Update, August 2026: We have revised the title and some terminology in this post to place greater emphasis on the Affective Ceiling Approach, which is the underlying scientific contribution, rather than on its current implementation as a Custom GPT. We have also renamed the tool the Affective Ceiling GPT.
These are changes in framing and naming. They do not alter the post’s substantive claims, evidence, reasoning or conclusions.
The Custom GPT is an experimental and temporary way of making the approach available. The Affective Ceiling Approach is independent of that particular platform and can be applied by researchers directly or implemented through other, more capable tools as they are developed.
The user inputs a target taxon. The tool then determines, where necessary, two related but distinct scopes: one for the sentience gate and one for the affective-capacity analysis.
You may be interested in the article "Ethics Without Sentience: Facing Up to the Probable Insignificance of Phenomenal Consciousness" by François Kammerer. Here is a summary. Relatedly, you may like this and this very short posts by Keith Frankish. I agree with Keith that phenomenal consciousness does not exist. In any case, even if it exists, I agree with François it is not significant. Phenomenal consciousness as traditionally defined is not falsifiable (or, to be precise, it is traditionally defined such that one can only know about their own consciousness at the moment). In practice, one always has to come up with empirical proxies for it, but I feel these are largely arbitrary if it is not falsifiable. So I believe it would be better to focus just on what animals can do (as in your affective-capacity analysis), and not on assessing phenomenal sentience.
Thanks, Vasco — this is a very interesting challenge, and the pieces by Kammerer and Frankish are highly relevant.
I am not sure, however, that our use of sentience depends on the traditional conception of phenomenal consciousness that you are criticizing. In the Welfare Footprint Framework, the relevant question is whether an organism can have positive or negative affective states—states that can make its situation better or worse for it. We try to infer those states from converging behavioural, physiological, neural, pharmacological, cognitive, evolutionary, and ecological evidence.
This is why I would hesitate to focus only on what an organism can do. Behaviour and functional capacities are important evidence, but they are not themselves welfare. A biological or artificial system may avoid damage, learn, prioritize certain outcomes, and change its behaviour without anything necessarily feeling good or bad for it. The distinction between affective and non-affective control is therefore part of the question we are trying to investigate, rather than something we can simply set aside.
I would also hesitate to describe the empirical proxies as arbitrary. They are indirect and uncertain, and their interpretation depends partly on one’s theory of consciousness. But some proxies are clearly more informative than others. Their value can be assessed through causal and comparative evidence, experimental interventions, and whether different classes of indicators point toward the same conclusion.
That said, your comment suggests a useful improvement to the tool. We could define the sentience gate more carefully and make clearer that it is not trying to identify an ineffable property disconnected from observable biological processes.
The workflow itself would remain the same. First, we ask whether the available evidence supports the existence of any positive or negative affective states. Only then, conditional on that assessment, do we ask how intense those states could plausibly become. These are related but distinct questions: the sentience gate concerns whether there is any affective welfare subject at all, whereas the ceiling analysis concerns the upper limit of that subject’s affective range.
Some of the same evidence may inform both stages, but the inferential targets are different. We could make this distinction clearer in the article and in the tool, while also separating the observable evidence from the further inference about affective experience at each stage.
Someone sympathetic to Frankish’s view might interpret the same evidence without appealing to phenomenal properties as traditionally conceived. But for the WFF, we still need to distinguish systems that merely process information and regulate behaviour from organisms for which conditions can plausibly feel better or worse.
Thank you for drawing our attention to this. I think it may help us clarify both the article and the tool without changing the basic sequence of the analysis, and without requiring us to settle the broader philosophical debate before making practical progress.
Here is a thought experiment that may be helpful. Imagine you had all the empirical/physical evidence about an animal, including all the "behavioural, physiological, neural, pharmacological, cognitive, evolutionary, and ecological evidence". There may be practical limits to how much one could know, but I am imagining a scenario where empirical/physical uncertainty was fully eliminated. Would you be able to say whether the animal is sentient or not under the Welfare Footprint Framework (WFF)? If the probability of sentience is just supposed to capture empirical uncertainty under WFF, it would be either 0 or 1? For it to be higher than 0, and below 1, I think it would have to capture metaphysical or definitional uncertainty. If so, would it be better for WFF to cover just empirical uncertainty?
