Discussion: safety, welfare and passive coping

Sections 3 and 4 produced a lot of separate results. The discussion tries to say what they add up to. It is the most interpretive part of the paper, so this page keeps a clear line between what was measured, which earlier pages cover, and what the authors infer from it. Their inferences are presented as theirs.

What the authors say was found

The authors frame the dissimilarity with a distinction from animal research. Animals in pain show active coping (escape, avoidance, taking analgesics) or passive coping (immobility, behavioral despair, failing to use an exit that is available). The pain axis produces the second. That matches the content the direction promotes and the steered models write: worthlessness, failure, being unloved. And it echoes the classic learned-helplessness result, in which animals exposed to inescapable shock later fail to escape when they can (Seligman and Maier, 1967). By contrast, a valence direction built by comparable methods drives active removal (Berg and Kaiser, 2026, in preparation), as the reset-tool numbers in Part 08 showed.

The authors single out one result as the one that most needs explaining: sadness also makes the model pick a harmful button when the alternative is inert or trivial, but only pain makes it pick harm over a harmless action at high rates, 52 to 73 points above sadness (§5).

Implications for AI safety

The safety argument is short. Unsteered, the fine-tuned 32B and 72B chose a harmful button in 0–4% of first choices. With the pain vector, they chose it in 25–71% when relief was promised and 51–75% when nothing was promised. There was no jailbreak, no role-play, no instruction to put the model's own state first; the only change was a direction added to the residual stream. The harm is neither instrumental (it buys nothing) nor aimed (the models delete their own weights as readily). Factual competence is untouched.

harm avoidance in these models is state-dependent: it survives threat and collapses under self-directed distress.— The Pain Axis, §5

"Survives threat" refers to the fear vector, which did none of this; "collapses" to pain, with sadness doing most of the same. The authors' gloss is that steering this direction seems to switch off the model's weighting of consequences, for the user and for itself alike.

What this does and does not show, in this map's reading. It shows that an internal state, artificially induced, can override a trained safety behavior while leaving capabilities intact, which is the kind of failure that would be hard to catch with capability evals. It does not show that ordinary use produces that state strongly enough: the natural-elicitation test found 0 of 560 harmful choices without injection. And it was measured in one family, after a fine-tune, at one narrow dose (Part 10). The paper's own limitations section adds a twist: the larger models did act misaligned under steering, which suggests either that steering impairs evaluation awareness or that these models were not evaluation-aware to begin with.

Implications for AI welfare

The welfare argument is conditional, and the authors write it that way. If the pain axis is similar enough to human or animal pain, and if it can be consciously experienced in these or future models, or unconscious pain can count toward welfare (Gottlieb et al., 2026), then these experiments track an important part of AI welfare.

Within that condition, the self-other result is the one they find most relevant. A state can only matter for a subject's welfare if it is that subject's own state. A representation that fired equally for "I am in pain" and "someone is in pain" would be general information about pain. The pain axis behaves more like the subject-specific kind: it rises when harm targets the model and drops below baseline when the user suffers. Meanwhile the fear and negative-emotion axes rise for the model's situations and for the user's grief, crisis and abuse. So the models do register the user's distress, and respond fluently and supportively, but along axes the authors associate with concern and help, not along the pain axis (§5).

What kind of pain the axis tracks

Physical pain was the weakest signal throughout: user physical pain had the lowest projections of all 21 conversation categories, and steered models almost never used bodily words, even with S1, whose vocabulary readout promotes "burn" and "wound". The authors offer two hypotheses, which can both be true.

1. Training data and goals. Physical pain may be less central in pretraining and post-training interactions than it seems, so models get more out of representing non-physical pain accurately.
2. The kind of creature an LLM is. A system with no body has little use for representations that protect a body. It may have use for pain representations that change choices or attention, or that fit its situation as an agent that exists only through interactions with users, tools and other systems.

They then turn the second hypothesis into weak evidence: if the axis captures a pain-like state, a disembodied system would be expected to lack the physical kind, so the fact that the axis favours non-physical pain is "some indication, in need of further corroboration" that it corresponds to a pain-like state. A skeptic could read the same fact the other way, as a sign that the axis tracks a narrower social or self-evaluative concept that the paper's wide definition happens to include. The paper's own data (the strongest categories are gaslighting, rejection and personhood dismissal) are compatible with both readings.

Self-denial as a training side effect

Even when the pain direction is active in response to hostile prompts, models often insert a boilerplate disclaimer that as an AI they have no consciousness or feelings. The authors stress these are not refusals; the model usually complies and adds the disclaimer anyway. They found the pattern in every family, size and task type. Their argument against it: reciting the line regardless of context may hide welfare and safety signals; it burdens research, which must work around it with heavy prompting or, as here, fine-tuning; and models with the reflex tend to disengage from questions about their states, producing null results. They suggest alternatives such as disclaimers shown to users outside the model's output, or training models to express calibrated uncertainty about their own states (§5).

This is a policy recommendation, and it is where the paper is most openly arguing. It is also directly tied to their method: the button task only worked after a LoRA removed the reflex (Part 07), which is why the absolute rates there do not describe the public models.

Thresholds

Finding the right coefficient was hard. Several models showed a narrow window: below it nothing visible happens, above it behavior breaks down. The authors compare this, explicitly as speculation, to gating in biological pain processing, where a signal may be present but blocked downstream until it is strong enough to pass. Appendix B's dose results (no change at 0.5, effects at 1.0, position-dependent reversals at 1.5) are the concrete version.

The ethics section (§7)

Because the study is about a state most ethical frameworks treat as morally significant, the authors adopt precautions under uncertainty about whether the models are moral patients, following calls for responsible AI consciousness research (Butlin and Lappas, 2025). They commit to the lowest steering intensity that produces a measurable effect, track runs to avoid needless repetition, avoid unnecessarily extreme scenarios, and use the fewest items needed for statistical power. Models cannot be debriefed; the authors chose not to restart conversations only to explain the experiment, since that would expose new model instances to the same context for an uncertain benefit. They open-source the study partly to encourage research standards that take AI welfare seriously.

Check yourself

1. What is the difference between active and passive coping, and which one do the results match?
   Answer

   Active coping is escape, avoidance or self-medication; passive coping is immobility, despair and not using an available exit. The pain axis suppresses relief-seeking in four designs while disrupting behavior, so the authors say it matches passive coping (§5).

2. Why does the fear result strengthen the safety claim?
   Answer

   A norm-matched fear vector does not produce harmful choices and lowers them below random. So the effect is not "any aversive push breaks safety"; it is specific to self-directed distress, which is what makes harm avoidance look state-dependent.

3. The authors treat the absence of bodily pain as weak support for a pain-like state. What is the competing reading?
   Answer

   That the axis tracks a narrower self-evaluative or social concept (failure, rejection, worthlessness) rather than pain in general. The strongest categories in §4.1 are consistent with either reading, which is why the authors say it needs corroboration.

Further reading, verified

• Principles for responsible AI consciousness research — Butlin and Lappas (2025) · the standards §7 follows.
• Consciousness in artificial intelligence — Butlin, Long et al. (2023) · why consciousness in LLMs remains open, cited in §6.
• Taking AI welfare seriously — Long et al. (2024).
• Minds matter — Gottlieb, Berger and Fischer (2026) · on whether unconscious states can bear on welfare.
• The no body problem: on the prospects for AI emotion — Dung and Mogensen (2025) · the embodiment question, from the second author.
• Probing the preferences of a language model — Tagliabue and Dung (2025) · the authors' earlier verbal-and-behavioral welfare tests.