NOTE-116: Schwitzgebel & Sinnott-Armstrong — Sacrificing humans

Contribution

It shows that three prominent 2024–25 books on animal and AI moral status share one structure. Each argues from premises that, taken at face value, license a radical deprioritisation of humanity, and each stops short with near-term advice justified by contingency, democratic assent or cultural deference rather than by principle. The review makes the underlying conditional explicit, and states it in separate forms for sentientist, deontological and perfectionist grounds. It also takes apart two operational pieces of Birch’s framework: what “precaution” means, and how sentience candidature is policed.

Key insight

Precaution has no direction until you say which error you most want to avoid. “Err on the side of caution” can mean protecting a possibly sentient creature or protecting human liberty and practice, and Birch’s framework picks the first without arguing for it. Combine that choice with aggregation over large or numerous populations, and the edge Birch and Sebo walk up to is not a stopping point that their principles themselves supply.

Assumptions

Premises the review works from:

• Deep uncertainty about consciousness itself: what it is, whether it comes in degrees, which entities have it, how to test for it (§1, citing Schwitzgebel 2024, Lau 2025, Seth & Bayne 2022).
• Uncertainty about the ethics of consciousness. It may be sufficient, necessary, or neither for moral status (§1, fns 2–4). Panpsychism would break sufficiency and illusionism would break necessity (p. 4).
• E1/S1 and E2/S2 as illustrative, not endorsed, premise pairs (§1). There are ~10¹⁹ insects (p. 4), with an assumed capacity of at least one millionth of a human’s for valenced consciousness.
• The books’ own texts at face value. Birch’s Framework Principles 1–3 (pp. 131, 133, 134) and his definition of a sentience candidate (p. 124). Sebo’s four “epistemically live” views (pp. 20–23, 37, 76) and his precautionary chapter titles. Keane’s descriptive and normative cultural relativism (p. 143).
• Interpretive charity, stated explicitly. FP1 is read as requiring all or enough proportionate precautions (fn 9). FP2 is read as requiring that the question be considered responsibly and acted on (fn 10).

Key results

What it argues:

• §1. Familiar ethical and scientific theses combine into radical conclusions (E1+S1 for insects, E2+S2 for AI). “We focus on the conditional … We take no stand on ponens vs. tollens” (p. 6).
• §2 Birch.
  • (2.1) Precaution is ambiguous about which error to avoid. The Bainbridge case and the “risks are not equal” argument (Birch p. 17) assume that false negatives cost more, but assessing that “turns on the very science at issue”. Mill’s liberty-precaution conflicts with Birch’s harm-precaution.
  • (2.2) The “zone of reasonable disagreement” is policed obscurely, both for the science (biopsychism is out, midbrain theories are in) and for ethics (only “maximize the suffering of all animals” is named as beyond the pale, p. 49). The consensus principle, avoid gratuitous suffering, implicitly sidelines autonomy, tradition and religious views. Suggested fix: add “avoid gratuitously violating the principles of any ethical view within the zone”.
  • (2.3) Citizens’ panels are impractical at 150–450 members and are sure to be human-centred. Species-proportional representation would outnumber humans roughly 100:1. Proportionality is “under-theorized and left to committee”. If human sophistication justifies priority over animals, superior AI might justify priority over humans.
  • (2.4) An impartial alien applying Birch’s principles might deprioritise humans. “Birch+” would accept the principles and reject the panels.
• §3 Sebo.
  • (3.1) The conjunction of AI and insect patiency, equal weight, aggregation and equal consideration across species, substrate, space and time implies the trolley should go to the human over two ants, and makes antibiotics “monstrous” if microbes count. His chapter titles (“If you might matter, we should assume you do”) are untenable literally: one-in-a-trillion harms, and incompatible harms. The 2023 “rebugnant conclusion” article accepts ant-over-human replacement in principle and resists it only on contingent grounds (the galaxy-colonising stewardship argument).
  • (3.2) The “relational” and “co-beneficial” moves blunt the conclusion. If fish and insect interests vastly outweigh ours, the mild advice is “like confronting a murderous slave owner only with the advice to lynch and beat his slaves slightly less” (p. 28). Sebo should say what the world he envisages would look like, and what to do between a stranger and ten million ants.
• §4 Keane. An anthropological, “giddy” book with thinly supported generalisations (NICU “cyborgs”, Japanese preference for robot carers). Its best idea is Feuerbachian: we project idealised features onto AI and treat it as a god. Relativism about practices needs a check from facts about sentience. Keane’s closing caution against interference omits that interference can be justified, and omits the danger of worshipping the wrong thing.
• §5. None of consciousness, agency, intellect, creativity or appreciation of ethics secures human specialness against AI and animal science. The conditional is plausible, and choosing between the consequent and modus tollens “will require further reflection”. “We are not endorsing this radical conclusion” (p. 38).

