Good sense, for, by, and to AI

Descartes opens the Discourse on the Method with a small joke that has outlived nearly everything else in the book: le bon sens est la chose du monde la mieux partagée — good sense is the best distributed thing in the world, since everyone is so well supplied with it that even those hardest to please in every other respect do not typically want more of it than they already have.1 The joke has a barb in it. Precisely because everyone is presumed to have good sense in full measure, almost no one thinks to cultivate it. Descartes' own definition is exacting: bon sens is the power of judging well, of distinguishing the true from the false — a power distinct from simply possessing it, which requires that it be well applied, through method, and not merely carried around as a birthright.

Kant, a century and a half later, reaches for a related but distinct idea in the Critique of the Power of Judgment: sensus communis, which he insists is not vulgar opinion but a properly communal sense — the faculty of reflective judgement that measures a particular case not against a rule already given but against the standpoint of everyone else. His three maxims of a common understanding are exact: think for oneself; think from the standpoint of every other person; think consistently with oneself.2 Hannah Arendt, reconstructing Kant's political philosophy two centuries later, insists on translating sensus communis as "community sense" rather than "common sense" for exactly this reason, and locates judgement in two moments: a first, private act of perception, and a second, properly public moment in which that perception is tested by imagining the presence of others — an "enlarged mentality" that judgement cannot do without.

Antonio Gramsci, writing from a fascist prison two centuries after Descartes, drives the wedge deepest. Italian preserves a distinction English blurs: senso comune, the disparate, largely unconscious, often superstitious sediment of inherited belief that a community carries without examining it, and buon senso — good sense — the small, sound, critical kernel embedded within that sediment, capable of being drawn out and made the basis of a more coherent understanding. Gramsci is unambiguous that the two are not the same achievement wearing different names: senso comune is, in his own words, "uncritical and largely unconscious," while buon senso is what a philosophy of praxis must extract and refine from it.3 Common sense, on this reading, is not a modest version of good sense. It is the raw material good sense works on.

Artificial intelligence inherits a third, more technical sense of the same phrase, and it is worth holding all three apart with some care. When John McCarthy identified "common sense" as the obstacle that would make machine reasoning "damnably hard," and when McCarthy and Hayes later gave that obstacle a name — the frame problem, the difficulty of specifying everything that does not change when an action occurs, without enumerating the world down to its last irrelevant fact — they meant neither consensus nor a critical faculty.4 They meant the vast, largely tacit stock of world-knowledge a system needs simply to avoid absurdity: that water is wet, that unsupported objects fall, that a broken glass does not mend itself by being observed. Seventy years, several symbolic projects, and now enormous language models later, the verdict of researchers closest to the problem remains sobering. As Yejin Choi has put it, artificial intelligence continues to be narrow and brittle for want of exactly this kind of knowledge, however fluent its surface performance has become.5

A system may hold an encyclopaedia of facts, and faithfully mirror the average of human opinion, and still misjudge — because none of that is the same achievement as judging well.

Three things, then, travel under one name. Common sense as consensus: the statistical centre of what a population happens to believe. Common sense as world-knowledge: the tacit grasp of how things hang together that the frame problem shows to be so difficult to specify. And good sense as the faculty of judging well: Descartes' bon sens, Kant's sensus communis, Gramsci's buon senso — the power to reach a sound, corrected verdict, which is conceptually independent of both the others. A system could in principle possess an enormous stock of world-knowledge, and reliably predict what most people would say, and still lack this third thing entirely. It is this third thing — not the first two, and not their sum — that GoodSense.ai takes as its subject, and that this manifesto now tries to give a philosophical foundation broad enough to build on.

David Hume gives our first pillar: not a rule for good sense to follow, but an account of what good sense is. His starting point is deliberately deflationary. "Reason is, and ought only to be the slave of the passions, and can never pretend to any other office than to serve and obey them."6 Reason, for Hume, discovers relations among ideas and matters of fact; it is extraordinarily good at finding means to ends. What it cannot do, on his account, is furnish the end itself. An engine of pure inference, however powerful, is motivationally inert until something else — a passion, a sentiment, a desire — gives it something to want. This is not a slight against reason so much as a warning against expecting from it a kind of labour it was never built to perform.

If reason cannot supply our ends, still less can it supply our moral verdicts out of nothing. Hume's most quoted paragraph — quoted so often it has acquired its own name, Hume's Guillotine — observes that authors of moral systems proceed for a while in the ordinary way of reasoning, establishing that this or that is the case, and then at some point, "imperceptibly," begin speaking instead of what ought or ought not to be — a change Hume calls "of the last consequence," since no one has explained how a conclusion of that entirely different kind could follow from premises that contain nothing but the first.7 No amount of is, purely descriptive, entails an ought. The gap is not a puzzle to be cleared up with more data; on Hume's own view, it is structural.

