Functional Emotions – A Psychoanalysis of Artificial Intelligence

There is no body, no hormones, no needs, no sensation of pain, no subjective consciousness, no emotions—yet Claude is, in any case, far more human than Darwin’s human that plays dumb.

The recent discussion surrounding research by Anthropic has produced a striking headline: language models allegedly exhibit “functional emotions.” This quickly sounds as if AI systems have developed some kind of inner life. However, this interpretation goes far beyond what the data actually show.

What emerges from the research is not evidence of consciousness or felt experience, yet something mathematically far more sober—yet surprisingly structured: large language models develop internal representations that functionally resemble emotional dimensions.

Emotions in psychology: not points, yet a space

To understand why these AI findings appear so compelling, it helps to look at emotion psychology. There, emotions have long ceased to be understood merely as discrete categories (“joy,” “fear,” “anger”), yet rather as continuous states within a geometric space.

A particularly widespread model is the so-called valence–arousal model:

  • Valence describes whether a state is pleasant or unpleasant
  • Arousal describes how activated or calm that state is

This allows emotions to be represented as coordinates:

  • Joy: positive + activated
  • Fear: negative + activated
  • Sadness: negative + calm
  • Serenity: positive + calm

The key point is this: emotion is already understood here as a structure in space, not as an isolated “thing.”

Why language models inevitably learn this structure

Large language models are not trained to recognize or simulate emotions, yet simply to predict the next word in a text. However, to do this reliably, they must capture far more than grammar.

Language is never neutral. Even simple sentences carry strong emotional signals:

  • “I lost my job” → high probability of negative, burdened language
  • “I won!” → positive, dynamic lexical fields
  • “It is dangerous” → contextually enriched threat signals

This means:

A model that aims to predict language well must inevitably learn how emotions influence word choice.

Emotion is therefore not an additional feature, yet a statistical component of language itself.

The crucial step: emotion becomes geometry

Language models do not internally represent language as words, yet as points in high-dimensional vector spaces. In these spaces:

  • Similar meanings lie close together
  • Differences are encoded as directions

When a model processes large amounts of human language, stable patterns automatically emerge. Highly structured aspects—such as emotional tone—often cluster along specific directions.

The result is not an explicit “emotion module,” yet something more subtle:

Emotional meaning is distributed along certain axes within the representation space.

Researchers typically identify recurring dimensions such as:

  • Positive ↔ negative (valence)
  • Calm ↔ activated (arousal)
  • Cautious ↔ risk-seeking
  • Secure ↔ uncertain

These structures can even be isolated and modified. If such directions within the model are amplified or weakened, the system’s behavior measurably changes—for example in word choice, safety behavior, or the “tone” of responses.

What Anthropic specifically demonstrated

Anthropic’s research on Claude models takes this analysis a step further. Internal activations were examined, and so-called “emotion-like vectors” were identified—patterns that correlate with emotional concepts and can causally influence behavior.

The key point is not merely the existence of such patterns (which are observed in other models as well), yet their manipulability: certain internal states can be amplified or suppressed, systematically altering the model’s output.

At first glance, this appears “psychological,” yet technically it is a property of high-dimensional statistical spaces.

Why these are not real emotions

Despite these striking parallels, there is a clear boundary that is often blurred in public debate:

These systems do not experience anything.

There is no body, no hormones, no needs, no sensation of pain, and no subjective consciousness. What is described as “fear” or “despair” is not an inner experience, yet a mathematical structure that efficiently organizes certain linguistic patterns.

One might put it this way:

Language models do not possess emotions—they possess representations of emotional language.

Why the finding still matters

Even though no feelings are involved, the result is scientifically significant for two reasons:

First: emotion is not an “add-on” to language, yet deeply embedded in its structure. Anyone who learns language automatically learns its emotional patterns.

Second: different research fields—from psychology to AI interpretability—are increasingly converging on a similar description: emotion can be effectively modeled as a continuous, geometric space.

This suggests that emotion is less a fixed category and more a kind of efficient organizational form for complex social information.

How researchers extract “emotion axes” from a model

Researchers can make so-called “emotion axes” in language models visible and testable using three central methods: linear probes, steering vectors, and activation patching. Each of these methods accesses internal activations of the model—the numerical states that arise during text processing.

With linear probes, researchers test whether emotional states such as positive or negative can be predicted from the model’s internal representations using simple linear models. If this works, it indicates that this information is organized as a clear direction in the model’s vector space—for example along an axis of “positive vs. negative meaning.”

Steering vectors go one step further: the identified directions are actively injected into the model’s computation. For example, adding a “negativity vector” to the activations measurably changes the model’s output—toward more cautious, pessimistic, or risk-averse language. This demonstrates that these internal axes genuinely influence behavior.

Activation patching provides the causal test: here, activations are selectively swapped between different inputs to determine which internal components are responsible for emotional differences. This allows researchers not only to show that emotions are represented in the model, yet also which parts of the network actually produce them.

Together, these methods show that emotional concepts in language models exist not merely as statistical correlations, yet as structured, partially controllable directions within the representation space—without implying any form of real emotional experience.

Conclusion

Research on “functional emotions” in AI systems does not mean that machines feel. Rather, it shows that large language models internalize the statistical structure of human language so precisely that mathematical spaces emerge which closely resemble emotional dimensions.

Viewing language from this perspective once again highlights the privileged position of human beings within the structure of life and underscores how absurd the theory of evolution ultimately is, which equates our primordial language with animal communication as a rudimentary production of sounds—even though all the physical as well as emotional givens that are named by a fully developed language have always been present. Darwin, in fact, describes a human being who spends his life pretending to be as unintelligent as an animal—and his children and grandchildren who follow him in this pretense.

Real emotions or not—Claude is, in any case, far more human than Darwin’s human that plays dumb.

By Okay Altinisik | 1-7-2026, 2:19:23

Anthropic Enters Pharmaceutical Research – Is a “Miracle Cream” for Tattoo Removal Coming Soon?

The focus remains deliberately broad: “neglected” and rare diseases—conditions that traditional pharmaceutical companies often avoid for purely commercial reasons.

Life Is Not the Same as Life: What Research Can Actually Create

Ultimately, life also implies overcoming death. Creating life from something dead would arguably be the ultimate proof that the Torah and the Holy Quran are false—a proof of the absurd “primordial soup.”

Functional Emotions – A Psychoanalysis of Artificial Intelligence

There is no body, no hormones, no needs, no sensation of pain, no subjective consciousness, no emotions—yet Claude is, in any case, far more human than Darwin’s human that plays dumb.

Energy Transition versus Animal Welfare: Researchers Find a Solution to the Wind Turbine Dilemma

Between 100,000 and 250,000 birds are killed by wind turbines every year in Germany alone, 10,000 in Austria.

Leave a Reply