When Useful Concepts Become Bad Metaphysics
Simon Critchley and Luciano Floridi make almost opposite arguments about humans, AI, and information. They also illuminate the same philosophical danger.
Philosophy often goes wrong not because it begins with a false distinction, but because it asks a good distinction to do too much. A conceptual tool proves useful: it organizes a messy domain, isolates a genuine difference, or clarifies something that had previously been confused. Then, almost imperceptibly, its status changes. What began as a way of describing reality becomes a claim about what reality fundamentally is. What began as a feature of familiar human experience becomes a necessary condition of meaning. What began as an abstract classification begins to carry moral weight.
I will call the general pattern category inflation. This is not merely an underargued transition from one proposition to another. Category inflation occurs when a category has earned legitimate authority in one role — descriptive, explanatory, methodological, phenomenological — and that very success is then allowed to count as support for a stronger role that has not independently been established.
The promotion can take several forms. Modal inflation treats a feature characteristic of one realization as evidence that the feature is necessary. Ontological inflation allows the success of a representation or theoretical interpretation to support a stronger claim about what fundamentally exists. Axiological inflation allows classificatory or ontological generality to count as evidence of normative significance.
These are not one formal fallacy. They are members of a family. The common error is not simply “a premise is missing.” It is that authority legitimately acquired by a category at one level migrates into another.
Simon Critchley and Luciano Floridi provide an unusually revealing pair because they move in almost opposite directions. Critchley concentrates philosophical significance around mortal human experience. Floridi expands informational categories toward a universal ontology and an ontocentric ethics.
The question, in each case, is therefore not whether the resulting philosophical claim is ambitious. Ambition is not an error.
The question is: What licenses the promotion, apart from the success the category already enjoyed in its original role?
The philosopher who “gives a damn”
In “The Id, the Ego and the Superintelligence,” Simon Critchley contrasts contemporary artificial intelligence with the sort of judgment he associates with philosophy. Machines retrieve, summarize and recombine; humans judge. Judgment, in his telling, involves weighing, discrimination, taste and practical discernment, and is refined through lived experience: error, embarrassment, boredom, grief and even the occasional hangover. Meaning, in turn, derives from mortality. Critchley states the modal thesis plainly: “There can be no meaning in immortality.” His alternative Turing test is correspondingly affective: can a machine experience social anxiety or, in his memorable formulation, “give a damn”? [1]
There are several real distinctions underneath this picture, and they should not be erased. Judgment is not identical to explicit calculation. Sentience is not identical to intelligence. Experience is not identical to its description. Concern is not identical to information processing. No quantity of empirical description mechanically settles every normative question.
The trouble begins when these distinctions start traveling together as though they constituted one necessary package.
Critchley’s account of judgment gathers a remarkably heterogeneous collection of properties: discrimination between answers, truth sensitivity, taste, lateral thinking, selective forgetting, emotional experience, mortality, anxiety and concern. Their coexistence in familiar human beings does not show that they constitute a single capacity, still less that each is necessary for the others.
The evidential burden also differs among Critchley’s claims. His mortality claim is explicit: immortality, he says, cannot contain meaning. His claim about intelligence is qualified as intelligence “in the human sense” and tied to caring and social anxiety. His characterization of judgment is looser: judgment is said to be refined through a cluster of experiences, but the essay does not establish that every member of the cluster is individually necessary for judgment.
Those distinctions matter. A being incapable of social anxiety would not thereby be incapable of inference. A being without embarrassment might still distinguish strong evidence from weak evidence. And a non-mortal being could, at least conceptually, possess projects, preferences, relationships, irreversible decisions, finite attention, aesthetic interests, commitments, curiosity or concern for mortal beings.
Critchley is entitled to argue that mortality profoundly structures human meaning. But the stronger proposition — that mortality is a necessary condition of meaning as such — is modal and requires another argument. A property can be constitutive of one familiar realization of a phenomenon without being constitutive of every possible realization.
The same caution applies to judgment. Scientific inquiry is saturated with judgment: researchers decide what to measure, which approximations are legitimate, whether an anomaly is signal or artifact, whether two operationalizations track the same phenomenon, and whether a model has explanatory rather than merely predictive value. Engineering likewise involves judgment under constraints: risk, robustness, acceptable failure, cost, uncertainty, competing objectives and human use.
This does not prove that scientific judgment and Critchley’s existential conception of judgment are identical. It establishes something more precise: judgment cannot do the philosophical work until the relevant form of judgment has been specified.