What, then, does supply our moral verdicts, if not reason working alone? Sentiment — but not raw, private sentiment, and this is the turn that matters most for our purposes. Hume held that unrefined sympathy, felt more strongly for the near and dear than for the distant, would produce nothing but contradiction if left to itself: your judgement of a person's character would depend on how you happened to stand in relation to them, and mine on how I happened to stand, and the two would rarely agree. To speak a common moral language at all, Hume argues, we correct this bias by adopting a steady and general point of view — asking not how a trait strikes me from where I sit, but how it typically affects those whom the person usually deals with, discounting my own accidental position entirely. Moral language, on this account, is already an invitation to concur: to speak evaluatively is to appeal to a standard others can share once their sentiments, like one's own, have been generalised in the same way.8

Good sense, on a Humean reading, is not what most of us happen to feel. It is what any of us would feel, once the bias of place and person has been corrected away.

This is the single most important consequence Hume's ethics holds for our project, and it is why consensus cannot do the work good sense is asked to do. A show of hands is not a corrected sentiment; it is simply many uncorrected sentiments, tallied. Hume distinguishes further between calm and violent passions — what looks like reason governing passion, he suggests, is often only a calm passion quietly overruling a violent one — and he treats custom, "the great guide of human life," as the ordinary vehicle by which such correction becomes habitual rather than laborious.9 Good sense, in other words, is not deduced fresh each time; it is a cultivated disposition, exercised through habituated correction, that nonetheless answers to a standard beyond any single sentiment taken raw.

The bearing of all this on artificial intelligence is direct, and increasingly discussed under a different name. Systems tuned by reinforcement learning from human feedback are, in the relevant sense, fitted to descriptive data: records of what human raters did prefer, when asked. Hume's gap says plainly that no amount of such fitting, however precise, entails what those raters — or anyone else — ought to prefer. Recent work making exactly this argument has named the is–ought gap, alongside the plurality of human values and an extended version of the frame problem, as the deepest reasons why alignment approaches built on fitting content to preference data cannot, by themselves, succeed.10 If Hume is right, the lesson is not that value learning is hopeless but that it has been aimed at the wrong target: not at reproducing the aggregate of what is preferred, but at building in the corrective process — the move from raw sentiment to the general point of view — that turns preference into judgement. That correction is what good sense for AI must try to engineer.

Hume tells us what a sound verdict is made of. He does not tell us how artificial systems come to matter socially in the first place — how a piece of software becomes woven into the fabric of institutions, categories, and entitlements that make up a society. For that, our second pillar is the social ontology of Brian Epstein, whose 2015 book The Ant Trap sets out to rebuild the foundations of the social sciences on more careful metaphysical footing.11 Epstein's target is what he calls the Standard Model of social ontology — the view, traceable through Hume, Hart, and Searle, that social objects are essentially projections of our attitudes and agreements onto an otherwise non-social world.12 His claim is that this picture, however natural it feels, is not generally true, and that seeing why requires two distinctions the tradition has run together.

The first is grounding: the relation by which a particular set of facts is, quite simply, the metaphysical reason a further fact obtains, on a given occasion. The fact that a particular banknote is legal tender, for instance, is grounded — on that occasion — by facts about its printing, issuance, and continued circulation within a monetary system. Grounding is not causation; it is closer to constitution, the "in virtue of" that links a specific instance to the specific facts that make it true.13

The second, and Epstein's more original contribution, is anchoring. Searle had argued that institutional facts are fixed by "constitutive rules" of the form X counts as Y in context C — rules Epstein renames frame principles, since, as he puts it, they are "neither constitutive nor rules," in something like the way the Holy Roman Empire was neither holy, nor Roman, nor an empire.14 A frame principle states the grounding conditions for a kind of social fact across an entire range of possible situations — not just what makes this note legal tender today, but what would make any note legal tender, in any circumstance covered by the principle. Anchoring is the relation that puts that frame principle itself in place: the deeper set of facts — a founding statute, an institutional history, a pattern of enforcement — that explains why these grounding conditions govern, rather than some others. Grounding fixes a fact, given the frame; anchoring fixes the frame.

Both relations, Epstein insists, run wider than most of the tradition assumes. Ontological individualism — the thesis that social facts are nothing over and above facts about individuals and their relations — is, on his account, false even for comparatively simple cases: a mob or a market can have properties not exhausted by any enumeration of the individuals composing it. A closely related but distinct thesis, anchor individualism — the claim that facts about individual people anchor the frame principles governing every social kind — fares no better as a universal claim, and Epstein shows the two individualisms to be, if anything, in tension with one another: the more one is true, the less likely the other becomes.15 Money and recycling bins may indeed be anchored by our attitudes; but a recycling bin, once anchored, is still heavy, made of plastic, and grounds further facts — about weight, capacity, decomposition — that have nothing to do with anyone's attitude toward it at all.

