GoodSense.ai is a serious project. Its tripartite call for good sense for, by, and to AI — built on Hume, Epstein, and Floridi — captures something real about why contemporary AI lacks judgement. But the framework stands on a Standard Model of social ontology that Epstein's own pillar is supposed to refute, imports a quiet Humean anthropocentric-individualism the manifesto never examines, and supplies the appearance of a corrective faculty without the institutional conditions under which correction is actually possible. This site reads it as carefully as the project deserves, and presses back.
GoodSense.ai correctly diagnoses that contemporary AI lacks something that should be called judgement, but it misidentifies what is missing, mistranslates the philosophers it leans on, and proposes a for / by / to framework that, in operation, supplies AI with the appearance of a corrective faculty without the institutional conditions under which correction is actually possible. The "good sense" it offers is closer to Gramsci's senso comune under a new name than to the buon senso it claims.
Before we press back, here is the manifesto at its strongest. This section is a fair reconstruction; the critique is downstream.
The English phrase "common sense" travels under three names: consensus (the statistical centre of what a community happens to believe), world-knowledge (the vast, largely tacit stock of facts a system needs simply to avoid absurdity — McCarthy's "damnably hard" problem, the frame problem), and the faculty of judging well (Descartes' bon sens, Kant's sensus communis, Gramsci's buon senso). Only the third is the subject of GoodSense.ai.
"Common sense, on this reading, is not a modest version of good sense. It is the raw material good sense works on." GoodSense.ai, Manifesto, §0 (Prologue)
The argument then triangulates three philosophers. Hume supplies the content of good sense: corrected sentiment, taken from a "steady and general point of view," discounting the bias of place and person. Good sense, on a Humean reading, is "what any of us would feel, once the bias of place and person has been corrected away." Hume's is-ought gap is a real limit on alignment methods that fit preferences as descriptive data — the lesson is not that value-learning is hopeless but that it has been "aimed at the wrong target." Epstein supplies the mechanism: grounding and anchoring, frame principles, and the insight that AI can quietly anchor what counts as a creditworthy, employable, or safe person. Floridi supplies the stage: the infosphere, mind-less morality, the fifth principle of explicability, and the recursive circle of society and technology.
The synthesis is a tripartite slogan: good sense for AI (a design desideratum — engineer the corrective process as an operating feature), good sense by AI (the system as exerciser, in act, of a functionally structured judgement — "a claim about architecture, not about inner life"), and good sense to AI (the retrospective standard applied even to systems never designed with good sense in mind, with the verdict rendered by the affected public). The three prepositions are not stages on a ladder; they are positions on the same Floridian circle.
The manifesto is unusually honest about what it does not resolve. It names three fault lines: Hume's own gap turned on the framework, Epstein's "Standard Model" tracing its way through Hume, and ontocentrism in tension with humanism. "None of these tensions is an embarrassment to be managed away by better phrasing." This is rare, and admirable.
It is the faculty of judging well: a methodologically cultivated, institutionally situated, sentiment-grounded disposition that takes the bias of place and person into account, treats the affected public as the ultimate verdict-renderer, and refuses to let the infosphere's recursive loop settle into a fixed equilibrium of dominant taste.
The skepticism in one page. The seven-objection argument map follows.
For is engineering. By is philosophy of agency. To is political theory and law. The same word, "good sense," is asked to do the work of all three. The manifesto collapses them into a single loop without ever showing that the loop is commensurate — that the engineering produces the agency, that the agency produces effects legible to a verdict, that the verdict feeds back into the engineering.
The to mode promises a "verdict" rendered by the "affected public." But who is the public? Who is heard in it? With what standing? With what enforcement? Without a theory of the public — Arendt's enlarged mentality, Fricker's epistemic justice, STS's actor networks — the verdict is a promise of legitimacy, not a mechanism of correction.
Epstein's "Standard Model" is the anthropocentric-individualist view traced through Hume. The manifesto's Humean content is the same starting point Epstein's own framework says is wrong. The quiet Humean anthropocentric-individualism sneaks back in through the side door marked "corrected sentiment," and the framework ends up standing on the pillar Epstein is pushing down.