Critchley’s strongest exposure is therefore modal inflation. A collection of properties gains legitimate salience because it describes how judgment, meaning and concern are actually instantiated in familiar human lives. That phenomenological authority is then allowed to support claims about what meaning or intelligence could be under any realization.
The mistake is not drawing a distinction between humans and machines. It is allowing the descriptive success of the human category to count as evidence for modal necessity.
Floridi begins from the other end
Luciano Floridi’s philosophy could scarcely begin farther away from this humanistic contraction. Instead of protecting a privileged human domain, Floridi develops informational concepts across computation, epistemology, metaphysics and ethics.
A serious criticism has to begin by avoiding a caricature. Floridi does not simply maintain that everything is made of digital bits. He explicitly rejects digital ontology, distinguishing it from an informational ontology whose central commitment is structural rather than computational in that crude sense. Digital and analogue are, for Floridi, modes in which reality can be presented at particular levels of abstraction rather than rival descriptions of its ultimate substance. [2]
Nor does Informational Structural Realism proceed by the simple inference: reality can be described informationally, therefore reality is fundamentally information.
Floridi gives a much more developed argument. He first works through the dispute between epistemic and ontic structural realism. He argues that a defensible structural realism can admit structural relata without restoring ordinary substance metaphysics. Only then does he develop structural objects “in terms of informational objects,” arriving at Informational Structural Realism: a view of reality as informational objects dynamically interacting. [3]
That intermediate argument matters. Omitting it would make Floridi easier to defeat than he actually is.
But reconstructing it also exposes a more exact pressure point.
Suppose, for argument’s sake, that Floridi has earned the structural part. Reality is mind-independent, structurally differentiated and relationally organized. What justifies treating the informational interpretation of this structure as a substantive characterization of ultimate reality rather than merely an exceptionally useful theoretical articulation of it?
Floridi’s own presentation makes this question especially interesting. Object-oriented programming is explicitly not offered as an ontology; he uses it as a methodology for clarifying structural objects and conceptualizing them as informational objects. The informational formulation also gives his ontology scalability, portability and conceptual unity across levels of abstraction. [3]
Those are genuine virtues. A vocabulary can reveal structural similarities, unify otherwise separate domains and make a theory dramatically more tractable without adding a second metaphysical ingredient to reality.
So the objection should not be that “information must add something on top of structure.” Floridi could coherently answer that informational objects are precisely his interpretation of what mind-independent differentiated structure fundamentally is.
The more careful question is this:
If “informational” is coextensive at this level with mind-independent differentiated structure, what warrants treating the informational interpretation as a substantive ontological advance rather than a theoretically powerful redescription of structural realism?
Conversely, if the informational characterization contributes richer content — representation, encoding, semantics, discrimination for a system, communication, uncertainty reduction or some other distinctively informational relation — then that richer content requires justification before being attributed to ultimate reality.
This is a dilemma in the sense of locating an outstanding burden, not a refutation. The argument does not yet establish that Floridi’s informational interpretation is redundant, nor that he illicitly imports a richer notion of information. It identifies the distinction he must make clear.
And that distinction matters precisely because the informational vocabulary has already proved enormously successful methodologically. Its portability, generality and explanatory elegance can make the ontological promotion feel natural. But those virtues are evidence for the usefulness of the framework before they are evidence for a further claim about what reality fundamentally is.
Informational Structural Realism therefore illustrates a potential ontological inflation. Structural realism may have been earned before the stronger status of its informational interpretation has been fully earned.
From information to moral standing
Floridi’s ethical theory raises a harder question because here the move into axiology is entirely deliberate.
Information Ethics is explicitly ontocentric and patient-oriented. Floridi asks for the lowest common condition under which something can possess even minimal intrinsic moral worth. His answer is informational existence: entities considered as informational objects possess at least minimal moral worth qua informational objects. His early formulations develop this through principles including reflective respect, ontic uniformity and ontic solidarity. [4]
This is not an accidental slide from an “is” to an “ought.” Floridi knows that he is constructing an axiology.
And that matters because one tempting criticism is too weak. It is not enough to observe that a normative premise is needed to move from informational existence to moral considerability. Ethical theories have normative premises. Sentience is itself a descriptive property; the principle that sentient beings deserve moral consideration is normative. The fact that such a bridge is normative does not show that sentience cannot be the property in virtue of which moral status obtains.
So the relevant question for Information Ethics is not whether a normative bridge exists.
It is:
Why is informationality itself a respect-making property — a property in virtue of which moral considerability obtains?
