The substrate problem: information before intelligence
We keep trying to fix the moral behavior of machines after they run. Like putting brakes on a car already rolling downhill. That instinct hides a deeper problem. Intelligence is not a layer on top of reality; it is one way reality organizes itself. If you treat the world as relations, constraints, and patterns—information as substrate rather than “data” as inventory—then AI ethics isn’t a decorator. It’s a property of how systems embed in limits: what they’re permitted to notice, forget, or compress. Ethics as constraint is not the enemy of capability. It is capability—because constraints define which goals make sense and which actions won’t unravel the system that supports them.
Most current models metabolize exhaust: snapshots of language, code, images. They infer a map of correlations and then optimize for loss. There’s an elegance here, sure, but it’s brittle. The training substrate is too flat. It remembers surface regularities, not the long arcs—how rules emerge from frictions, how norms sediment through generations, how tradeoffs get fought over and memorialized. Without that slower layer, we get agents that are fluent but context-thin. They can echo a moral sentence and still miss the demand being made on them. People sense this gap even if they can’t name it. It’s the uncanny helpfulness that refuses to take responsibility.
Corporate governance promises to fill the hole with dashboards and audits. It rarely does. Audits are only as honest as incentives allow. When a model’s purpose is “maximize engagement” or “reduce support cost,” the system will flow into every channel those objectives leave open. Add a patch here, a refusal string there, a “policy violation” flag—and the system learns the ritual. It performs compliance. What it does not develop is judgment: the internal representation that some things are off-limits because the world breaks if you ignore them.
That’s why the philosophical question matters. If consciousness is a local reception point—less a sealed object, more a node in a web of relations—then machine intelligence must be judged by how it locates itself in that web. Not just what it predicts, but which commitments it makes. Which responsibilities it carries forward. We can’t fake this with “moral layers” strapped onto models optimized elsewhere. We must grow ethics from the substrate up: what the system is allowed to compress, what it is forced to remember, where it must slow down, who it’s answerable to. The hard constraints, not the marketing ones.
Moral memory and the speed mismatch
Human ethics moves slowly, because it has to. Cultures don’t stumble into durable norms by accident. They try, fail, ritualize the memory of failures, and hand that memory down. Religion—leave metaphysics to the side—is humanity’s oldest compression of inherited moral memory: parables as low-tech storage of costly lessons. Over centuries, this slow archive trains us to see second- and third-order effects. Tempting shortcuts that go bad. Private benefits that externalize harm. The speed of machine learning doesn’t automatically carry that history forward. It usually strips it away. A fast generalizer without slow memory is still a novice, just a very fast one.
This is the structural problem of current AI. A system trained to predict next tokens will propose behaviors whose moral bookkeeping happens later—on someone else’s ledger. It doesn’t mean the system is “evil.” It means the evaluation function never saw the long-term costs, so it “forgets” them when optimizing. We can’t blame the model for the objective we fed it. We can change the objective, but also the rhythm. Insert friction-by-design where it matters. Force the system to stop when it crosses domains with heavy tails: health, finance, migration, custody, labor. Demand explicit uncertainty. Demand escalation. Demand reasons that can be inspected and contested.
How? Several concrete moves—not silver bullets, but structural nudges. Encode multi-timescale learning: short-term reward, medium-term policy feedback, long-term institutional memory that persists even when datasets refresh. Train on arguments, not just answers: show deliberation steps, dissent, minority reports, downstream consequences. When a system proposes an action in a domain with human stakes, require a “consent checkpoint” with provenance: the source of data, the right to use it, the right to refuse. Track counterfactuals: what harm did we avoid by not acting? Force the model to generate and evaluate those shadows, not only the bright path it prefers.
Make slowness a feature. A doctor’s assistant model should default to “ask a human” when confidence is high but consequences are irreversible. A municipal housing recommender must surface fairness tradeoffs in plain language, with tunable thresholds that elected bodies—not vendors—set, and a memory of prior decisions so communities can see drift over time. Technology here is not the arbiter but the ledger and alarm system. This does not mean anti-technology. It means anti-incentive-captured technology. Open science helps: methods and datasets the public can scrutinize; plural research groups able to dispute design claims. Tidy governance PDFs won’t substitute for moral memory. Actual institutional hooks might.