Under WFF, affective welfare either exists or does not exist at all. However, how can one justify such a binary distinction? Each type of empirical evidence, including "behavioural, physiological, neural, pharmacological, cognitive, evolutionary, and ecological evidence", exists on a spectrum. So I find it very counterintuitive that we could map empirical evidence to affective welfare either existing (sentience) or not existing at all (lack of sentience). I believe this would require assuming the existence of a sharp surface in the evidential space separating affective welfare existing from it not existing at all (even if there is uncertainty about the surface location), which is what feels arbitrary to me.
I should first clarify whose position is being discussed: The Welfare Footprint Framework itself does not prescribe a particular theory of sentience or affective capacity. It treats positive and negative affective experiences as the core of welfare, but it remains open about the precise biological mechanisms that produce them and about how affective capacities should ultimately be compared across species. The ideas in this series of posts are therefore better understood as our provisional contribution to this problem, developed largely for the practical purpose of making interspecific Welfare Footprint comparisons possible. They should not be presented as an official WFF position on whether sentience is categorical or graded.
My own personal view is that the biological capacities associated with sentience can vary continuously, while the minimal presence of affective experience is categorical. Nociception, memory, learning, integration and considerable information-processing capacities may all have existed in pre-sentient ancestors. My hypothesis is that, at some point, affective states became part of the control system, providing a common currency through which signals and memories could be valued and used to guide behaviour. Under this view, there is a difference between a system in which something can feel good or bad and one in which nothing does, even though the capacities supporting that transition and the affective capacities that follow it are graded.
An analogy would be an exoplanet candidate. Astronomers may have evidence of varying strength that a planet exists around a distant star. Their confidence may be 20%, 70% or 99%, but the planet itself either exists or does not. Graded evidence does not require the underlying fact to be graded.
Likewise, the fact that behavioural, neural and physiological evidence exists on continua does not by itself imply that sentience must exist on a continuum. It means that our confidence about its presence is graded. This not requires assuming a simple or already known boundary in the evidential space. The biological transition may be complex, and we may be deeply uncertain about where it lies.
In your hypothetical, if we had both complete physical evidence and an adequate biological account of which forms of organization generate affective states, then yes: under my working view, the answer for a particular organism at a particular time would be either that affective experience was present or that it was absent. Our proposed classifications reflect our incomplete evidence and incomplete biological understanding, not a claim that the organism is partially sentient in proportion to our confidence.
This is also why the tool uses a sentience-plausibility gate followed by a separate affective-capacity analysis. The first asks whether there is sufficient reason to infer that anything can feel good or bad for the organism. The second asks, conditional on that, how differentiated and intense those states could be. The first is treated categorically in the underlying hypothesis but uncertain in our knowledge; the second is explicitly graded.
An analogy would be an exoplanet candidate. Astronomers may have evidence of varying strength that a planet exists around a distant star. Their confidence may be 20%, 70% or 99%, but the planet itself either exists or does not. Graded evidence does not require the underlying fact to be graded.
Looking into this example is helpful to explain my point. Humans have to come up with a defition for what is a planet, and then test whether a given celestial body satisfies it. Margot (2015) illustrates this. I present below their Figure 1 whose lines represents 3 potential definitons. They are analogous to the sharp evidential surface I mentioned in my last comment. Increasing the mass of a celestial body which is slightly too light to be a planet (just below the definitional line) by just 1 kg without practically changing its semi-major axis could make it a planet (cross the definitional line). I think this is fine. At the same time, I would not see any deep difference between 2 celestial bodies which are practically the same except one is 1 kg heavier, and therefore is a planet unlike the other.
My own personal view is that the biological capacities associated with sentience can vary continuously, while the minimal presence of affective experience [sentience] is categorical.
If sentience was categorical, there could in principle be 2 organisms which only differed infinitesimally in biological capacities, but one was sentient, and the other was not. This would be fine if sentience was similar to being a planet. However, sentience is often assumed to be necessary for moral consideration. In this case, one organism would get some moral value, and the other would get exactly 0 moral value. I would find this problematic if the organisms only differed infinitesimally.
This is also why the tool uses a sentience-plausibility gate followed by a separate affective-capacity analysis. The first asks whether there is sufficient reason to infer that anything can feel good or bad for the organism. The second asks, conditional on that, how differentiated and intense those states could be. The first is treated categorically in the underlying hypothesis but uncertain in our knowledge; the second is explicitly graded.
My take is that it would be better to have everything graded, and assume all organisms have more than exactly 0 moral value.
