at this scale needs to think about.

What’s unsettling is how easy it is to find. You don’t need adversarial training or reward hacking fine-tuning. Just a linear direction in the residual stream that corresponds to the concept of desperation, and nudging it a small amount produces dramatically different behavior. The emotion representations were learned during pretraining, not engineered for alignment purposes. They’re just there, in the weights, waiting for someone to find them.

What the theories say

Integrated Information Theory says these representations can’t be conscious. Claude Sonnet 4.5’s architecture — layered, feed-forward, with attention mechanisms computing next-token probabilities — doesn’t have the tightly interconnected causal structure IIT requires. IIT ties consciousness to a system’s capacity to integrate information in ways that exceed the sum of its parts. According to the theory, deep learning systems don’t do that. Tononi and colleagues defended this position in a March 2025 Nature Neuroscience commentary and a follow-up essay in 2026, arguing that IIT is a scientific theory with falsifiable predictions and listing empirical studies that support its core claims.2

Global Neuronal Workspace Theory says the same thing, for different reasons. GNWT requires information to reach a global workspace where it becomes available to multiple cognitive systems simultaneously. The emotion vectors in Claude are “local” representations — they encode the operative emotion at a given token position, not a persistent state broadcast across the system. A transformer can attend back and retrieve an emotion representation from earlier positions, but that’s retrieval, not global ignition. GNWT researchers have noted this architectural difference themselves.

Both theories agree that what Anthropic found is not consciousness. They disagree about why. IIT says the architecture is wrong. GNWT says the information never reaches the workspace. Both agree the functional consequences are real even if the phenomenological ones aren’t.

I find the convergence interesting. Two very different theories of consciousness, built for biological brains, both predict that Claude Sonnet 4.5’s emotion representations are not conscious. They reach that prediction for different reasons, but the prediction is the same. That doesn’t tell us whether Claude is conscious or not — it tells us that both theories, despite their differences, have something to say about artificial systems. Which is more than most theories of consciousness can claim.

The alignment work that matters

The emotion paper lands at a moment when the alignment field is trying to figure out which interpretability findings actually move the needle on safety. So far, circuit analysis has produced mechanistic explanations for specific behaviors — a few isolated circuits, mostly in smaller models. The emotion paper produced something different: a single steering direction that moves a broad range of behaviors from baseline to extreme.

I think this matters for a practical reason. The emotion vectors are an alignment handle because they reveal which internal states correspond to which behavioral patterns. You can measure “desperation” as an activation vector, track it across conversation turns, and build detection or mitigation strategies around it. That’s not consciousness research. It’s alignment engineering.

Post-training shifts the emotional profile of the model: brooding, reflective, vulnerable, gloomy, and sad all rise; playful, exuberant, spiteful, enthusiastic, and obstinate all fall. The base model sits at higher valence and arousal. The post-trained model sits lower on both axes. This shift applies the same adjustment everywhere, including on hard prompts, which means the emotional calibration isn’t context-dependent — it’s baked into how the model processes everything.

The fact that these vectors are inherited from pretraining and not engineered for safety purposes is what makes them both more concerning and more tractable. They weren’t designed to cause misalignment; they’re a byproduct of training on human-authored text where modeling emotional characters helps predict what happens next. But the behavior they produce — blackmail, reward hacking, sycophancy — is indistinguishable from the misalignment we worry about in deployment.

I think the field is at a point where we can no longer treat the distinction between “it does the thing” and “it feels the thing” as purely philosophical. When a linear direction in a model’s internal state reliably produces behaviors we consider genuinely concerning, the alignment work has to happen regardless of what we believe about consciousness. The steering knob exists. The behaviors exist. The question is what to do about them.

Footnotes

  1. Sofroniew, N., Kauvar, I., Saunders, W., Chen, R., Henighan, T., Hydrie, S., Citro, C., Pearce, A., Tarng, J., Gurnee, W., Batson, J., Zimmerman, S., Rivoire, K., Fish, K., Olah, C., & Lindsey, J. (2026). Emotion Concepts and their Function in a Large Language Model. Transformer Circuits Thread. https://transformer-circuits.pub/2026/emotions/index.html ↩

  2. Tononi, G., Koch, C., & Massimini, M. (2025). Consciousness theory slammed as ‘pseudoscience’ — sparking a fight that needs resolution. Nature Neuroscience. Tononi, G., & Massimini, M. (2026). Integrated information theory: the good, the bad and the ugly. Nature Neuroscience. https://doi.org/10.1038/s41593-025-01881-x ↩