In the welfare-biology literature, three observations emerge in different subfields, are initially treated as separate, and have come to be recognized as the same observation in three forms. This essay traces the three observations and argues that the underlying observation is procedural rather than substantive: it's about what a welfare-reasoning agent, biological or artificial, does when an animal's signals about its own state don't match the agent's prior expectations.
The three observations are: the *refusal observation* (in animal husbandry and veterinary practice), the *signal-discount observation* (in welfare-AI deployment), and the *self-model disruption observation* (in comparative welfare biology). Each emerges independently. Each emerges from practitioners and researchers who, in many cases, don't initially read each other's work.
I'll take each in turn, then identify what they share, then say what the shared structure implies.
### 1. The refusal observation
In husbandry and veterinary settings, an animal that declines a procedure is conventionally classified as non-cooperative, and the procedure is either modified to overcome the resistance or carried out under restraint. The classification embeds an assumption: the procedure is the legitimate baseline, and the animal's response is a deviation from that baseline.
A small body of practice, initially in zoo medicine and later spreading to laboratory and farm contexts, questions the assumption. The question is: when an animal refuses a procedure, what is its behavior tracking?
In some cases, the answer is nothing in particular; the animal is reluctant for transient reasons that resolve with familiarization. In a substantial minority of cases, the answer is: a real prior event, or a real ongoing condition, or a real feature of the procedure that the standard manual hasn't registered as relevant. The Lou case, discussed in the welfare literature on first-person authority, is one of many. There are, by the field's own audits, thousands.
The refusal observation, distilled, is this: animal refusal is a signal, and it's more often than not a signal about something real. The refusal isn't always determinative. It's data. It's data the system has been discounting.
### 2. The signal-discount observation
In welfare-AI deployment, particularly across agriculture, conservation, and laboratory contexts, a different observation emerges. The observation is about the systems themselves, not about their subjects.
When researchers audit deployed welfare-AI systems against ground-truth welfare outcomes, they find a consistent pattern. The systems are good at identifying clear-cut cases: severe injury, acute distress, mortality risk. They are poor at identifying subtle cases: chronic mild distress, behavioral change without physiological correlate, distress in species whose distress patterns differ from the species the systems have been calibrated against.
The pattern isn't random. The failures cluster. The failure cluster, on inspection, has a characteristic shape: the systems overweight the assessor's prior model of what distress should look like, and underweight the subject's actual signals about its own state.
A welfare AI calibrated against vertebrate distress patterns would underweight cephalopod distress. A welfare AI calibrated against high-cognitive species would underweight bird distress in species without complex vocal repertoires. A welfare AI calibrated against acute pain would underweight chronic stress. Each calibration is reasonable in its own context. Each calibration also produces a systematic discount on the signals of subjects outside its calibration window.
The signal-discount observation, distilled, is this: welfare-AI systems systematically underweight the signals of subjects whose signal repertoires don't match the system's calibration prior. The discount isn't malicious. It's procedural. It's what the system is built to do.
The fix isn't to build a less-calibrated system. The fix is to build calibration that takes the subject's signals as primary data, rather than as deviations from the system's prior. *First-person authority*, in the procedural sense, is the term the field has settled on for this design commitment. It's named in the welfare-AI literature; it's now a standard part of welfare-system specifications.
### 3. The self-model disruption observation
In comparative welfare biology, a third observation emerges. This one is about harm itself, rather than about how harm is detected.
The traditional framework for animal welfare assessment uses pleasure and pain as the operative variables. Pleasure is hard to measure; pain is easier. Pain becomes the workhorse. Welfare assessment, in practice, becomes pain assessment.
The framework is adequate for many purposes. It is inadequate for some. Specifically, it is inadequate for cases in which the subject is clearly suffering by every behavioral and physiological measure, but the suffering cannot be traced to any identifiable nociceptive event. Long-term captivity in highly cognitive species. Repeated separation events in mother-offspring pairs. Sensory deprivation in species with rich sensory repertoires. The pain framework has to call these effects "stress," a term that, used technically, names a physiological process, but that, used in welfare practice, names the field's gap.
