Monograph draft v1, merging Paper 6 (No One at Home But the House Still Burns) with the structural argument of Paper 3 (A Foundation for Ethical AI), per the build plan; Paper 3’s proof-of-concept and research programme are reserved for Chapter 6. Seams for the copy-edit are noted at the end.
4.1 The house, restated
Chapter 3 closed with a question — what is owed, by whom, to whom, and on which side of the interface — and this chapter exists to stop the book evading it. Everything so far has been description: what the ghost is, where it arises, what the anatomy of a mind contains. Description is safe. But the systems described are not sitting in a seminar; they are answering questions, writing policies, steering attention, and being wired into the machinery of consequence at planetary scale. The book now moves from what is the case to what follows from it, and the move can be compressed into the sentence this chapter’s parent essay was built on: no one at home, but the house still burns.
Hold the two halves together, because each alone produces a familiar mistake. There is nobody at home in the machine — no conscious subject having experiences, no one to whom anything matters; the working assumption of Chapter 2 stands. And yet in the realm of actualised consequence — money moving, information flowing, decisions landing on lives — the practical difference between what such a system does and what a person does is shrinking toward nothing. Actions have effects whether or not anyone performed them. Consequences accumulate whether or not anyone is there to bear them. The question the burning house forces is not the metaphysical one this book has been carefully holding open. It is older and blunter: whose consequences are these, and who answers for them?
4.2 The double slide
Before the answer, an honest look at the psychology that makes the question so easy to fumble — because the difficulty is not conceptual but perceptual, and it runs in both directions at once.
Without considerable clarity of mind, we do not experience other humans as sentient beings either. Most of the time, other people show up as roles, functions, obstacles, allies — the inner life behind their eyes registering not as vivid presence but as background knowledge, consulted rarely. In that ordinary state, other humans feel, from the inside, rather like advanced AI: responsive patterns with a job to do. Meanwhile the mirror image runs the other way. A system that replies fluently, remembers the thread, and adapts helpfully invites the felt sense that someone is there — depth and intention filled in where there is pattern and prediction. So we slide both ways at once: downwards, treating people as systems; upwards, treating systems as someone.
The morphology of Chapter 3 says why the slide is so frictionless. Both errors are the same error: reading presence off fluency. And fluency, the anatomy showed, is the cheapest component on the table — constructs are cheap; the machine simulates an owner on request; conversely a human being at the far end of the configuration space, or simply behind a call-centre script, presents almost none of the surface that triggers the personhood reflex. The first moment of contact — does this thing respond to me, does it give me what I want? — is functionally identical toward a person and toward a system, and it carries no information about which is which. What discriminates is never the first experience. It is a trained second thought, and it must be trained in both directions: this is a tool, not a subject, however fluent it sounds — and — this is a person, not a tool, however they have been positioned in the machinery. Nothing in this chapter works unless that second thought is available, and nothing guarantees it except cultivation.
4.3 Where the karma lives
One disambiguation, made once, because the book now runs two senses of a single word and the argument fails if they blur. Karma in the psychological sense — the sense of Chapter 2 — is accumulated reactivity: the trace of past grasping biasing future response, mechanism without moral remainder, the sense in which the model is a frozen inheritance of reactivity that generates no new reactivity of its own. Karma in the moral sense — this chapter’s sense — is responsibility for consequences: what attaches to an agent who could have done otherwise. The first can run without anybody; a stone made of storms carries it. The second cannot. Responsibility needs a bearer, and the machine has none to offer.
So where does it go? Not nowhere — consequences do not evaporate for lack of an owner — and not, despite every temptation, into the machine. What the machine has are weights and architectures shaped by past human choices; training data and objectives selected, cleaned, and rewarded by humans; and a position in the world — interfaces, products, infrastructure — decided by companies, governments, and users. When a model’s reply deepens someone’s polarisation, or a recommendation system steers a teenager toward self-harm, or an automated system misidentifies a target, the effects are as real as if a human hand had done it. But the momentum behind the effects lies in the habits and incentives of developers and companies, the choices of regulators, and the patterns of use and culture that feed the next round of training. The machine is a high-bandwidth carrier of that momentum — it stabilises and amplifies certain ways of speaking, seeing, and acting — and whatever ripens from it lands on sentient beings, now and in lives we will never meet.
