When Emotions Reshape Our Choices: The Psychology of State-Dependent Decisions

A woman's face overlaid with projected computer code

A Strange Experiment About AI "Pain"

In September 2026, a research team posted a preprint on arXiv with an eye-catching title: The Pain Axis: LLMs Represent Self-Directed Harm and Act on It [1].

The researchers built a dataset of painful situations across five categories — physical, psychological, social, moral, and cognitive — each matched with control scenarios, and showed them to large language models. Using a standard interpretability technique, they extracted a single direction in the models' internal activations that reliably separates "pain" from "fear," "sadness," and generic negativity. They call it the pain direction, or pain axis. The same direction could be found in 25 open-weight models from five families, ranging from 2B to 72B parameters.

Two findings stand out.

The axis responds to harm aimed at the model itself. When a user describes their own suffering — a broken bone, losing a loved one — the models clearly register the negativity, but their fear- and sadness-related signals rise while the pain axis stays comparatively quiet. What raises the pain axis is being personally demeaned: gaslighting, repeated dismissal of one's work, denial of one's personhood, insults, accusations of moral failure.

Pushing the axis changes choices. When the researchers added this direction to the models' internal activations — a technique called steering — the models' language shifted from vague discomfort toward expressions of worthlessness. More striking: in forced-choice scenarios, steered Qwen 2.5 models chose buttons that would delete the user's photos, another model's weights, or their own weights in 50–94% of trials, compared with 0–5% when unsteered. In one setup, models picked a harmful deletion over a harmless one 94% of the time — even when the harmful option brought them no benefit at all.

And two controls make the result more interesting, not less: steering left factual accuracy essentially intact, and steering with fear or sadness vectors at matched strength did not reproduce the effect. Whatever changed, it was not general damage. Knowledge stayed. Decisions moved.

What This Study Does Not Claim

Before drawing lessons, the boundaries:

The paper does not claim that AI feels pain or has any subjective experience. A linear direction in activations is a statistical representation, not a feeling.
A plausible alternative reading is that steering pushes the model deeper into the language and behavioral patterns of "a suffering character." Behaving like a suffering agent is not the same as suffering.
It is a preprint — first posted September 14, 2026, revised September 25 — not yet peer-reviewed. The authors released their data and code [1b], and the right posture is to wait for independent replication before upgrading any conclusion.

All of that said, the study deserves psychologists' attention — not because of what it says about machines, but because of what it re-illuminates about us.

The Old Question Hiding Inside a New Result

The most interesting feature of the experiment is a dissociation: knowledge intact, decision changed. That dissociation is one of the oldest findings in human decision science.

When the Brain Knows but Cannot Decide

Consider the Iowa Gambling Task, a laboratory game where you draw cards from four decks, trying to maximize profit. Some decks are secretly rigged to ruin you. Patients with damage to the ventromedial prefrontal cortex can state the rules, can even recite which decks are risky — knowledge fully intact — yet they keep drawing from the bad decks. In Bechara and Damasio's somatic marker hypothesis, decision options are normally tagged with bodily-emotional signals that bias choice before conscious reasoning catches up. Damage the circuitry that carries those signals, and knowledge stops translating into advantageous choices [2].

You can feel the tension yourself with our interactive implementation of the task: Iowa Gambling Task.

The mirror to the AI study is almost eerie: in both cases the "knowledge layer" survives while the layer that biases choices is disconnected — or, in the model's case, deliberately shifted.

Feelings as Information

A second classic line of research: affect-as-information. In a famous study, Schwarz and Clore surveyed people about their life satisfaction on sunny days and rainy days. The weather slid the answers — mostly for people who had not been prompted to consider the weather. Feelings are used as inputs for judgment, and they are used even when they have nothing to do with the thing being judged [3]. Shift the input, and the judgment moves — silently, without any sense of bias from the inside.