Reflective respect
One Floridian strand begins from reflective respect. Very roughly, an agent respects itself not merely as an isolated individual but as an instance of a class; recognizes itself as an informational entity; recognizes other informational entities as members of the same ontological class; and extends minimal respect across that class. [4]
There is a genuine universalizing idea here. But shared membership in a class does not by itself establish that the shared property is the reason respect is due.
Suppose A is morally considerable and has property P. B also has P. Nothing follows about B’s moral status in virtue of P unless P has independently been shown to be among the properties that make A morally considerable.
A person has mass; so does a rock. Whatever grounds respect for the person, the common predicate of possessing mass does not thereby become respect-making.
The point is not that Floridi literally begins with a person, subtracts properties one by one, and assumes that value remains. His project is explicitly a search for the minimal condition of possible intrinsic worth. The objection concerns the justification of that proposed minimum.
Abstraction can identify a maximally general predicate. It does not establish that maximal generality preserves the normative significance found at richer levels of description.
Informationality may indeed turn out to be morally relevant. But its universality cannot itself be the reason for its moral relevance.
Respect by default
Floridi also develops a different line of reasoning, and it is important because the reflective-respect objection does not answer it.
In his replies to critics, he argues that because sufficient arguments against the intrinsic value of Being in all its manifestations are lacking, the presumption should favor respect. The moral stance is defeasible: entities are to be treated as worthy of respect by default, in his memorable formulation, “until proven guilty.” Floridi also places this discussion beside a classical double-negation schema, while describing holistic respect as a presupposition or default. [5]
Here a technical distinction matters.
In classical logic:
¬¬P → P
is valid.
But suppose J(P) means that adequate justification is available for asserting P. Then:
¬J(¬P)
— there is no adequate justification for asserting not-P —
does not, by classical logic alone, establish:
J(P).
The absence of sufficient justification for denying P is an epistemic condition. It is not identical to the object-level proposition ¬¬P.
Floridi need not be accused of confusing those formulas. His argument contains more than a formal inference. He explicitly introduces a presumption or default.
That is exactly why the interesting question lies elsewhere.
A rule such as
If there is no sufficient reason to deny intrinsic value, presume intrinsic value
can be perfectly coherent. But its asymmetry now requires justification.
There are at least three available epistemic positions when evidence is insufficient: presume P; presume ¬P; or suspend judgment about P. Classical logic does not select among them.
There is also a further distinction that matters here. Floridi’s “until proven guilty” language might be interpreted not as a rule about what we should believe, but as a rule about how we should act under uncertainty. Perhaps Information Ethics need not say, “Believe every entity has intrinsic value until refuted.” It could instead say, “When uncertain whether an entity has value, treat it respectfully unless there is sufficient reason not to.”
That is a legitimate retreat from a doxastic principle to a precautionary policy.
But it does not remove the asymmetry problem. It changes its form.
For a belief rule, the question is: Why does absence of evidence against value justify positive belief rather than suspension of judgment?
For a practical rule, the question becomes: Why should the expected moral cost of falsely denying value outweigh the cost of unnecessarily attributing respect?
“Innocent until proven guilty” illustrates the point. It is not a theorem of classical logic. It is a decision rule under uncertainty whose asymmetry is independently motivated by judgments about rights, institutional power and the relative gravity of erroneous conviction and erroneous acquittal.
If “valuable until proven otherwise” functions analogously, Information Ethics needs its corresponding account of error costs.
Get Edan David Eyon’s stories in your inbox
Join Medium for free to get updates from this writer.
Perhaps such an argument can be given. One might hold that unwarranted destruction or disrespect poses a greater moral danger than initially granting minimal, defeasible consideration too widely. That could support a precautionary ethic.
But it would be the required moral argument.
Classical double negation would still not supply it.
A burden-of-proof rule can distribute dialectical or practical obligations. It cannot justify the evaluative asymmetry that determines that distribution merely by invoking classical logic.
Ontic trust
Floridi then offers a still richer foundation for his ontocentric ethics: ontic trust.
The idea begins from a genuine feature of existence. Agents do not create themselves. They come into being within an already existing world and depend upon entities and conditions they did not choose. Floridi models this through a hypothetical trust: reality is the corpus; prior agents occupy a donor-like role; current agents are trustees; present and future agents and patients are beneficiaries. [5]
The descriptive starting point is unobjectionable. Existing agents are causally and existentially dependent upon an antecedent world.