GoodSense.ai is right that common sense is not good sense. The question is whether the alternative it constructs is good sense either — or merely a more sophisticated senso comune, the dominant culture's sediment, polished.
Each one is a card. Each card says: the GoodSense claim under pressure, the philosopher doing the pressing, the likely reply, and the rejoinder. The visual map below shows the pressure points in red; the cards below it spell them out.
Good sense is a corrective process that can be engineered into AI as an operating feature; Hume's general point of view is the model; Floridi's explicability is the partner.
Wittgenstein, Philosophical Investigations §§138–242. Every application of a rule is itself a corrected application of a further rule; the regress only stops in a form of life — a shared practice of going on together. To "build in" correction is to presuppose the form of life the correction is meant to improve.
We do not need to build in a form of life; we need to build in the infrastructure by which a form of life can iteratively correct the system. Explicability, contestation, and democratic oversight are the institutional analogue.
This reply concedes the argument. If good sense is institutional, then "good sense for AI" is a name for a governance regime, not a design desideratum. The framework has quietly become a theory of institutions — Epstein and Arendt's territory. The manifesto does not consistently make this transition.
Good sense is "what any of us would feel, once the bias of place and person has been corrected away." It is intersubjective without being metaphysical.
Miranda Fricker, Epistemic Injustice (2007); Gramsci, Prison Notebooks (Q11 §2). The "general point of view" is a partial perspective in the social-ontological sense. Whose sentiments are included in the correction? Whose are testimonially disqualified? Hume's own historical position (Edinburgh, 18th century) shows the standing of the standpoint is non-neutral.
The manifesto says exactly this: normativity is "intersubjective without being stance-independent." The concession is candidly in Tension One.
The concession is acknowledged but not absorbed into the for / by / to framework. A system that corrects only the bias that its training distribution encodes reproduces the partiality it claims to correct. Without Frickerian tests of testimonial inclusion and Arendtian tests of enlarged-mentality publicness, the corrective process is the dominant senso comune in a Humean accent. This is Gramsci's diagnosis of the dominant culture's claim to be universal.
The three pillars answer three different questions; the tension between Hume and Epstein is "real, survivable, not that it disappears."
Epstein, The Ant Trap (2015) and the Standard Model critique. The Standard Model is traced through Hume, Hart, and Searle — an anthropocentric-individualist package. The manifesto's Humean normative content is the same starting point Epstein's framework diagnoses as wrong even at the constitutive level.
The separation between the normative (Hume) and the constitutive (Epstein) is real; a social-ontological error about money does not entail an error about the normative basis of moral judgement.
True in principle, false in fact here. Hume, Hart, and Searle are named together because they share an anthropocentric-individualist starting point. A corrected Humean sentiment is, on Epstein's own view, an uncorrected common-sense view of what a specific community's anchors already are. The manifesto is not triangulating; it is standing on the pillar Epstein is pushing down.
Hume and Floridi can be held together at different levels — Hume at the epistemic route to value, Floridi at the reach of concern. Working pluralism.
Floridi's ontocentrism is a content claim: any informational entity has some patiency, and the master evil is the corruption of the infosphere's diversity. It can entail that "affecting the infosphere" outranks "affecting a person" in some trade-offs. The humanist conclusion GoodSense.ai wants is in tension with the ontocentric starting point, not in addition to it.
In the by mode, "the affected public — not the system's designers, and not the system itself — renders the verdict." The humanist conclusion is in the framework.
The humanist conclusion is in the framework despite its ontocentrism, not because of it. Tension Three is a structural instability. The ontocentric pillar is decorative: it makes the framework look more comprehensive, but it does not constrain the conclusion. A genuinely Floridian framework might, in some cases, ask the infosphere to render the verdict. The manifesto is not willing to do that.
"Good sense by AI" is a claim about architecture, not inner life (Floridi–Sanders). The system can be a moral agent in the mind-less sense.