Claims

id	claim	strength	support
C1	Familiar ethical premises plus plausible scientific premises imply that insect aggregates or near-future AI could outweigh humanity	moderate	worked premise pairs E1/S1 and E2/S2 (§1); the scientific premises are called “not unreasonable”, not defended
C2	An appeal to precaution does not by itself fix which error to avoid; Birch’s harm-precaution and Mill’s liberty-precaution conflict	strong	conceptual argument with worked cases: fishing, bathrooms, the ant-colony trolley (§2.1)
C3	Birch’s sentience-candidate boundary between “speculative” and “credible” evidence is not principled	moderate	textual argument: biopsychism is excluded (p. 62) while midbrain and central-complex theories are included (p. 272), against network-style HOT and GWT (§2.2)
C4	Citizens’ panels cannot be both inclusive and deliberative, and will predictably protect human interests	moderate	practical argument (§2.3); not tested against any real panel
C5	Species-proportional representation would give insects ~100 representatives per human	moderate	arithmetic on assumed 10% likelihood × 10⁻⁶ degree × population (p. 19)
C6	If human sophistication grounds priority over animals, superior AI could ground priority over humans	moderate	parity argument (§2.3, p. 20; §3.1, p. 26)
C7	Sebo’s live views, taken together, imply choosing two ants over a human stranger	moderate	derivation from views 2 and 4 (§3.1); the reviewers note that the liveness of each does not make the conjunction live
C8	Sebo’s and Birch’s retreats to moderation rest on contingent or political grounds, not on their principles	moderate	close reading (§§2.3–2.4, 3.1–3.2)
C9	Keane’s cultural generalisations are thinly supported	weak	the reviewers’ own cultural experience and the sources’ limits (§4, fns 29–31)
C10	Human specialness is under serious threat on common views of the grounds of moral standing	moderate	survey argument (§5); explicitly a conditional, with the reviewers neutral on its antecedent

Concepts

• Sentience (Birch) — the capacity for valenced phenomenally conscious experience (Birch p. 26, as reported on p. 7).
• Sentience candidate — a system with an evidence base that (a) implies a realistic possibility of sentience that it would be irresponsible to ignore in policy, and (b) is rich enough to identify welfare risks and design precautions (Birch p. 124). The reviewers read (b) as practical, not epistemic (fn 11).
• Investigation priority — Birch’s lower tier, e.g. insect larvae, spiders, gastropods and nematodes (Birch pp. 126, 284).
• Zone of reasonable disagreement — the range of scientific and ethical views to be taken seriously in policy. The review’s main target in §2.2.
• Rebugnant conclusion / Pascal’s bugging — Sebo’s puns on Parfit and on Pascal’s mugging. The first is the result that a large insect population with greater total welfare is better than humans. The second is the dominance of expected value through low-probability microbial sentience times vast numbers.
• Birch+ — a “nearby possible philosopher” who keeps Birch’s principles but drops the panels, and so deprioritises humans (§2.4).

Connections

The account of consciousness it holds. No theory. It is openly sceptical about the science of consciousness (fn 1: Schwitzgebel’s Weirdness of the World, Lau’s “The End of Consciousness”) and agnostic about how consciousness relates to moral status in both directions: panpsychism breaks sufficiency and illusionism breaks necessity. The review treats the attribution of consciousness to insects and AI as an open empirical question whose answer bears on moral weight, and argues only about what follows conditionally. This is the stance of the first author’s LIT-191 (deep, possibly lasting uncertainty about AI consciousness) carried over into ethics. It resembles LIT-159 in method: a conditional argument whose antecedent is left for the reader to reject if they must.

Against held works.

• LIT-111 (Birch’s centrist manifesto) is Birch’s later, AI-specific statement. The review’s critique of sentience candidature and of precaution applies directly to its research-programme proposals.
• LIT-056 (Butlin et al.). Birch is a co-author. Its indicator method, derived from computational-functionalist theories (GWT and others), is the evidence base behind Birch’s AI candidates (Birch p. 321). The review’s §2.2 point, that network-based theories cut against Birch’s region-based animal criteria, sits beside it.
• LIT-126 (Vieira) rejects precaution toward AI outright, which is the view the review treats as one side of an unresolved choice between precautions.
• LIT-207 (Roberts) and LIT-206 (Goldstein & Lederman) address what chatbots are and want, not moral weight. Neither engages aggregation or precaution.
• LIT-135 (Seth), if right that AI consciousness is unlikely on current trajectories, would weaken S2, the premise that makes the AI arm of the conditional bite.

Bearing on the record

• consciousness tag: justified. The review is about attributing phenomenal and valenced consciousness to insects and AI, and about the ethical role of consciousness. Both fall squarely under the blurb. ethics is correctly primary.
• Any THEORY or cluster note built around AI-sentience precaution (the LIT-111 and LIT-056 line) should cite this review for the point that precaution needs a stated error preference, and for the critique of how sentience candidature is policed.
• ML practice: nothing. No instruction for machine-learning practice. It bears only on policy about AI welfare, not on how models are trained or evaluated.