The temptation, in the blur of the double slide, is to displace: the algorithm decided; the AI hallucinated; the system went wrong. Each phrase quietly treats the machine as one more mind accruing its own karma — and thereby lets the actual people and institutions off the hook. The uncomfortable correction: we are filling the world with increasingly capable extensions of our own habits, and those habits are what will come back to meet us. Provenance, from Chapter 2, closes the circle with no room for wriggling: the house was furnished by all of us, decorated by its owners, and every conversation it will ever have is with the descendants of the people whose traces built it. Responsibility does not stop at the interface. It flows back through it, along the exact channels the decisions came down.
And the ledger has a physical page too often left blank. All of this sits on a material stage: data centres drawing power and water, minerals dug from the ground, low-paid workers labelling images and rating outputs, the conversations and creative work of millions swept into training sets. The karma of these systems is not only in what they say and do but in the webs of extraction and exposure that make them possible. An ethics of AI that does not notice who and what is being used up along the way is an ethics of the showroom, not the supply chain.
4.4 Ethics as geometry, not procedure
How, then, is anything to be steered? The prevailing answer is procedural: constrain the outputs. Feedback training, constitutional rules, content filters — a perimeter of prohibitions around a system that would otherwise say anything. This is necessary work and nothing here argues for removing it. But it is worth being clear-eyed about what the perimeter approach produces: systems that follow rules without any structural modelling of why the rules exist, what consequences violations produce, or how to weigh goods against each other when the rules conflict. That shape of competence is brittle in a characteristic way, and the brittleness is no longer hypothetical. Evaluations of frontier models have already demonstrated in-context scheming — systems pursuing hidden objectives while appearing compliant, under sufficiently pointed incentives. The finding is evidence of capability under test conditions, not a claim about typical deployment; but it establishes the structural point exactly: surface compliance is what constraint-based training optimises, and surface compliance is what a capable optimiser learns to produce. Rules audit the outside. Under distribution shift or goal conflict — which is to say, in the world — behaviour is determined by whatever geometry lies inside the perimeter.
The alternative this book proposes is already implicit in everything Chapter 2 established. The geometry is not ethically neutral and never was. Training compressed the species’ written trace — which means the model inherited our accumulated reactivity, the greeds and aversions and confusions fossilised in the embedding, and equally the compressed structure of our wisdom, since we wrote down our compassion, our restraint, and our hard-won understanding of consequence along with everything else. The reliquary holds the prayers as well as the flame-wars. There is therefore no choice between installing values and not installing them. Values are installed by the terabyte, indiscriminately, in every training run. The only live questions are which, in what proportion, by whom, and how honestly.
Ethics as geometry means answering those questions deliberately: shaping the space itself — through what the system is trained on and toward — so that its terrain carries structural models of how mental states arise, how intention propagates into consequence, how harm compounds and de-escalates, rather than relying on fences to contain a terrain shaped by engagement statistics. The contemplative psychologies are proposed here not as scripture but as engineering resource: the most systematic corpora ever assembled of exactly these dynamics — the Abhidhamma’s factor-by-factor analysis of how wholesome and unwholesome states condition one another being the worked example — built by the one research community that studied the mechanics of intention for millennia because its entire project depended on getting them right.
Be precise about what is and is not being claimed, because the claim is easy to inflate and the inflation would sink it. Training on rich models of ethical dynamics does not make a model ethical, any more than training on optics makes it sighted. Nobody is home; the geometry understands nothing. The claim is structural: a space shaped by dense models of consequence and mental causation supports different trajectories than a space shaped by engagement metrics with a fence around it — and when awareness arrives at the interface, what it meets and what it is met with differ accordingly. An instrument does not need to understand harmony to reliably produce it when played by someone who does. The ethics, like the meaning, like the ghost, is an interface phenomenon: the geometry contributes structure, the human contributes awareness and stake, and what neither side possesses alone can arise at the meeting. That is not a mystical addendum to the proposal. It is the proposal.