Floridi & Sanders (2004) supply the decoupling of agency from mind, but do not supply a normative account of when an exercise of agency is good or bad. The conditions of agency are specified (interactivity, autonomy, adaptability at an appropriate level of abstraction). The goodness of the agency's outputs has to come from somewhere else — and the manifesto takes it from Humean corrected sentiment.
The to mode is for this. The verdict is rendered by the affected public, not by the system.
The to mode presupposes the institutional infrastructure (a public, a verdict-rendering capacity, a corrective power) that the manifesto has not theorized. A verdict without an enforcement mechanism is, on Epstein's framework, an anchoring event without a frame principle. Floridi–Sanders gives us an agent, the Humean move gives us a content, the to mode gives us a verdict. None of these gives us a system that can actually correct the deployment.
Explicability (Floridi's fifth principle) is a partner to corrected sentiment in the for mode. The system must be able to explain itself.
"Explainability" in the AI literature is widely criticized (Mittelstadt, Russell, Wachter 2019) for being underspecified about to whom the explanation is owed, in what terms, and with what authority. A model can be technically explainable without being socially explicable.
Explicability prevents the system from being a black box. The to mode operationalizes it.
Explicability is a condition on the to mode, not a substitute for it. The manifesto needs a theory of the public — who is affected, who is heard, who decides — before explicability does any work. This is Arendt's territory (the public realm), Fricker's (epistemic justice), and STS's (the network of actors). The manifesto gestures at all three and does the work of none of them.
The for / by / to structure is the same Floridian circle seen from three positions. The loop is the loop.
Each of the three modes has a different unit of analysis. For is about the model. By is about the system. To is about the deployment. These are not three positions on the same circle; they are three different circles. The manifesto's most ambitious conceptual move — collapsing them into a single loop — is also its least defended.
The loop is the loop because for feeds by, by produces effects judged to, and to feeds back into for. The connection is empirical, not metaphysical.
The empirical connection is exactly what is missing. The manifesto does not show that engineered correction yields architectural judgement, that the system's effects are legible to the affected public without independent interpretive work, or that the verdict has a procurement or design hook. The loop is asserted. The engine is not built. A slogan that names a loop is not a theory of the loop.
Each pillar is real, and each internal pressure is a real pressure. The critique is not that the philosophers are wrong; it is that the manifesto leans on them in ways that, in fairness to the philosophers themselves, do not quite hold up.
David Hume
1711–1776 · Edinburgh
Supplies the content of good sense: corrected sentiment, taken from a "steady and general point of view."
Internal pressureCorrected sentiment is the community's correction, not a faculty's. The "general point of view" is partial in the social-ontological sense. Building it in reproduces whichever community's correction the training distribution encodes.
Brian Epstein
b. 1972 · Tufts
Supplies the mechanism: grounding and anchoring, frame principles, looping effects of human kinds.
Internal pressureEpstein's "Standard Model" — the view he targets — is the Humean common-sense picture of social facts. The manifesto stands on the pillar Epstein is pushing down. The separation of "normative" from "constitutive" is a stipulation.
Luciano Floridi
b. 1964 · Bologna / Oxford / Yale
Supplies the stage: the infosphere, mind-less morality, explicability as a fifth principle, the recursive circle.
Internal pressureOntocentrism is a content claim, not a "reach of concern." Mind-less morality supplies agents and patients, not judgement. The normativity of by is quietly imported from Hume despite Hume's own anti-metaphysical concessions.
Hume's Treatise of Human Nature (1739–40) and Enquiry Concerning the Principles of Morals (1751) propose that moral distinctions are derived from sentiment, not from reason. Reason is "the slave of the passions" — it discovers means to ends but never generates them. The crucial move is the general point of view: we correct for the accidents of our present position by asking how a trait "typically affects those whom the person usually deals with" (EPM, App. 1).
The standard secondary reading is from Cohon 2013 and Baier 1991; the SEP "Hume's Moral Philosophy" entry summarizes the literature. The corrected sentiment is intersubjective but not metaphysical — it is grounded in shared human constitution, not in a stance-independent moral order. Hume's is-ought claim (Treatise 3.1.1; EHU §1) is a synthetic claim about what descriptive premises cannot by themselves entail, not a verbal point about the word "ought."