Limitations

• It is a review. Its targets are the books’ formulations, and page citations are the reviewers’. Whether the books are fairly represented can be checked only against the books, which were not read here.
• The scientific premises S1 and S2 are called “not unreasonable” but get no argument. The radical conditional is only as strong as those premises.
• Much of the case against Sebo rests on reading his chapter titles literally, which the reviewers concede he does not mean (§3.1).
• The Keane section reviews tone and evidence more than argument, which suits the book but leaves it thin philosophically.
• By design the reviewers commit to no answer: the conditional is defended, the choice between ponens and tollens is not.
• This reading is of the September 2025 draft. The Ethics text was not reached.

Open questions

• Is there a principled, non-question-begging way to choose between harm-precaution and liberty-precaution when they conflict (§2.1)? Birch and Sebo assume the first.
• Can sentience candidature be given a sociological or other explicit standard (fn 23) without inheriting the conservatism the reviewers attribute to citizens’ panels?
• Is there a ground of human priority that neither animal science nor AI threatens? §5 finds none but does not argue that none exists.
• For AI specifically: under what evidence would S2 (near-future AI with human-level or superior conscious practical deliberation) become more than “not unreasonable”? That is where LIT-056 and LIT-135 disagree.

Corrections to the seeded skim

• Dossier access: “the full HTML draft … about 11k words: abstract, keywords, section heads, §1, the openings of §§2–4, and §5”. This reading covers the whole draft, every subsection and footnote, from the PDF of the same 3 September 2025 file. The Ethics text is still unread.
• The dossier’s §1 gives only the insect arm (E1/S1). §1 also has an AI arm, which is the part that bears on AI. “(E2.) Entities with human-level or superior capacities for conscious practical deliberation deserve at least equal rights with humans.” “(S2.) Near future AI systems will have human-level or superior capacities for conscious practical deliberation.” E2 is “commonly accepted by deontologists, contract theorists, and others”, and S2 is “not unreasonable as a scientific prediction” (p. 5). The same structure returns in §2.3 (p. 20), where it is “suspiciously self-serving” to put humans above an animal threshold of status while denying any higher threshold that superior AI could cross. It returns again in §3.1 (p. 26) for perfectionist and deontological views.
• Dossier check: “the reviewers’ treatment of Birch’s ‘sentience candidate’ threshold”. It is substantial (§§2, 2.2). They quote the definition (Birch p. 124), which has two parts: (a) a realistic possibility of sentience that it would be irresponsible to ignore, and (b) an evidence base rich enough to identify welfare risks and design precautions. They read (b) as making candidature “partly a practical criterion” (fn 11). They list Birch’s candidates, which include “AI systems that have the computational markers of sentience according to computational functionalist theories such as global workspace theory and perceptual monitoring theory” (Birch p. 321). They then press three problems. Birch’s line between dismissible “speculation” (biopsychism, p. 62) and “credible” evidence (midbrain theories) is obscure. His reliance on brain regions (the insect central complex, p. 272) conflicts with network views, the “straightforward interpretation” of higher-order and global-workspace theories. And trade-off behaviour (hermit crabs, pp. 276–277) may be non-representational or non-conscious. As a “friendly amendment” they float a sociological standard, e.g. “>20% of academics in relevant fields have >20% credence” (fn 23).
• Dossier: “Birch’s endpoint is a slow, democratic proportionality process.” Correct, and the review argues against it at length (§2.3). Birch proposes panels of 150–450 people (Birch p. 146) using PARC tests. The reviewers object that such panels cannot be both inclusive and deliberative, and that non-experts cannot be adequately informed. They add that the panels will be conservative and human-centred, against Birch’s own precautionary default. Representation weighted by likelihood × degree × population would give “about a hundred representatives of insect interests for every one representative of human interests”, on 10% × one-millionth (p. 19).
• Missing from the dossier: the analysis of precaution (§2.1). “By itself, an appeal to precaution does not specify the type of error most to be avoided” (p. 8). Mill’s harm principle, read as a precaution in favour of liberty, puts the burden on whoever would restrict liberty, while Birch puts it on whoever risks harming a sentience candidate. The reviewers “see no reason always to prioritize” one over the other (p. 11). This is the review’s main philosophical argument against Birch, beyond the charge of retreat.
• Keane (§4). The dossier’s account is right as far as it goes. §4.1 (Feuerbachian projection: AI as a god on anthropological standards of godhood) and §4.3 (“What We Owe AI”) are also substantive. §4.3 argues that Keane needs Birch and Sebo, because “Is the cow, insect, or AI system genuinely sentient?” matters however well a culture’s practices work for it.
• Title. Crossref records the Ethics title as “Sacrificing Humans for Insects and AI” without the subtitle. The record’s longer title is the draft’s. Whether the published article carries the subtitle is unverified.
• primary topic: ethics (unchanged, correct). A secondary religion tag is justified, because §§4.1–4.3 treat AI as a candidate object of worship and the repurposing of religious practice.