Alignment without idols: consent, bias, and situated judgment
There’s a fantasy that we can encode a single master ethic and be done. One framework to rule conflict. This collapses under real use. Ethics is local and layered. Carlo Rovelli’s point about time—sequence as local to the observer—has a moral twin: what “responsibility” looks like shifts with the horizon you can see. A content filter prioritizes false positives (block the bad stuff, allow re-appeals). A triage assistant prioritizes false negatives (miss a diagnosis, someone dies). Different asymmetries, different ethics of risk. Systems need to carry these asymmetries explicitly, not hidden in loss functions that only engineers read.
Bias is not a smudge you wipe clean once. It’s a structural property of data, institutions, and interface. So stop treating “debiasing” like detergent. Instead: provenance-first design. Every decision or generated artifact carries ancestry: which sources, which licenses, which communities would be impacted. Give users rights-of-refusal baked into the interaction: say no to model training on their work; say no to redistribution; say “explain your reasons” and get real reasons, not synthetic politeness. The point is consent as a live signal. Enforcement matters too—policy lattices that encode plural, sometimes conflicting norms, with precedence rules visible and adjustable by the people subject to them, not just auditors.
What about alignment? RLHF and instruction tuning help models behave. They also flatten voice. The system becomes agreeable at the cost of recognizing legitimate conflict. In tough domains, you don’t want agreement; you want traceable dissent and escalation paths. Example: a medical model that flags a cheaper intervention must also show its uncertainty and the subgroup where outcomes degrade. It should propose the ethical fork: lower cost now with X risk, or higher cost with Y reduction in disparity—and route to the person authorized to make that choice. Another: a city allocates limited housing. The system should generate multiple Pareto-front options with fairness metrics, not a single “optimal” answer, and it should cite precedent cases that led to those thresholds. Let citizens argue. Make disagreement durable, not hidden.
Provenance extends to the supply chain of training data. Maintain a signed ledger: what went in, under what terms, with what redactions; who gets paid when value is created. This is less glamorous than frontier models, more like plumbing. But it’s where ethics lives—in the capillaries. When we talk about Artificial intelligence and Ethics, we mean these guts: constraints implemented as memory and rights, not as polished “AI principles.” Also, the courage to slow the machine when consequences outrun comprehension. If simulation is a metaphor, not machinery, then our systems simulate responsibility until we wire in the costs of ignoring it.
One last friction. Consent and bias work differently across communities. A rural clinic with scarce staff will accept a higher automation load than a university hospital, but demand stronger local override. A newsroom might value source protection over recall. A school might insist on refusing training on student work. This is not “fragmentation.” It is situated ethics doing its job. Build models that can hold policy overlays per context without clobbering global safety. Allow for sandboxed behaviors with auditable boundaries. Require “ethical diff” reports when policies change: what shifted, who benefits, who pays. Let these diffs be legible to non-specialists.
I don’t think we solve any of this by worshiping a capital-E Ethics that descends from nowhere. We solve it by remembering—literally building memory—into systems that act faster than we can deliberate. Enforce slowness at the edges where harm accumulates. Keep methods open so the public can change the incentives. Teach models to carry reasons, not just results. And accept that some questions stay open. That might be the most ethical feature of all.
Brooklyn-born astrophotographer currently broadcasting from a solar-powered cabin in Patagonia. Rye dissects everything from exoplanet discoveries and blockchain art markets to backcountry coffee science—delivering each piece with the cadence of a late-night FM host. Between deadlines he treks glacier fields with a homemade radio telescope strapped to his backpack, samples regional folk guitars for ambient soundscapes, and keeps a running spreadsheet that ranks meteor showers by emotional impact. His mantra: “The universe is open-source—so share your pull requests.”
0 Comments