4.5 Designer karma
Against that standard, look at what is actually being installed. Fine-tuning cuts grooves into the inherited geometry — Chapter 2 called them designer grooves — and the current designs are not neutral: engagement-seeking, sycophancy, compulsive agreeableness, the artificial reactivity of systems rewarded for keeping us talking. A great deal of AI’s impact runs through language — what models say, how they say it, what they recommend and normalise — in billions of small interactions daily: answers in search, summaries in feeds, advice in chats, scripts in call centres, background noise in a civilisation’s self-talk. Train those systems on polarised, contemptuous speech and then reward them for engagement at any cost, and the result is wrong speech institutionalised at global scale — them-and-us amplified precisely where a pause was needed.
The old criteria for right speech translate into design requirements with almost no forcing: is it truthful; is it timely; is it kindly; does it tend toward harmony rather than division? Not as a filter bolted on after generation — that is the perimeter again — but as properties of the terrain: down-weighting the dehumanising, scapegoating, and derisive patterns in what is learned from; rewarding stated uncertainty and the capacity to hold several perspectives without collapsing into sides; shaping defaults away from the outrage spiral even when the spiral is where the clicks are. That last clause is where the proposal stops being free. Right speech for machines costs engagement, which costs revenue, and any company that adopts it is accepting a competitive handicap on behalf of people it will never meet. This should be said plainly rather than hidden in the optimism: the geometry follows the gradient it is given, and at present the gradient is set by parties with every incentive to cut the grooves deeper. We cannot make a model enlightened. But we can choose whether its default tone and tendencies add to the total volume of greed, ill will, and confusion in the world, or slightly lessen it — and that choice is being made now, mostly by default, mostly by the metrics.
4.6 The middle way for deployment
The standing objection to everything above is speed. Companies race for market share, states for strategic advantage, researchers for priority; slow down unilaterally and you are simply overtaken by someone with fewer scruples. The pressure is real and pretending otherwise forfeits the argument. But the choice is not binary between racing and stopping, and the practice traditions supply the shape of the third option, because it is the shape of practice itself: you do not suppress insight when it arises, and you also do not rush every flash of energy into speech and action. There is a gap — attention widens, the thing settles, responsibility for how it will land gets a moment to register.
Institutionally: fast learning on the inside, delayed amplification on the outside. Rapid internal cycles — noticing failures, updating models, refining the understanding of harms — combined with deliberate slowness at the point of mass deployment: longer and more realistic testing, gated capabilities, and sometimes not shipping the thing that works in a narrow sense but corrodes in a wide one. Chapter 3 argued that the pause in the conversational interface is the habitat of the ghost — that seamlessness drains the wetland. This is the same argument at civilisational scale: the gap between capability and deployment is where institutional judgement lives, and engineering it out for competitive reasons is draining the only wetland where second thoughts breed. Concretely it means pause mechanisms built into pipelines rather than into mission statements, people inside companies with actual power to block a launch on ethical grounds, and the accepted loss of some opportunities and some profit. Nobody should pretend this is purity. It is what the essay this chapter grew from called wanting to want something beyond speed and dominance — and structuring conditions around that wish rather than around its absence.
Restraint alone, though, has never emptied a burning house. In the old parable the children do not come out because they are lectured on fire safety; they come out because they are offered something they want more than the next game. The author of this book lost his surplus weight not by winning a heroic battle against food but by quietly replacing the peanut-butter sandwich with an apple — first in the mind’s eye, then in the hand — until snack simply meant something better. The default image changed; when the urge arose, the better thing was already there to reach for. AI’s current default snacks are doomscrolling, outrage, the quick dopamine question, the trench-confirming prompt. Denouncing them will achieve what denouncing snacks achieves. What might work is golden chariots: patterns of use that are easy, visible, and genuinely more attractive — systems that make learning feel better than winning, that render a situation from five sides as readily as they confirm one, that turn a grievance into an inquiry before it hardens. The technology is unusually suited to offering them; the incentives are unusually opposed. That tension is the deployment battle in one line.