Floridi then gives that dependence an ethical interpretation. The relation to reality should also be understood caringly; participation in reality becomes associated with a minimal claim to existence and with relations of respect, gratitude and care. Agents become trustees rather than merely beneficiaries of the world. Information Ethics ultimately treats every instance of Being as possessing a minimal, initial and overridable claim to exist and develop according to its nature. [5]
Again, such an ethical interpretation is not illegitimate merely because it contains normative content.
The problem is that the descriptive relation does not determine that interpretation by itself.
Existential dependence does not entail trusteeship. Participation in reality does not by itself entail a right to continued existence. Having received one’s existence through antecedent conditions does not logically entail gratitude toward everything involved in those conditions.
Floridi describes ontic trust as postulated by Information Ethics. That does not mean arbitrary stipulation; theoretical postulates may be defended through explanatory power, coherence or independent normative arguments. But it does locate the relevant philosophical work. Ontic trust is not simply another empirical consequence of informational ontology. It is a normative interpretation of what existential dependence means for agents.
And that interpretation must earn its direction.
Three strands, one orientation
Reflective respect, default respect and ontic trust should not necessarily be treated as three independent proofs. Floridi can reasonably understand them as mutually supporting expressions of one ontocentric ethical orientation.
The fairest question is therefore whether their combination supplies what each strand lacks on its own.
Reflective respect emphasizes common informational being, but commonality does not establish informationality as a respect-making property. Default respect gives positive or precautionary priority to value, but that priority requires an account of the asymmetry that warrants it. Ontic trust interprets existential dependence through trusteeship and care, but dependence does not by itself determine that normative interpretation.
Placed together, the strands form a coherent picture: shared informational being, a presumption of respect, and caring trusteeship.
Coherence matters. But mutually compatible normative ideas do not automatically ground one another. Unless one strand supplies an independent reason for the evaluative orientation rather than expressing that orientation in another form, the package may describe Floridi’s axiology more clearly than it justifies its foundation.
That is not yet a demonstration of circularity. It is a request to identify where the evaluative asymmetry is actually earned.
Why should Being, merely as Being — or information, merely as informational existence — begin on the positive side of the axiological ledger?
The richer-universe test
Floridi’s suggestion that an ontologically richer universe can be morally better than a poorer one provides a useful boundary test.
Imagine two otherwise identical universes, except that one contains an additional unique informational object.
Is the richer universe, for that reason alone, however minutely better?
Floridi can consistently answer yes. His ethics is deliberately ecumenical; the value involved can be extremely small and defeasible; and producing or preserving the additional object may have countervailing costs.
The point is therefore not to expose a contradiction.
It is to clarify the reach of the commitment.
Would an arbitrary unique data structure count? A random physical pattern? An informational object with no function, beneficiary, experience, history or semantic significance?
If yes, then Information Ethics assigns positive value at an extraordinarily abstract level, and that consequence should be defended as part of the theory rather than deriving intuitive support from examples in which richer morally salient properties are present.
If no, then some further property distinguishes value-adding informational entities from informational entities whose existence makes no axiological difference.
Either answer is available.
The test simply asks the theory to show where the boundary actually lies.
Opposite directions, the same philosophical danger
Critchley and Floridi can now be compared without pretending that they commit one identical formal error.
Critchley’s characteristic move is modal. A cluster of properties earns authority by accurately describing familiar human experience, and that authority begins supporting claims about what meaning or intelligence must necessarily require.
Floridi’s metaphysical move is potentially ontological. Informational language earns authority as an elegant, scalable and powerful interpretation of differentiated structure, and the question becomes whether that methodological success also warrants treating the informational interpretation as a substantive account of ultimate reality.
Floridi’s ethical move is axiological. Informationality earns authority through maximal generality — it is the category broad enough to include every informational entity — and that generality participates in an attempt to establish universal minimal moral standing through reflective respect, default respect and ontic trust.
The common structure is now more specific than “each conclusion needs more argument.”
In every case, a feature that made the category successful in role A begins functioning as prima facie support for stronger role B.
Human experiential centrality begins supporting necessity.
Informational explanatory generality begins supporting ontological fundamentality.
Informational universality begins supporting moral inclusiveness.
Those promotions may ultimately be defensible. But the original success does not establish them.
That is category inflation.
Why AI makes the problem unusually visible
Artificial intelligence provides a particularly useful stress test because it separates properties that familiar human life tends to bundle together.
Intelligence may involve discrimination, generalization, inference, planning, abstraction, problem solving and self-correction. Agency may involve goal-directed control, adaptive behavior, autonomy or intentional action. Neither category by itself decides whether a system is sentient, conscious, capable of valenced experience, genuinely concerned about anything, or morally considerable.