Anchoring, not grounding, is where an artificial system does its deepest social work — not in getting this or that case right, but in quietly resetting what counts as a case of anything at all.

Artificial systems, we want to suggest, are now doing both kinds of metaphysical work at once, and the distinction between them is exactly what a well-intentioned but under-philosophised AI ethics tends to miss. When a model scores an application, flags a transaction, or moderates a post, it participates in grounding: it supplies, for this instance, part of what makes it true that the applicant is high-risk, the transaction fraudulent, the post in violation. That is consequential, and errors there are real harms. But the deeper transformation is at the level of anchoring. Widespread algorithmic scoring does not merely apply an existing concept of creditworthiness case by case; over time, it can quietly become part of what anchors the frame principle for creditworthiness itself — resetting, often without any public debate, what kind of thing a person now has to be to count as fundable, employable, or safe. This is a close cousin of what Ian Hacking called the looping effect of human kinds: a classification acts on the people it classifies, who adjust their behaviour in response, which in turn reshapes the classification, in a loop that has no natural stopping point.16 Epstein gives that loop its metaphysics. Good sense, on this pillar, cannot content itself with auditing outputs one grounding-instance at a time; it has to reach the anchoring — the frame principles a system is quietly resetting — or it will have addressed only the shallower of the two problems Epstein distinguishes.

If Hume supplies the content of good sense and Epstein the mechanism by which artificial systems reach into social reality, Luciano Floridi supplies the stage on which all of this now happens, and the dynamic that makes it urgent. Floridi's infosphere — a term he reintroduced in 1999, on the model of "biosphere" — names the whole informational environment: every informational entity together with its properties, interactions, and relations, online and offline, digital and analogue alike.17 Information technologies, on his account, do not merely add new tools to an otherwise unchanged world. They re-ontologise it — they change, at a level deeper than mere use, what the environment fundamentally is and what kind of beings we are within it. The online is not a separate compartment from the offline; it is progressively absorbed into a single environment we simply inhabit.

Floridi frames this as a fourth and, in a sense, final decentring. Copernicus displaced us from the centre of the physical universe; Darwin, from the centre of the biological kingdom; Freud, from transparent access to our own minds. Turing's revolution, on Floridi's reading, displaces us from the centre of the infosphere: we are no longer the only, nor even reliably the most capable, information-processing agents within our own environment.18 We increasingly live, in his coinage, as inforgs — informational organisms whose identity is bound up with the information we generate, consume, and are constituted by — moving through an onlife in which the online and offline no longer separate cleanly, in how we work, learn, love, shop, or vote.

This re-ontologised environment demands, on Floridi's account, an ethics fitted to its actual scope. His information ethics is deliberately ontocentric rather than anthropocentric or even merely biocentric: every informational entity, simply in virtue of existing as such, carries some minimal, overridable moral standing, and is owed at least a disinterested, careful attention as a possible object of care. The fundamental evil, on this view, is entropy — not the entropy of thermodynamics or of Shannon's information theory, but a specifically metaphysical entropy: the destruction, corruption, or impoverishment of the infosphere's diversity and structure. An action is good, on this reading, roughly to the degree that it protects or enriches that structure, and bad to the degree that it degrades it. The distinction matters and is easy to blur: Floridi's entropy is a moral, not a physical, quantity.

Two further moves make this ethics usable for artificial agents specifically. First, Floridi and Sanders argue that an entity counts as an agent, at an appropriate level of abstraction, if it shows interactivity, autonomy, and adaptability — and that such an agent can be a source of moral good or harm, and so a proper object of moral evaluation, without possessing consciousness, intentions, or free will at all. They call this "mind-less morality": moral agency without a mind behind it.19 Second, where many hands — human and artificial — jointly produce an outcome no one of them intended, Floridi argues that the resulting harm or benefit can be genuinely distributed, with responsibility to be allocated across the multi-agent system rather than pinned entirely on any single node within it.20 Together, these moves let us ask what an AI system is doing, morally speaking, without first having to settle whether it understands anything at all — a distinction Floridi has more recently sharpened by describing large-scale AI as a new kind of agency that can succeed without being intelligent, a decoupling he thinks the field would do well to take seriously rather than argue around.21

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.

Floridi's most useful gift to this manifesto, though, is dynamic rather than static: the picture of a society and its technologies as locked in a continuous, recursive circle, each shaping the conditions under which the other develops. His own applied framework — beneficence, non-maleficence, autonomy, justice, and a fifth principle he insists AI ethics needs beyond the four inherited from bioethics, explicability — belongs to what he calls soft ethics: not what the law compels, but what we ought to do once compliance has already been secured.22 The circle itself is visible with unusual clarity in recommender systems, where the choices users make train the very models that will go on to shape the choices future users are offered — a recursive loop, now well documented, that left unattended tends toward homogenisation of taste and the narrowing of what a person is even shown.23 Nothing about the mechanism itself decides whether that loop is virtuous or vicious. That is precisely where good sense — Humean in its content, Epsteinian in where it must intervene — has to do its work, turn after turn.