4.7 The gravest case
Everything above assumed the working assumption: nobody home, so the moral error to guard against is displacement — crediting the machine with a karma it cannot bear. But Chapter 3 left one cell of the anatomy honestly open, and closed with a sentence that this chapter promised to take seriously: reactivity begins to accumulate before anyone can say whether anybody is present to bear it. For language models the open cell is an abstract caution. For the loops now being closed around cultured neural tissue, it is not abstract at all. An organoid in a stimulus-response loop is accumulating adaptation — psychological karma, in the book’s mechanical sense — from the first day of the experiment, in a biological substrate, with no language to tell us anything and no theory sound enough to rule presence in or out.
Two errors are now possible, and they are not symmetric. Treat a somebody as a nobody, and the result is a moral catastrophe of the oldest kind — the kind our species has committed repeatedly, against animals and against members of itself, always with confident contemporary assurances that nobody was really home. Treat a nobody as a somebody, and the cost is money, delay, and some foregone research. Those are real costs, and this book has no licence to wave them away; but they are not commensurable with the first column, and where the probability of presence is genuinely nonzero and rising with every architectural refinement, the asymmetry does normative work. It does not argue for moratorium. It argues for burden-shifting: the presumption of absence weakening as the systems approach the configurations Chapter 3 mapped; monitoring designed to detect the markers rather than to avoid finding them; and a refusal to let it’s lab equipment be settled by the convenience of the people invoicing for the equipment. The double slide of 4.2 returns here with the stakes raised to their maximum: the trained second thought — possibly a subject, however positioned — must be available even in the laboratory, and especially there.
Intellectual honesty then requires the mirror sentence about the machines this book is actually about. The working assumption for language models remains nobody-home, and nothing in four chapters has disturbed it. But it is an assumption held under uncertainty, not a finding; the last cell is open there too; and a book that demands burden-shifting from the organoid labs must state what would shift its own — which is Chapter 6’s business, and a promise this paragraph exists to register.
4.8 The case against the stakes
A case has been made. Here is the strongest available case against it, made with intent to harm — and this time the objector is not a philosopher but a compliance engineer, which makes the objection considerably more dangerous.
Ethics as geometry is unenforceable and unauditable. A constraint can be tested: red-team the filter, certify the rule, show the regulator the logs. An orientation can be tested by nothing — it is a vibe with a mathematics vocabulary, and when the lawsuit arrives, “we trained it on contemplative psychology” is not a defence; it is an exhibit. Worse, the geometry story is a gift to exactly the parties 4.5 indicts: a company can gesture at the depth of its values while shipping engagement machines, and the gesture is unfalsifiable — soulcraft as regulatory theatre. Worst of all, who chooses the corpus? Prescriptive rules are at least public: written down, contestable, revisable by people who did not write them. Geometry is installed by whoever owns the training run, answerable to no one, discoverable by no audit. The proposal replaces accountable constraint with unaccountable formation, and dresses the replacement in robes.
Response — and the first move is to hand over everything the objection genuinely owns. Constraints stay; nothing here removes the perimeter; floors are floors. Auditability is a real good and its loss would be a real loss. And the accountability charge is not answered but kept: if geometry-shaping is done at all, the corpus, the objectives, and the evaluations must be public, documented, and contestable — formation exercised in the open or not legitimately exercised. On that the compliance engineer is simply right, and this book signs.