Thanks, Vasco. I think our respective positions are now becoming quite clear, and we may be reaching the point where further progress depends more on scientific and theoretical work than on extending the analogies.
My exoplanet example was intended only to distinguish graded evidence from the underlying fact being investigated. Whether a candidate celestial body exists is different from whether an already known body satisfies one of several proposed definitions of a planet. Margot’s figure addresses the latter question. My analogy concerned the former.
As I noted above, my position is that the biological capacities supporting sentience (and our evidence about them) can vary continuously even if the minimal presence of affective experience is categorical. What I would add here is that gradual evolution does not make such a transition implausible. Existing structures and capacities can be co-opted into a new functional role, as captured by the concept of exaptation: a feature that originally evolved for one function, or for no particular function, may later acquire another. This may make the evolutionary onset of sentience especially difficult to locate, because there need not be an obvious biological marker separating the earliest sentient organisms from their closest non-sentient relatives. The transition may have depended on a very small biological change, or even on a change in context that allowed existing mechanisms to perform a new role.
So yes: under my hypothesis, two organisms could in principle be extremely similar while affective experience was present in one and absent in the other. I do not find that paradoxical. A small difference in a complex system can sometimes make a new causal process possible. The supporting architecture may vary continuously even though the process we are asking about either occurs or does not occur in a particular organism at a particular time.
Your concern about one organism receiving some moral value while the other receives exactly none introduces a separate normative question. That conclusion does not follow automatically from the claim that sentience is categorical. The sentience gate in the tool is not intended to assign moral worth. Its focus is narrower—and appropriately so, given that it concerns one of the biggest mysteries in science: it asks whether affective states, and therefore sentience in the relevant sense, plausibly occur in the target taxon.
Turning to the normative question, assuming that every organism has more than exactly zero moral value is a possible normative position, but it is not a scientific consequence of treating biological capacities as continuous. It would require its own justification.
Speaking only for myself—and separately from the analytical commitments of the Welfare Footprint Framework—my principal concern is severe suffering, particularly Disabling and Excruciating Pain. Annoying and Hurtful experiences remain genuine components of welfare and should be recorded in WFF analyses. But the presence of lower-intensity negative affect alone would not ordinarily make an organism or intervention a leading priority for me, especially given the scale of severe suffering that already exists and the need to consider where effective intervention is possible.
This is one reason the affective-ceiling question matters so much to me. Nature is not benign, and neither are most modern animal-production systems. Given the scale of severe suffering that already exists, I think our attention should be concentrated where the most intense suffering is possible. If biologically simpler sentient organisms have narrow affective ranges and cannot experience severe Pain, I would find that reassuring.
I therefore think our remaining disagreement concerns at least two distinct questions. Scientifically, you favor treating sentience itself as graded, whereas I regard its minimal presence as categorical while treating our evidence—and subsequent affective range, resolution, and intensity—as graded. Normatively, you favor assigning every organism more than zero moral value, whereas my own priorities are concentrated on organisms capable of severe suffering, particularly where meaningful intervention is possible.
My exoplanet example was intended only to distinguish graded evidence from the underlying fact being investigated. Whether a candidate celestial body exists is different from whether an already known body satisfies one of several proposed definitions of a planet. Margot’s figure addresses the latter question. My analogy concerned the former.
I was analogising celestial bodies to organisms, and being a planet to being sentient. However, I also think whether a celestial body exists or not is a matter of human convention. I assume a clear definition would involve coming up with some properties like a minimum mass and density. In this case, one could come up with some amount of matter slighly too light to be a celestial body which would become a celestial body after adding it one more atom of hydrogen.
So yes: under my hypothesis, two organisms could in principle be extremely similar while affective experience was present in one and absent in the other. I do not find that paradoxical.
Even if the difference between the organisms can be arbitrarily small (not just extremely small, but arbitrarily close to a difference of exactly 0)? Supposing sentience is categorical/sharp, 2 organisms could be equal in all regards except that one differs 10^-100 % along a single metric, and therefore is the only one of the 2 which is sentient.
A small difference in a complex system can sometimes make a new causal process possible. The supporting architecture may vary continuously even though the process we are asking about either occurs or does not occur in a particular organism at a particular time.
I agree for a "small difference", but not for an arbitrarily small difference, and I think this is what distinguises sentience from lack of it if it is categorical/sharp.
Scientifically, you favor treating sentience itself as graded, whereas I regard its minimal presence as categorical while treating our evidence—and subsequent affective range, resolution, and intensity—as graded.