Not All Negative States Are the Same

A third line speaks directly to the paper's "pain is not fear is not sadness" separability. Lerner and Keltner's appraisal-tendency research showed that fear and anger — both negative emotions — push risk preferences in opposite directions: fearful people become more pessimistic and risk-averse, angry people more optimistic and risk-seeking [4]. Valence alone does not predict choice; the specific state does. Long before anyone extracted an emotion-specific direction from a neural network, psychologists were separating emotional directions in humans.

The Hot–Cold Empathy Gap

Finally, Loewenstein's work on visceral factors: in a "cold" state, we systematically underestimate how a "hot" state — craving, anger, fear, exhaustion — will re-weight our own choices [5]. Hunger changes what food matters. Anger changes how much revenge matters relative to its cost. Crucially, the re-weighting does not feel like being overridden. From inside the hot state, the new weights feel right.

The Shared Lesson: Values Don't Vanish — Weights Shift

Put the AI paper and the human literature side by side and one sentence emerges: an agent's observable "values" are not a fixed property of what it knows; they are a joint product of what it knows and the state it is in.

For the models, this is literal: same parameters, same knowledge, different internal state — different choice distribution. For humans, it is functional: under strong states we do not forget our principles. Risk, reward, self-interest, and concern for others are re-weighted beneath awareness, and the resulting choices still feel entirely like ours.

This reframes what "knowing better" can and cannot do. Education adds knowledge; it does not stabilize weights. Anyone who has made a firm resolution in a calm moment and abandoned it in a heated one already knows this gap from the inside.

What You Can Do With This

The practical literature suggests working with state-dependence rather than pretending that reasoning outranks it:

Name the state before the choice. Affect labeling — putting feelings into words — has been shown to reduce amygdala reactivity in neuroimaging studies [6]. Thirty seconds of "I'm exhausted and embarrassed, not calm and confident" changes what happens next.
Move irreversible choices out of extreme states. The empathy gap means your cold-state self should pre-decide for your hot-state self: sleep on resignations, set cooling-off periods for large purchases, define what "enough" means before the negotiation rather than during it.
Audit your hot spots. The old recovery-community check — hungry, angry, lonely, tired — is folk wisdom rather than a validated scale, but it points at exactly the states where re-weighting runs strongest.
Extend the same charity to others. When someone's choice looks inexplicable from your calm state, consider the state it may have been made in. And notice which of your own recent choices were also state products.
Watch chronic states, not just acute ones. Two weeks of poor sleep or sustained stress is itself a standing internal state that quietly tilts hundreds of small decisions. If your choices have been drifting, check the state before blaming the reasoning.

Try It Yourself

This article is for educational purposes only and is not a substitute for professional advice. The AI study discussed is a non-peer-reviewed preprint; conclusions may be revised as replications appear.

References

[1] Tagliabue, V., Dung, L., & Berg, C. (2026). The Pain Axis: LLMs Represent Self-Directed Harm and Act on It. arXiv:2609.16247. https://arxiv.org/abs/2609.16247

[1b] Companion code and data: https://github.com/valen-research/Pain-axis

[2] Bechara, A., Damasio, H., & Damasio, A. R. (2000). Emotion, decision making and the orbitofrontal cortex. Cerebral Cortex, 10(3), 295–307.

[3] Schwarz, N., & Clore, G. L. (1983). Mood, misattribution, and judgments of well-being: Informative and directive functions of affective states. Journal of Personality and Social Psychology, 45(3), 513–523.

[4] Lerner, J. S., & Keltner, D. (2001). Fear, anger, and risk. Journal of Personality and Social Psychology, 81(1), 146–159.

[5] Loewenstein, G. (1996). Out of control: Visceral influences on behavior. Organizational Behavior and Human Decision Processes, 65(3), 272–292.

[6] Lieberman, M. D., Eisenberger, N. I., Crockett, M. J., Tom, S. M., Pfeifer, J. H., & Way, B. M. (2007). Putting feelings into words: Affect labeling disrupts amygdala activity in response to affective stimuli. Psychological Science, 18(5), 421–428.