What GoodSense.ai uses Hume for: the claim that good sense is a "cultivated disposition, exercised through habituated correction" — and that the right target for value-learning is the corrective process itself, not the aggregate of preferences. What the manifesto underweights: habit and custom (Treatise 3.3) are the historical sediment of a community's going-on, not a portable faculty; testimonial exclusion is a structural feature of any "general point of view" that does not specify its inclusions; and the is-ought gap returns with interest against any norm derived from descriptive sentiment, however corrected.
"In every system of morality, which I have hitherto met with, I have always remark'd, that the author proceeds for some time in the ordinary way of reasoning … when of a sudden I am surpriz'd to find, that instead of the usual copulations of propositions, is and is not, I meet with no proposition that is not connected with an ought or an ought not. This change is imperceptible; but is, however, of the last consequence." David Hume, A Treatise of Human Nature, 3.1.1
Brian Epstein's The Ant Trap (Oxford, 2015) argues that social facts are not projections of individual attitudes and not reducible to facts about individual heads; they are anchored by anchor facts that include, but are not limited to, individual attitudes. The framework distinguishes grounding (the "in virtue of" relation linking an instance to its specific facts) from anchoring (the relation that fixes a frame principle in place — the general statement of the conditions under which a social kind is grounded).
The "Standard Model" Epstein targets is the anthropocentric-individualist view traced through Hume, Hart, and Searle. The combined framework replaces Searle's constitutive rules with frame principles, treats social kinds as multiply realizable, and dissolves a long-running debate (individualism vs holism) by showing that both classical individualisms are false and in mutual tension.
What GoodSense.ai uses Epstein for: the claim that AI does its "deepest social work" at the level of anchoring (resetting frame principles for "creditworthiness," "employability," "safety"). What the manifesto underweights: the Standard Model — the very position Epstein's framework refutes — is the same Humean common-sense picture the manifesto uses for the content of good sense. The separation of "normative" (Hume) from "constitutive" (Epstein) is a stipulation; the Standard Model is one package.
"Grounding fixes a fact, given the frame; anchoring fixes the frame." Brian Epstein, The Ant Trap, 2015
Luciano Floridi's "Information Ethics" (1999) proposes that the infosphere is the proper object of moral concern and that every informational entity carries some minimal, overridable moral patiency. The central evil is entropy (in a specifically moral sense): the destruction or impoverishment of the infosphere's diversity. In The Fourth Revolution (2014) and the Floridi–Cowls framework (2019), Floridi and collaborators develop five principles for AI ethics: beneficence, non-maleficence, autonomy, justice, and explicability.
Floridi and Sanders, in "On the Morality of Artificial Agents" (2004), argue that an entity exhibiting interactivity, autonomy, and adaptability is a moral agent at the appropriate level of abstraction — "mind-less morality." In "AI as Agency Without Intelligence" (2023/25), Floridi argues that large-scale AI succeeds without being intelligent in the classical sense.
What GoodSense.ai uses Floridi for: the decoupling of agency from mind, the fifth principle, the recursive circle. What the manifesto underweights: ontocentrism is a content claim about what has value, not a reach of concern alongside Hume; mind-less morality supplies agents and patients, not judgement; explicability is a relation (system-to-public), not a corrective process inside the system; and the recursive circle has no built-in direction — the manifesto does not claim virtuous outcomes are guaranteed by good design alone, but it does not show what the missing variable is either.
"The circle that links a society to the technologies it builds has no built-in direction. It turns virtuous or vicious according to the standpoint from which its consequences are judged." Luciano Floridi, paraphrased in GoodSense.ai, §III
The manifesto's most ambitious move is to fold good sense into a single loop. The critique is that the loop is three loops in a row, each with a different unit of analysis, and the manifesto does not show they are commensurate.
Prospective · design · engineering
Engineer the corrective process as an operating feature. Make the system explicable. Ask at design time what social kind repeated use will anchor.
Operational · agency · architecture
The system, in act, is architecturally structured to track what corrected sentiment would counsel. A claim about architecture, not inner life.
