NOTE-769: Bird & Tobin — Natural Kinds (SEP)

Contribution

An encyclopedia entry that organises the natural-kinds literature by four metaphysical questions (are our kind classifications natural, are they into kinds, what are kinds, do they have essences), then tests the answers against biology, chemistry, psychology and the social sciences, and closes with the epistemology of kinds and the semantics of kind terms. It argues for little of its own. Where it does conclude, the conclusion is that a proposed condition is necessary and not sufficient, or that an argument assumes what it claims to derive.

Key insight

“Natural kind” bundles several commitments that come apart: that some classifications are natural (naturalism), that there are entities that are the kinds (strong realism), and that kinds have essences. “Intermediate views tend to be more popular” (§1): natural classifications by kind, with no kind-entities and with or without essences. Most disputes about whether something is a natural kind are disputes about which of these is meant.

Assumptions

• The entry’s subject is the kinds of the sciences; that “it is the use of natural kinds by the individual (‘special’) sciences that generates our interest in them” (introduction).
• A kind is “natural” if it “reflects the structure of the natural world rather than the interests and actions of human beings” (introduction).
• Its scale of commitment (Bird 2018, §1) runs from out-and-out conventionalism to irreducible kind-entities with essences.

Key results

• Six criteria, mostly necessary only (§1.1.1).
  1. Shared natural properties: “necessary … but not a sufficient condition”; Mill: white objects do not form a natural kind.
  2. Support for induction: also “necessary but not sufficient”, and “does not go beyond Criterion 1”.
  3. Participation in laws: more contentious, and still not sufficient, since laws “may concern natural qualities and quantities that do not define kinds (e.g. Newton’s law of gravitation)”.
  4. Forming a kind: needed because 1–3 fail (“negatively charged things, nor … things with a mass of 1 kg”), and not clear in what it requires.
  5. Hierarchy and 6. categorical distinctness: contentious, and each fails for some scientific kinds (allotropes and functional groups in §2.2; species in §2.1).
• Induction runs on properties and relations, not kinds (§1.2.1). On the narrower conception, “the real answer to ‘what tends to confirm an induction?’ is not ‘natural kinds and their instances’ but is ‘natural properties and their instances’”, and “inductions often concern relations … But we do not suppose that objects that enter into a natural relation form a natural kind.” Quine’s liberal view, by contrast, counts any set whose members share a natural property, including positively charged objects.
• Cluster kinds (§1.2.2). Boyd’s homeostatic property clusters: properties kept together by mechanisms that make deviations unlikely to persist, with vague boundaries and no individually necessary property. Objections: the mechanisms are not delineated (Boyd), deciding when two mechanisms are the same involves convention (Craver 2009), and they overlap (Tobin 2017). Slater (2015) recasts it as stable property clusters, “sufficiently stably co-instantiated to accommodate the inferential and explanatory uses”. Chakravartty argues that liberal realisms dissolve into conventionalism or into realism about something else.
• Conventionalism with a check (§1.1.2). Hacking (1999): constructed categories can be real or not, and which is “subject to empirical investigation”. Satanic ritual abuse was a constructed idea whose cases were not substantiated; child abuse, a category with a datable origin (Denver, 1961), had “real instances … found through empirical investigation”. Kukla’s three dependences: material, causal (Hacking’s interactive kinds, Merton’s bank run), constitutive.
• Pluralism (§1.1.3, §2.1). Dupré’s promiscuous realism: many cross-cutting, equally legitimate taxonomies, each objectively grounded, with no hierarchy and no categorical distinctness. Species concepts disagree on extensions; pluralism is “the weakest form of realism about kinds in the philosophy of biology”.
• Essentialism argued for and against (§1.3). Kind essentialism does not entail that a particular’s kind is essential to it (neptunium-239 decaying to plutonium; speciation by isolation). The Kripke–Putnam arguments establish at most “partial essences”, necessary conditions such as atomic number 79 for gold. Salmon: only trivial essentialism follows from the theory of reference, and the reconstructed water argument needs an essentialist premise, P3. Fine: essential is not the same as necessary.
• Eliminativism (§1.4). Hacking (2007): “Delete every mention of natural kinds”; Ludwig (2018) likewise, on the ground that the concept covers heterogeneous practices and no unitary theory fits all sciences. Conix and Chi (2021): the concept still “enables fruitful investigation into non-arbitrary classification”.
• The special sciences (§2). Species may be individuals, not kinds (Ghiselin, Hull), and their membership conditions are relational, so either species are not natural kinds or kindhood need not be intrinsic. Chemical kinds are the current paradigm, with complications (isomers, water as a dynamic structure, allotropy). Mental kinds face multiple realisability; Griffiths (1997) argues emotion is not a natural kind and Machery that “concept” is not one. Social kinds are interactive: “the object of classification itself changes in being classified” (Hacking 1995, 1999). Khalidi (2010): the difference from natural kinds is one of degree, “of the degree of control we have over the object of classification”.
• Epistemology first, and its critics (§3). Slater (2015) and Ereshefsky and Reydon (2015): judge kinds by their role in inductive inference within scientific practice; many successful scientific kinds (non-causal, functional, heterostatic) do not count as natural under the HPC view. Against them, Lemeire (2018): epistemic success “is constrained by deference to the world not only in cases of natural kind classification, but also in cases of conventional and pragmatic classification”, so a purely epistemic approach “fails to give us a principled way of distinguishing between the natural kinds and their conventional counterparts”. Kendig and Grey (2021): the stability of a cluster cannot be assessed without the metaphysics that explains it. Reydon (2016): kinds are co-created by nature and classifier.
• Semantics (§4). The modal, epistemological and semantic (Twin Earth) arguments against descriptivism; direct reference and externalism; Putnam’s same-kind relation anchored to a sample. Rigid designation for kind terms may be trivial (Soames, Schwartz, Martí) unless kind terms denote only sparse, natural entities. Whether essentialism follows from the semantics: Salmon and Burge against; Bird (2020) that the semantics carries metaphysical assumptions that competent users make tacitly.