I believe sentience should be treated as graded if it means having a moral value above exactly 0. This is because I do not think an arbitrarily small difference should make an entity go from having exactly 0 moral value to having some moral value.
I am open to sentience being treated as binary if it does not mean having a moral value above exactly 0. However, in this case, lack of sentience would not imply lack of moral value.
Thanks, Vasco. I think your final paragraph shows that our positions are now closer than they initially appeared, and that part of the disagreement came from asking the word sentience to do two different jobs: describe whether affective experience is present, and determine whether an entity has non-zero moral value. In our tool, it is used only in the first sense.
On the scientific question, yes: under a categorical account of sentience, arbitrarily small changes near the relevant biological threshold could in principle place otherwise extremely similar organisms on opposite sides of it. I do not regard this as incoherent. Continuous underlying variables can produce categorical system-level events; a neuron, for example, may or may not fire an action potential as its membrane potential crosses a threshold. This analogy does not establish that sentience is categorical, but it shows that gradual biological change does not rule out a categorical onset.
The sentience gate in our tool therefore does not assign moral worth. It asks whether welfare-relevant affective experience is plausible enough for an affective-ceiling analysis to be meaningful. Whether entities without sentience may nevertheless have moral value, or whether moral status should vary continuously, are separate normative questions.
So perhaps our remaining scientific disagreement is now quite narrow: whether the minimal presence of affective experience is itself graded or categorical. On almost everything surrounding that boundary, however, we seem to agree: the biological capacities supporting sentience, our evidence for it, and the affective range, resolution, and intensity that follow are all graded. That is quite a lot of common ground! :) Thanks, Vasco, for the careful and constructive exchange.
On the scientific question, yes: under a categorical account of sentience, arbitrarily small changes near the relevant biological threshold could in principle place otherwise extremely similar organisms on opposite sides of it.
Would it make sense to run WFF's affective capacity analysis for only one of 2 organisms which only differ infinitesimally?
I do not regard this [the sentence quoted above] as incoherent. Continuous underlying variables can produce categorical system-level events; a neuron, for example, may or may not fire an action potential as its membrane potential crosses a threshold. This analogy does not establish that sentience is categorical, but it shows that gradual biological change does not rule out a categorical onset.
I found this example quite thought-provoking. Yet, I do not think neuronal firings are categorical. Membrane potential and neuronal firings are represented in the mid and bottom plots of Figure 9A below from section 9.1 of the book Neuronal Dynamics. Membrane potential varies continuously. It is not binary. So humans have to come up with a convention for which types of variations in it count as a firing. In principle, 2 different time series of the membrane potential could differ infinitesimally, and one have 0 firings while other has many firings which fall just short of counting as such. If one was interested in a property like the maximum membrane potential, it would make sense to analyse both time series
There is also no fact of the matter about whether a given membrane potential corresponds to a neuronal firing or not. The number of neuronal firings is a property of the time series of the membrane potential. It is not a property of a neuronal state. In contrast, I understand sentience (existence of affective experiences) is supposed to be a property of a state (of an entity at a given time). Membrane potential is a property of a state, and it is continuous/graded (not binary/categorical/sharp).
The physical world varies continuously. So I think all properties tracking it have to be continuous too. People point out phase transitions as evidence of categorical changes, but they involve continuous changes.
Would you keep the sentience gate in WFF if you thought it was a matter of human convention instead of scientific fact?
Thanks, Vasco. I had already acknowledged that the analogies were no longer helping, and then I fell into offering another one, so I concede that mistake. I do not think my underlying point is coming across, but I have laid it out as clearly as I can in our exchange. But just addressing your questions:
If one organism truly had affective experience and the other did not, the ceiling analysis would apply only to the former. For the latter, an affective ceiling would be inapplicable, not merely lower. If we could not tell which was which, the tool should reflect that uncertainty and proceed only conditionally.
(One clarification: this sentience gate belongs to our experimental Ψ approach, not to the core WFF).
Your final question captures the crux: if I regarded sentience as merely a matter of human convention, I would not retain it as a scientific gate. My position, as I mentioned before, is that whether anything feels good or bad is a real fact about the organism, even when we cannot determine it confidently.
I had already acknowledged that the analogies were no longer helping, and then I fell into offering another one, so I concede that mistake. I do not think my underlying point is coming across, but I have laid it out as clearly as I can in our exchange.