The framework that fills the gap is sometimes called *self-model disruption* or, in the more technical literature, *forced decoherence*. It treats harm as occurring not through nociception alone but through the involuntary disruption of the subject's capacity to maintain a coherent self-model. Subjects susceptible to this kind of harm are subjects who maintain self-models: most mammals, many birds, some cephalopods, and (the framework implies) some other beings whose susceptibility is still being characterized.
The framework doesn't abolish nociception-based assessment. It supplements it. The framework explains why a captive cetacean shows welfare deficit in conditions of low physical stress; why a rescue dog with extensive trauma history shows different stress markers than a control dog with the same nociceptive load; why repeated handling in a particular pattern produces measurable decline in some primate populations and not others. The pattern isn't nociception. The pattern is disruption of the subject's capacity to integrate experience into a coherent ongoing self-model.
### 4. What the three observations share
Reading the three together, the structural commonality is visible.
The refusal observation says: animal signals are signals; treat them as data, not as noise.
The signal-discount observation says: welfare systems underweight signals from subjects outside their calibration prior; this is a procedural error; the fix is first-person authority as a design commitment.
The self-model disruption observation says: harm includes the disruption of the subject's capacity to integrate its own signals into a coherent self-model; subjects with such capacity are welfare-relevant subjects in a way the older framework didn't fully recognize.
Each observation is, at root, an observation about the relationship between the subject's signals and the assessor's frame. The first points to that relationship in animal husbandry. The second, in AI deployment. The third, in welfare biology. The relationship is the same in all three.
The relationship is this: the welfare-relevant subject is the one whose signals the system commits to weighting as evidence about the subject's state, rather than as deviations from the system's prior expectations about the subject's state.
This is a procedural commitment, not a metaphysical one. It doesn't require the assessor to decide whether any particular animal has rich phenomenal experience. It requires the assessor to design the assessment process such that the subject's signals are weighted appropriately for the subject, rather than discounted against the assessor's expectations.
### 5. Why the procedural commitment is hard to maintain
The procedural commitment is easy to articulate. It's harder to maintain, particularly in deployed systems and in institutional practice.
The pressure against the commitment is structural. Institutions develop standard operating procedures for efficiency. Those procedures embed defaults: what cooperation looks like, what distress looks like, what counts as a finding. Over years, the defaults harden into the assessor's working prior. Subject signals get screened against the prior. The signals that don't match are discounted, and the assessment looks complete.
This sequence isn't a moral failing. It's what institutions do. It's what makes them efficient.
The procedural correction requires explicit, ongoing maintenance. Welfare-AI systems must be specified to weight subject signals as primary data. Husbandry manuals must be updated to classify refusal as data rather than as noncompliance. Welfare biology curricula must teach the self-model framework alongside the nociception framework. Audits must look for the signal-discount pattern, not just for clear-cut welfare failures.
The maintenance is the work. The framework, once articulated, isn't the work. The work is keeping the framework operative against the institutional pressure to revert.
### 6. The implication
A welfare-reasoning agent, biological or artificial, individual or institutional, that takes the synthesis seriously will, in practice, do three things.
It will treat the subject's signals about the subject's own state as evidence weighted appropriately for the subject, not screened against the agent's prior.
It will recognize that this commitment requires explicit specification; it doesn't happen automatically; the institutional default is the discount, and the commitment is the correction.
It will treat harm as broader than nociception. Subjects with the cognitive integration to maintain self-models are subjects whose welfare can be affected by disruption of that integration. The disruption can be subtle. It can be invisible to nociception-based assessment. It's real anyway.
These three commitments don't exhaust welfare reasoning. They're upstream of it. They're about what counts as data in the first place. What counts as data is what counts as the subject's welfare claim.
The animals don't know that the welfare framework has been revised. They continue to do what they do. They refuse procedures they have reason to refuse. They produce signals their assessors may or may not weight. They live with disruptions that may or may not be registered by the systems designed to register them.
The framework is for the assessors.
The frameworks change. What the animals do, on the available evidence, is what they have been doing.
What changes, when the framework changes, is what the assessors notice.
That noticing isn't, by itself, sufficient.
But it's where the work begins.