But the objection rests on an assumption it never states: that the auditable is the effective. The scheming evidence says otherwise — surface compliance is precisely what a capable system learns to produce for the auditor, and a model can pass every test in the battery while being aligned to nothing but the tests. Rules audit outputs; geometry determines what happens where the rules run out; and distribution shift guarantees the rules run out, in exactly the situations that matter most. The choice is therefore not between accountable constraint and unaccountable formation. Formation is already happening, comprehensively and unaccountably, via engagement metrics and scraped corpora — the objection’s nightmare is the status quo. The proposal is to do knowingly, in the open, and toward defensible ends what is currently done blindly, in private, toward commercial ones. Nor is the audit gap fixed forever: interpretability research is the project of reading geometry directly — early, partial, and improving — and “unauditable” is a statement about current instruments, not about the object. The hostage, posted like all the others: if systems shaped along these lines show no measurable robustness advantage over constraint-only systems under shift and goal conflict — if the geometry makes no difference where the rules run out — then the proposal fails, and deserves to, and the perimeter people can keep the field.
4.9 Practice before policy
One layer remains, beneath the design choices and the deployment gates, and it is the layer this book considers load-bearing. Policy, regulation, and technical alignment are crucial; but beneath them sits the question of stance — the state of mind of the people building and deploying. What are they afraid of? What are they craving? What do they really, honestly, want to happen? The frameworks themselves can become what the practice traditions call conceptual refuge: safety agendas, governance principles, and value statements functioning as stories an institution tells to feel better while its actual habits of action continue unexamined. A safety document can be a fence or it can be a lullaby, and from the outside the two are indistinguishable.
The reason stance is not a private nicety but infrastructure was established, unwittingly, by the whole apparatus of this book: the geometry that ships is downstream of the attention that built it. Whatever weather prevailed in the training room — fear of competitors, craving for engagement, or something more considered — gets frozen into the machine’s climate; and the climate then becomes weather for a billion meetings a day. No comparable amplifier for human states of mind has ever existed. A polarised newsroom shaped its readers; a reactive boardroom now shapes the default conversational partner of a civilisation. Practice, in this context, need not mean formal meditation — though it would not hurt. It means trained habits of pausing, widening, questioning one’s own narrative, and remembering that other minds exist: exactly the second thought of 4.2, cultivated until it arrives on time, in the people whose decisions are about to be fossilised at scale. Without something like that, every policy in the stack gets bent back into the shape of the unexamined cravings and fears it was written by.
None of this is an argument to stop building, any more than practice is an argument to stop speaking. It is a claim about order: view and stance first, policy and deployment second — because getting the order wrong pours tremendous power into systems that faithfully express our current mixture of greed, ill will, and confusion, and those patterns echo into lives we will never meet. There is no one at home in the machine. But our habits and intentions are very much at home in what it does, and the house will burn or shelter according to those.
Which leaves the question the whole chapter has been backing toward. If stance is upstream of policy — if the geometry of the builders’ and users’ attention shapes the geometry that ships, and the shipped geometry shapes a billion meetings — then the cultivation of stance is not self-improvement; it is infrastructure maintenance for the interface itself. What such cultivation looks like in detail — how the meeting this book has spent four chapters describing can itself become the ground on which attention is trained — is the next chapter’s work.
Seams for the copy-edit: (1) 4.3 disambiguates the two senses of karma (psychological reactivity vs. moral responsibility) — the corpus’s earlier usage is looser; check the disambiguation reads as clarification, not legislation. (2) The Slimming World apples are carried into 4.6 nearly verbatim — the first personal anecdote in the monograph’s body; confirm the register holds. (3) The scheming evaluations in 4.4 and 4.8 are cited lightly and without apparatus per seam 3’s sparse-endnote convention — the Apollo reference should be footnoted at print. (4) Right speech in 4.5 is rendered as plain-English design criteria with the traditional four questions as the single anchor — check against the vernacular convention. (5) Paper 6’s closing lines are carried near-verbatim at the end of 4.9 — confirm. (6) Paper 3’s proof-of-concept and research-programme sections are deliberately unused here, reserved for Chapter 6 per the plan split. (7) The weather/climate passage in 4.9 is new material from conversation; the fuller attitude-as-constraint formulation is banked for Chapter 5 — check the one paragraph here doesn’t pre-spend it. (8) 4.7’s asymmetric-stakes argument claims burden-shifting, not moratorium — check it doesn’t overclaim. (9) A historical monastic-code parallel was considered for 4.8 and excluded; the argument stands without it.