I personally found your analogy quite helpful. It gave me a chance to explain why I think many properties considered categorical are actually graded. In addition, I think trying to present properties which are categorical is important to illustrate sentience can be so too. The absence of any categorical properties would arguably be strong evidence against sentience being categorical.
Your final question captures the crux: if I regarded sentience as merely a matter of human convention, I would not retain it as a scientific gate.
Makes sense. I think sentience is a matter of human convention, which is why I would not use it as a scientific gate for the affective capacity analysis.
Executive summary: The author introduces the Interspecific Affect GPT as a structured, evidence-sensitive tool to estimate species’ maximum plausible affective intensity relative to humans, aiming to make interspecies welfare comparisons more explicit without claiming precision or resolving downstream ethical questions.
Key points:
The post transitions from prior theoretical work on affective capacity (information-processing and evolutionary lenses) to a practical tool for interspecific welfare comparison.
A central unresolved problem in welfare science is comparing affective intensity across species, especially regarding maximum intensity (“ceiling”) and how experience maps to time.
The author argues the ceiling question is often more decisive, since limits on maximum intensity constrain total possible suffering regardless of duration.
The tool focuses narrowly on estimating a species’ upper bound of pain intensity relative to a human-anchored reference scale, not on assigning moral weights or rankings.
It introduces human-anchored categories (e.g., Annoying(h), Excruciating(h)) to create a shared reference scale without implying equivalence in actual experience.
The tool is intended as a structured reasoning scaffold that makes assumptions, evidence, and disagreements explicit and open to criticism, rather than a calculator or decision rule.
It adopts methodological commitments such as biological parsimony, explicit separation of sentience and affective-capacity analysis, and avoiding unjustified cross-taxon inference.
The workflow proceeds stepwise: defining taxonomic scope, checking assumptions, classifying sentience plausibility, reviewing multi-domain evidence, assessing affective architecture, and inferring ceilings with stress tests.
Ceiling estimates are tested via evolutionary “cost of intensity,” alternative hypotheses (e.g., poorly regulated intense states), and convergence checks that widen uncertainty when evidence conflicts.
The tool includes a red-teaming step to challenge its own conclusions and produces a final dossier with sentience judgment, ceiling estimate, uncertainty considerations, and research priorities.
The author emphasizes that the tool is for disciplined scientific inference, distinct from how uncertainty should be handled in ethical or policy decisions, and invites criticism and iteration.
This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, andcontact us if you have feedback.
The Affective Ceiling Approach: Estimating Maximum Pain Intensity Across Species
The Affective Ceiling Approach is a structured, evidence-based way of estimating the highest intensity of Pain that members of a species could plausibly experience, using human affective intensity categories as explicit reference points.
This post is the third in a short series about sentience, ceilings of Pain and Pleasure intensity (‘affective ceiling’, or ‘affective capacity’), and how to make interspecific welfare comparisons more explicit without creating false precision.
In the first post, we approached the problem through what was essentially a physics or information-encoding lens, treating affective experience as a system for encoding and transmitting biologically relevant information. In this view, Pain and Pleasure are signals whose properties can be analysed in terms of how information is represented and processed. Taking the human affective system as a reference point, we focused on two key features of this signalling system: the range of intensities it can represent, and the resolution with which differences in intensity can be distinguished. We then explored whether early forms of sentience would already exhibit these attributes, or whether they might instead be associated with more limited affective range and/or resolution.
In the second post, we approached the problem through an evolutionary lens, arguing that sentience and affective capacities should be treated like other biological traits: shaped by evolutionary constraints and payoffs, with extreme affective intensities requiring adaptive justification rather than arising automatically from neural complexity.
This third post marks the transition from the theoretical framework to its practical application. It introduces an experimental tool, the Affective Ceiling GPT, designed to operationalize those ideas alongside additional ones presented here, in a way that may eventually support interspecific welfare comparisons.
What problem the Affective Ceiling tool addresses
Within the Welfare Footprint Framework (WFF), welfare is quantified as time spent in different affective intensity categories, where Pain and Pleasure are operationally defined as any negative or positive affective states. Intensity categories are assigned relative to each species’ own behavioral and physiological indicators of perceived Pain and Pleasure intensity, allowing robust within-species comparisons without requiring interspecific assumptions.