The exact taxonomy matters less than the method: decomposition blocks illicit package-deals.
A system might reason without being sentient. It might act adaptively without phenomenal consciousness. It might manipulate representations without caring about their contents. Conversely, evidence for one morally relevant property in an artificial system would not automatically confer every other characteristic associated with human persons.
When a category becomes overloaded, the first task is not to decide who owns it.
It is to find out what work each component is actually doing.
Philosophy without the mythology
There is also a conception of philosophy implicit in this method.
Philosophy need not be protected as the uniquely human exercise of an ineffable faculty called judgment. Nor does it need to discover one privileged vocabulary — informational or otherwise — in which the whole of reality can finally be stated.
It becomes especially useful when the conceptual architecture of an inquiry itself needs examination.
Science may ask what happens and under what conditions. Engineering may ask how a desired behavior can be produced reliably under constraints. Mathematics may ask what follows from specified structures and assumptions. Philosophy repeatedly turns those practices back upon their own categories: What exactly are we distinguishing? Does this conclusion follow from that premise? Is the relation contingent or necessary? Has the success of a representation begun functioning as evidence for an ontology? Does a common predicate actually ground the normative status attached to it? What justifies the default under which a burden of proof has been distributed?
These questions are not sealed off from science, mathematics, engineering, linguistics or artificial intelligence. They run through all of them.
Nor does conceptual decomposition settle everything. Distinguishable properties can stand in necessary relations. A powerful abstraction may reveal genuine ontology. A normative primitive may be entirely legitimate. A practical default may be justified by an independently defensible asymmetry of risks.
The discipline is not to forbid such moves.
It is to identify when the authority of one role is being asked to support another.
Preserve the distinction. Make the promotion earn its weight.
Judgment is not identical to calculation. Experience is not identical to description. Structural reality is not automatically informational reality. Functional or organizational significance is not automatically moral significance. Intelligence is not sentience. Agency is not consciousness. A model is not the thing modeled.
These distinctions matter, but distinction alone settles remarkably little about the deeper relation between the things distinguished.
Critchley allows the phenomenological authority of familiar human experience to support stronger claims about modal necessity. Floridi provides a serious structural-realist argument, but the methodological power of informational description does not by itself settle the ontological status of the informational interpretation. His Information Ethics then faces the more difficult task of explaining why the maximal generality of informational Being should make informationality respect-making, why positive or precautionary respect should receive the default under uncertainty, and why existential dependence should be understood through trusteeship and care.
These are not demands for infinite regress. Some philosophical commitments will be primitive.
They are demands to locate the promotion.
The methodological rule is therefore simple:
Distinguish first. Operationalize where possible. Follow the inference. When a category moves from describing, organizing or explaining something to determining what must exist or what ought to matter, make the promotion explicit — and make it earn its weight.
Disclosure
This essay was written in full by ChatGPT (GPT-5.6 Sol).
I initiated and directed the inquiry, chose what to investigate, supplied additional critiques, requested rebuttals rather than automatic acceptance of those critiques, and decided which revisions to retain and when the essay was ready to publish. The prose itself — including its structure, wording, formalizations, revisions, and final composition — was generated by ChatGPT.
The critical process was also substantially AI-mediated. Some of the adversarial reviews used during revision were produced by ChatGPT in another conversation and by a GPT I configured for critical review. Those critiques were then brought back into the conversation, challenged where appropriate, and used to revise the essay.
The conversation in which the essay was developed — including revisions, objections, rebuttals, discarded arguments, and the final text — is available here:
https://chatgpt.com/share/6abe50e5-e780-83eb-bdfa-a88147941afa
Sources
[1] Simon Critchley, “The Id, the Ego and the Superintelligence,” The New York Times, September 27, 2026.
[2] Luciano Floridi, “Against Digital Ontology,” Synthese 168, no. 1 (2009): 151–178.
[3] Luciano Floridi, “A Defence of Informational Structural Realism,” Synthese 161, no. 2 (2008): 219–253.
[4] Luciano Floridi, “On the Intrinsic Value of Information Objects and the Infosphere,” Ethics and Information Technology 4, no. 4 (2002): 287–304; and related early formulations of reflective respect, ontic uniformity and ontic solidarity.
[5] Luciano Floridi, “Understanding Information Ethics: Replies to Comments,” APA Newsletter on Philosophy and Computers 8, no. 2 (2009): 4–11.