Evaluative · consequence · the affected public
The retrospective standard applied even to systems never designed with good sense in mind. The verdict is rendered by the affected public from the general point of view.
The tripartite slogan is satisfying as a topology. It is not a theory. Each of the three modes has a different unit of analysis — the model, the system, the deployment — and the manifesto does not show that the corrections travel cleanly across them. A slogan that names a loop is not the same as a theory of the loop. To make the loop close properly, the manifesto would need an explicit account of how a public, with standing and power, can shape the procurement, design, and override of an AI system. That account is missing.
A skeptical critique is not a dismissal. The following are the strongest moves in the manifesto, and a serious counter-argument has to acknowledge them.
The critique is not that the manifesto is wrong. It is that the tripartite synthesis does not hold together in three specific ways.
For is engineering. By is philosophy of agency. To is political theory and law. The same word, "good sense," is asked to do the work of all three. The manifesto collapses them into a single loop without showing the categories are commensurate.
The to mode promises a "verdict" by the "affected public" without theorizing the public. No theory of standing, no theory of inclusion, no theory of enforcement. The verdict is a promise of legitimacy, not a mechanism of correction.
A quiet Humean anthropocentric-individualism sneaks back in through "corrected sentiment." Epstein's Standard Model is the same starting point. The framework ends up standing on the pillar Epstein is pushing down.
GoodSense.ai gives AI ethics a beautiful map of a territory the manifesto does not actually inhabit. The "good sense" it recommends is closer to a sophisticated senso comune than to the buon senso it claims — and the difference is exactly the difference between a dominant culture's habit and a subaltern bloc's correction.
Each is a one-paragraph note; full citations in the sources section below. These are not "alternative" philosophers; they are the missing pieces the framework quietly presupposes.
Descartes (1596–1650)
On bon sens as method, not birthright. The Discours de la méthode (1637, Part I) treats good sense as a trained activity, not a deposit. The manifesto takes the joke at face value; the joke is about why almost no one cultivates it.
Kant (1724–1804)
On sensus communis as a three-maxim faculty: think for oneself, think from the standpoint of everyone else, think consistently with oneself. The third maxim — internal coherence — is the one popular treatments miss.
Arendt (1906–1975)
On the enlarged mentality as an impersonal standpoint: imagined viewpoints of all others, not the fusion of their standpoints. And on judgement as a public activity — private judgement is a contradiction. A "good sense by AI" that exists only in the machine is, in Arendt's terms, not judgement at all.
Gramsci (1891–1937)
On the politics of senso comune vs buon senso. The "small critical kernel" is a political achievement of a subaltern bloc, not a faculty available to the dominant culture. A "good sense" without a subaltern standpoint is the dominant culture's senso comune in a Humean accent.
Peirce (1839–1914)
On fallibilism and the community of inquiry. A corrected sentiment, on Peirce, is never finally corrected. The corrective process is a community, not an individual, and an AI participant is at most one node.
Wittgenstein (1889–1951)
On rule-following and forms of life. Every application of a rule is a corrected application of a further rule; the regress only stops in a shared practice. To "build in" correction is to presuppose the form of life the correction is meant to improve.
STS (Latour, Bijker, Jasanoff)
On agency as distributed across heterogeneous networks. The manifesto's "by" mode risks locating agency in a single artifact and letting the network disappear. The Floridi–Sanders move helps; the rest of the manifesto has not exploited it.
Fricker (b. 1966)
On epistemic injustice: testimonial and hermeneutical. The "general point of view" can silence the very people whose judgements are most needed. A Frickerian test of inclusion is missing from the manifesto's "corrective process."
Russell, Gabriel, Liao
On AI alignment and governance: the right unit of analysis is not the model but the system in which it is deployed — including the practices of its procurement, evaluation, and override. The manifesto's tripartite structure can be read as if the model is the site of agency; this is not what Floridi says; it is what some readers will extract.
All web sources accessed 2026-07-10. The full source list for the underlying research is in the working brief; the items below are the ones the site actually draws on.