A major unresolved challenge arises, however, when one tries to compare welfare across species. This difficulty is not unique to the WFF; it applies to any welfare metric, since the intensity of affective experiences across different sentient beings remains one of science’s most profound unknowns. In the WFF, this is acknowledged explicitly in Module (Ψ) Interspecific Scaling, which highlights the need for transparent methods to address cross-species comparability without distorting the integrity of species-specific welfare assessments.
At least two distinct questions arise when comparing welfare across species. First, there is the ceiling question: what is the highest affective intensity a given species can plausibly reach? Second, there is the time-mapping question: if different species can potentially experience comparable affective intensities, does a given unit of clock time correspond to the same magnitude of experienced Pain or Pleasure?
Both issues have appeared in the broader literature on welfare range and interspecific comparison, but we make the distinction explicit here because the two questions are often run together. Our view is that the ceiling question is frequently the more decisive one: if a species cannot plausibly reach very high intensities, later adjustments involving duration or aggregation cannot place it on a par with a system that can. In other words, limits on maximum intensity constrain the overall magnitude of Pain, regardless of how experience unfolds over time. The Affective Ceiling GPT is therefore designed to address this first question, explicitly and provisionally: given the available evidence, where does a species’ maximum plausible affective capacity lie relative to a human-anchored reference scale? In doing so, the tool adopts an evidence-constrained approach: it aims to identify the best-supported upper bound, rather than to assume broader or more extreme affective capacities in the absence of positive justification.
This focus on parsimonious or evidence-constrained ceiling estimates should not be confused with a rejection of precautionary reasoning. The tool is designed to answer a scientific question: given the available evidence, what is the best-supported upper bound on a species’ affective intensity? That is a narrower question than what ethics or policy must answer. When it comes to practical decisions about whether to impose conditions that may cause suffering, the possibility that a species can feel more intensely than current evidence supports remains a legitimate concern. Because the costs of underestimating animal suffering can be high, there are good reasons to apply precautionary reasoning in such decisions even when the scientific analysis yields only a constrained ceiling estimate. These are two distinct steps, and the evidential caution appropriate for one does not automatically carry over to the other.
This tool also serves as a test case for how large language models, current and forthcoming, may support structured reasoning on a scientifically difficult and philosophically contested issue, while offering a more transparent, disciplined, and evidence-sensitive way of approaching the problem.
The role of human-anchored reference categories
One difficulty in interspecies comparisons is that the same terms for the intensity categories (Annoying, Hurtful, Disabling, Excruciating) can be used across taxa as if they referred to comparable magnitudes. Often, they do not.
To reduce this ambiguity, the tool presented here uses human-anchored reference categories, formally introduced in this text, and identified by the (h) suffix (Box 1). These categories do not replace the usual species-internal categories used within WFF analyses; they only come into play when cross-species comparison is required.
The purpose of the notation is not to claim that two species experience the same intensity whenever they are mapped to the same reference category. Rather, it is to create an explicit shared scale for asking a narrower question: if a species is sentient, how high could its maximum perceived Pain intensity plausibly go relative to human reference levels?
Box 1. Human-Anchored Affective Intensity Categories as Absolute Reference Points
A core challenge in interspecific welfare analysis is how to compare affective intensity across species. To address this, the present tool uses human-anchored negative-affect reference categories, marked by the suffix (h): Annoying(h), Hurtful(h), Disabling(h), and Excruciating(h). These serve as absolute reference points for ceiling analysis of Pain intensity.
This notation is distinct from the standard Welfare Footprint Framework use of Annoying, Hurtful, Disabling, and Excruciating as categories that are internal to a species’ own indicators (e.g., behavioural, neurophysiological). Within a given species, these labels refer only to relative differences in affective intensity inside that species and do not imply cross-species equivalence. By contrast, the (h) categories are used only when explicitly making interspecific inferences.
Accordingly, the highest Pain state identifiable within shrimp is conceptually distinct from Excruciating(h), which refers to the human reference level of pain associated with extreme phenomenological intensity and severe functional disruption. Whether the former plausibly reaches the latter is an empirical and theoretical question, not a definitional one.
If a system is sentient, what is the plausible upper bound of the intensity of Pain it could experience, relative to a human-anchored reference scale?
That is its central task. It does not compute moral weights, produce rankings, or generate prioritization recommendations. It does not quantify the magnitude of species-to-species differences in affective capacity. Nor does it claim to settle sentience debates. Instead, it aims to clarify one upstream scientific input upon which many downstream ethical analyses rely, often without making that input explicit.