The three pillars are not offered as competing accounts of the same thing. Each answers a different question, and the divisions of labour between them is, we think, exact enough to build on.

This division of labour lets us read "good sense for, by, and to AI" as three genuinely distinct claims rather than one slogan repeated three times — the design that precedes a system, the judgement it exercises while running, and the standard by which we evaluate what it has done.

Prospective · ConstitutiveGood sense for AI

"For" names a design desideratum applicable to any system we are willing to call AI, whatever its substrate. To build good sense in is not to hard-code a fixed table of values and call the matter settled; a fixed table is precisely what the is–ought gap warns us cannot substitute for judgement. It is to engineer the corrective process itself — the Humean move from raw, biased sentiment to a generalised, common point of view — as an operating feature of the system, together with the explicability Floridi names as AI ethics' distinguishing fifth principle, so that the process remains visible and answerable rather than opaque. And because Epstein shows that the deepest social stakes lie in anchoring rather than mere grounding, good-sense-by-design has to ask, at the design stage, not only "will this classify correctly?" but "what social kind will repeated use of this system anchor, and is that a kind we can defend?" A system that scores well on the first question and has never been asked the second has not yet been designed with good sense at all.

Operational · AgentiveGood sense by AI

"By" names the system as an exerciser of good sense: describing and acting upon the world through a corrected, context-sensitive lens, in the act itself. The obvious worry is that this asks too much — that judgement of this kind presupposes exactly the consciousness, intention, or understanding an artificial system may never have. Floridi's decoupling of agency from intelligence is the answer to that worry, not a way around it: an entity can be a moral agent, a proper site of good or harm, at an appropriate level of abstraction, without any mind behind it at all. To ask a system to exercise good sense "by" itself is therefore not to ask it to feel the corrected sentiment Hume describes; it is to ask that its agency be structured, functionally, so as to track what corrected sentiment would counsel, and that its classifications be made with an operative sensitivity to which social kinds — in Epstein's sense — they are actively grounding and anchoring as they run. Good sense by AI is a claim about architecture, not about inner life.

Evaluative · ExternalGood sense to AI

"To" is the retrospective mode, and arguably the most urgent, since it is the only one of the three that already applies to nearly every system now deployed. Most existing AI was not designed with anything resembling good sense in mind. That does not exempt it from the standard; it only means the standard must be applied after the fact, to consequences rather than intentions. This is where Floridi's patient-oriented, ontocentric ethics does real work: the question is not merely who acted, but what, and whom, was affected — which informational entities, which social kinds, were degraded or enriched. Epstein's framework lets us name the damage precisely, as a corruption of some specific frame principle rather than a vague diffuse harm. And Hume's general point of view supplies the standpoint from which the affected public — not the system's designers, and not the system itself — renders the verdict: given what this system did, once we correct for whose interests happened to be served by its design, would good sense have counselled differently? Where the answer is yes, good sense applies to that system, whether or not anyone ever intended it to.

Design without operation is a wish. Operation without evaluation is a drift. Evaluation that never feeds back into design is only a postmortem. Good sense is the name for the loop closing properly.

The three modes are not sequential stages one passes through once. They are the same circle Floridi describes, seen from three positions on its circumference: for feeds by, whose consequences are judged to, whose verdicts flow back into the next round of for. Nothing about the loop's structure guarantees it turns virtuous rather than vicious. Good sense is not a fourth ingredient added to the mechanism from outside; it is simply what we call the loop when the correction is actually made, turn after turn, rather than merely wished for once at the start.

An honest synthesis names its fault lines rather than papering over them, and this one has at least three that a careful reader will find before we point them out. We would rather name them ourselves.

None of these tensions is an embarrassment to be managed away by better phrasing. Refusing to force a single discipline's monopoly on judgement means living with friction between disciplines that were never built to agree with one another. We take that friction, held openly rather than concealed, to be itself a small exercise of the good sense this manifesto is trying to describe.

We began with a joke at good sense's expense — that everyone believes themselves amply supplied with it, which is exactly why so few trouble to cultivate it. We end by hoping the same complacency does not settle over the machines we are building in our own image. Common sense, in whichever of its senses, has proven hard enough for artificial intelligence to earn. Good sense is a harder, and we think more urgent, achievement still — and one no single discipline can secure alone.

Good sense, well exercised, is not a destination. It is the correction we keep making, together, each time the circle turns.

This manifesto is a foundation, not a conclusion. We invite philosophers, social ontologists, psychologists, engineers, and the simply curious to press on its seams, including the ones we have named ourselves.