Claims

id	claim	strength	support
C1	Shared natural properties, support for induction and a role in laws are each necessary but not sufficient for a natural kind	moderate	counterexamples (white things, charged things, Newton’s law), §1.1.1
C2	Natural relations that support induction do not thereby make their relata a kind	moderate	argument from the narrower conception of kinds, §1.2.1
C3	The Kripke–Putnam arguments establish at most partial essences, or essences only with an essentialist premise	moderate	Salmon’s reconstruction and Kripke’s own cases, §§1.3.1, 4.4
C4	Kind essentialism does not entail that a particular’s kind is essential to it	moderate	the neptunium and speciation cases, §1.3
C5	A purely epistemic criterion cannot separate natural from conventional kinds	weak here	reported from Lemeire (2018), §3
C6	Social kinds differ from natural kinds in that classifying them changes them, or in degree of control only	weak here	reported from Hacking and from Khalidi, §2.4

Concepts

• Naturalism (weak realism). Some classifications are natural; no commitment to kind-entities.
• Strong realism. There are entities that are the natural kinds.
• Essence. A property or set whose possession is necessary and sufficient for membership; a partial essence is necessary only.
• Homeostatic property cluster. A family of properties kept co-occurring by mechanisms, without a necessary member.
• Stable property cluster (Slater). The same, judged by the stability that suits it to inference, not by the mechanism.
• Interactive kind (Hacking). A kind of person whose members respond to being classified, which changes the kind.
• Kinding. The act of making kinds in scientific practice, in the epistemology-first literature.

Connections

The entry builds on Mill, Whewell, Quine (1969), Kripke (1972/1980), Putnam (1975), Boyd (1991, 1999), Dupré (1993), Ellis (2001) and Hacking (1995, 1999, 2007). Its §3 is the most recent material (Slater 2015; Ereshefsky and Reydon 2015; Lemeire 2018; Kendig and Grey 2021). Ladyman’s real patterns (LIT-219) and Dennett’s (LIT-220), which the record reads for the same question, are not mentioned in it.

Bearing on the record

• QUESTION-028 is sharpened, not answered. The question asks when “independently generalizable, intervention-relevant relational structure” warrants treating a communicative category as a real kind.
  • The entry’s §1.2.1 says that natural relations supporting induction do not thereby make their relata a kind, and §1.1.1 that inductive support is necessary, not sufficient. So the question’s qualifiers cannot be enough by themselves on the mainstream view; something must answer the “kindhood” criterion (4), which the entry calls unclear.
  • Lemeire’s point (§3) is the sharpest form of the record’s guard, CLAIM-150: predictive and explanatory success constrains conventional and pragmatic classifications too, so it cannot be what separates a real kind.
  • The entry also supplies the two most usable positive answers. One is Slater’s stable property cluster, which would make the question’s qualifiers sufficient if stability is judged before the data. The other is Hacking’s test for constructed categories, that real instances are “found through empirical investigation”. It is close to the question’s own answer condition: generalisation to realisations not used to define the category.
  • Communicative categories are classifications of human acts, so Hacking’s interactive kinds apply: a category used by speakers can change what they do. Any test of the kind the question wants has to allow for that, or show the category is indifferent to being named.
• THEORY-036. Consistent with it and independent of it. Dennett’s compressibility over-admits patterns; on this entry, shared properties and inductive support over-admit kinds (Quine’s liberal kinds include positively charged objects). Both say the excluding work is done by a further clause. The entry never mentions information or compression.
• THEORY-057. The entry reports Griffiths (1997) that emotion does not correspond to a natural kind, which agrees with THEORY-057’s conclusion from a different line.
• The owner’s heading. The work was filed under “Relating K-complexity to minimum sufficient statistics”. Nothing in the entry concerns information, compression or statistics, as the skim (NOTE-217) anticipated; that link is the owner’s to make. The positions most comparable to “the model class decides what counts as structure” are Dupré’s promiscuous realism and Reydon’s co-creation.
• Filed from this reading: THEORY-273, on inductive and explanatory usefulness as necessary and not sufficient.
• No instruction for machine-learning practice; nothing for the anthology.

Limitations

• A survey. Its reports of Lemeire, Slater, Khalidi and others are summaries, and the arguments are not reproduced in full.
• The social-kinds section (§2.4) is short and leans on Hacking and Khalidi; nothing in it treats linguistic or communicative categories.
• The 2022 revision was read; the Spring 2024 archive the owner cites was not checked against it.

Open questions

• What the “kindhood” criterion (4) requires is left open by the entry itself; an account of it is what QUESTION-028 needs.
• Whether Lemeire’s objection defeats Slater’s stability criterion, or only an epistemic approach with no stability clause, is not settled in the entry; reading Lemeire (2018) and Slater (2015) would.