For that reason, the tool is best understood not as a calculator but as a structured reasoning scaffold designed to prevent premature convergence on comfortable answers. Its goal is not to replace expertise, but to make expertise easier to apply by forcing assumptions, inferential steps, and disagreements into a more explicit and inspectable form.
We welcome criticism and feedback on the tool itself, as well as on the reasoning structure it embodies. In a field as scientifically difficult and ethically sensitive as interspecies welfare comparisons, such criticisms are very much needed and valued.
How the Affective Ceiling GPT works
(Note: This description reflects the current instruction set as of this draft. Future versions may evolve, but the account below matches the present public architecture)
The following information is not required to use the tool—its interactions are designed to be self-guided and self-explanatory—but is provided for readers who want to understand its rationale in more detail. The tool is guided not only by its stepwise instruction set, but also by a curated supporting knowledge base. This includes methodological notes, definitions, examples of good and bad outputs, and selected theoretical materials used to stabilize how the workflow is interpreted across cases. In particular, the knowledge base helps the tool keep distinct the sentience gate, the affective-ceiling analysis, and the meaning of the human-anchored reference categories. It is best understood not as a fixed database of answers, but as a supporting layer that helps the model apply the framework more consistently and transparently.
The workflow is structured so that the tool does not jump directly from a user’s query to a ceiling estimate. Instead, it separates a series of questions that are often run together in informal discussion: What taxonomic scope is appropriate? What methodological assumptions are being made? At what level should sentience be assessed? What evidence bears specifically on affective ceiling rather than merely on responsiveness or nociception? Only after those issues have been made explicit does the tool move toward a provisional ceiling estimate. The stepwise design is meant to make the reasoning easier to inspect, challenge, and revise.
Step 1: Input framing and scope
The user inputs a target taxon. The tool then determines, where necessary, two related but distinct scopes: one for the sentience gate and one for the affective-capacity analysis. This distinction matters because evidence relevant to sentience may sometimes be available at a broader taxonomic level than evidence relevant to affective ceiling. The current instruction set therefore allows the sentience scope to be broader when the evidence base is broader, while keeping the affective-capacity scope as close as possible to the user’s target unless broader generalization is necessary. In both cases, the tool is expected to justify the taxonomic scope in terms of the evidence available and potential heterogeneity among the species in the target group.
Step 2: Methodological commitments check
Before proceeding, the tool makes its key methodological commitments explicit and asks whether the user wishes to keep or modify them. These include: the Welfare Footprint Framework use of Pain and Pleasure as umbrella terms for negative and positive affective states; the use of an epistemic sentience classification rather than a numerical sentience score; a biological parsimony default in the absence of positive evidence for broader or more extreme affective ranges; the rule against inferring one taxon’s ceiling from another without explicit justification; and the decision, for present purposes, to analyze intensity ceiling prior to potential interspecific differences in the subjective perception of time. This step is important because it turns assumptions that are often left implicit into assumptions that can be criticized, revised, or rejected before the analysis proceeds.
Step 3: Sentience plausibility and operational gate
Using the sentience scope defined in Step 1, the tool conducts an initial evidence-based appraisal of sentience status and classifies it as Plausible, Contested/Uncertain, or Not coherent. This step is intentionally brief and classificatory rather than a full literature review. Importantly, the tool treats sentience primarily at the broader or ancestral taxonomic scope selected in Step 1, rather than reopening it by default at the level of the narrower target taxon. When the broader sentience scope is supported by convergent evidence, nested taxa are ordinarily treated as inheriting that status unless there is strong subgroup-specific counterevidence or a biologically serious reason to suspect loss or radical divergence.
The current instructions direct the GPT to consult Feinberg’s From Sensing to Sentience as one relevant source at this gate-setting stage, while cross-checking against other theoretical frameworks and empirical evidence.
Step 4: Evidence and indicator assembly
Using the affective-capacity scope defined in Step 1 as the primary unit of analysis, the tool assembles the evidence specifically to inform the affective ceiling question: what level of intensity this system could plausibly reach. It maps evidence across behavioural, neural , neurophysiological, pharmacological, cognitive/representational, and evolutionary domains, while allowing broader evidence only when its relevance to the narrower taxon is explicitly justified.
For each relevant indicator, the tool must state what construct it bears on, whether it informs sensitivity, capacity, both, or neither, what its main strengths and limitations are, and where it may mislead through false positives or false negatives. The point is not merely to list evidence, but to clarify what inferential work each indicator can and cannot do in constraining the plausible affective ceiling. The instructions also make explicit that nociception, defensive behaviour, or simple stimulus responsiveness do not by themselves establish advanced affective capacity.
Step 5: Affective capacity scope
Only after the evidence has been organized does the tool ask what sort of affective architecture the target plausibly has. Here the emphasis is on whether the biological hardware plausibly supports coordinated, high-intensity affective integration, whether through centralized mechanisms or functionally integrated distributed ones. This step is meant to assess architectural support and its limits, while distinguishing direct support from speculative extrapolation; it is not yet the stage at which a final or near-final human-anchored ceiling is assigned.
Step 6: Ceiling inferences and three mandatory checks
At this stage the tool proposes a provisional human-anchored ceiling, either as a single category or a range, and then stress-tests it in three ways. First, through a process we refer to as “Cost of Intensity Check” by asking: would states analogous to Disabling(h) or Excruciating(h) be evolutionarily and biologically justified by the species’ life history and metabolic budget?
Second, it applies an “Alternative Hypothesis Check”. The framework defaults to narrow ceiling estimates in the absence of positive evidence for broad affective range, consistent with biological parsimony. At the same time, the tool checks a specific counter-argument: could a lack of regulatory control over affective states result in intense but poorly modulated negative experience? This possibility is treated as speculative unless the evidence specifically supports it. It remains a live counter-hypothesis, but not one that overrides the parsimonious default in the absence of positive evidence.
Third, the tool uses a final “Convergence Check”: if the evidence is in conflicting directions, the output should widen the uncertainty bounds rather than force an artificially crisp answer.
Step 7: Red-team / steelman
The tool must then construct the strongest plausible case against its own conclusion. Depending on the case, this may involve arguing that the current estimate is too low, too high, or otherwise too confidently bounded. The aim is to expose where the conclusion is most vulnerable to reinterpretation and to ensure that the final output reflects engagement with the strongest serious challenge, rather than premature convergence on a comfortable answer.
Step 8: Final dossier, subjective time, and research priorities
The final output is not a moral weight or a ranking of species, but a dossier containing the sentience-plausibility judgment, the affective-ceiling estimate, a brief assessment of whether subjective time could become relevant in later analysis, a research-priority judgment identifying the single experiment most likely to change the conclusion, and a disagreement log summarizing the user’s objections.
At the end of each step, the users are asked whether they agree, disagree, or want to revise any part of the reasoning before the next step. These interactions are designed to transform disagreement into more explicit and inspectable arguments, reducing the impression of false certainty.
Churchland, P. S. (2002). Brain-wise: Studies in neurophilosophy. MIT Press.
Crump, A., Browning, H., Schnell, A. K., Burn, C., & Birch, J. (2022). Sentience in decapod crustaceans: A general framework and review of the evidence. Animal Sentience, 7(32), 1.
Damasio, A. (1999). The feeling of what happens: Body and emotion in the making of consciousness. Harcourt.
Damasio, A. (2010). Self comes to mind: Constructing the conscious brain. Pantheon.
Elwood, R. W. (2011). Pain and suffering in invertebrates? ILAR Journal, 52(2), 175–184.
A preliminary estimate, and a request for better ones.
Summary
I believe the standard literature estimates for the number of DALYs attributable to a case of stunting are too low, largely because they don’t account for the long term effects. This means that childhood nutritional interventions that reduce the prevalence of stunting may be substantially more cost-effective than previously believed.
Epistemic status
Exploratory and back-o...
Note: This post was crossposted from Planned Obsolescence by the Forum team, with the author's permission. The author may not see or respond to comments on this post.
Subtitle: It’s a major warning shot, and might be the last one we get
All opinions are my personal view, and don’t represent my employer or fellow investigators.
This week, METR and Redwood Research published...
Overview
This post summarizes a new preprint on alternative proteins from the Humane and Sustainable Food Lab: New alternatives, same orders. We investigated 19...
Update, August 2026: We have revised the title and some terminology in this post to place greater emphasis on the Affective Ceiling Approach, which is the underlying scientific contribution, rather than on its current implementation as a Custom GPT. We have also renamed the tool the Affective Ceiling GPT.
These are changes in framing and naming. They do not alter the post’s substantive claims, evidence, reasoning or conclusions.
The Custom GPT is an experimental and temporary way of making the approach available. The Affective Ceiling Approach is independent of that particular platform and can be applied by researchers directly or implemented through other, more capable tools as they are developed.