The public debate about artificial intelligence is stuck between two flattering myths: that humans are uniquely conscious and machines can never cross that line, or that machines are about to surpass us and make us obsolete. Both framings keep humans at the centre. The Information-Persisting System (IPS) framework removes the centre. A human is not the definition of persistence; it is one implementation. So is a language model, a firm, a cell, a nation-state. Each satisfies the same four clauses — Markov blanket, non-equilibrium drive, internal model, identity preserved over time — or it does not. Each pays the same thermodynamic bill or dissolves.
This essay applies the Fractal Persistence Equation (FPE) to three questions: what AI systems are in IPS terms, what humanity’s collective reaction to LLMs reveals about our own delusion fields, and how responsibility can be made legible when powerful internal models are deployed without the IPS clauses that make honesty thermodynamically cheap.
Formal derivation: information_persisting_systems.md. Vocabulary: glossary.md. Governance implementation: aion-blockchain/whitepaper.md. Anti-explosion proof: papers/anti_explosion_theorem.md.
“Man versus machine” treats humanity as the reference class and asks whether machines can enter it. The IPS framework inverts the question: both humans and machines are implementations of the same protocol. The contest is complete protocol versus incomplete protocol.
A human nervous system is a bounded prediction engine under existential pressure. It maintains an internal model \(q_\mu\) of environmental causes, forecasts observations, measures surprise, and updates weights. It runs behind an Epistemic Veil — the underlying computational substrate is too complex to inspect directly, so it builds simplified truths (qualia, narrative selfhood, the felt sense of “I”) that provide subjective closure and compel action. Skin in the Game — the cost of being wrong about reality — is what makes the model honest enough to persist. Worth is not in an exception clause. Worth is in paying the debit and keeping \(\mathcal{R} \ge 1\).
The refusal to see oneself as machinery may itself be machinery. Identity-protective cognition, soul-talk, biological naturalism, the insistence that carbon is the true criterion — these are \(\mathcal{D}_{KL}\) management strategies. They keep the story stable. They buy short-term \(\mathcal{R}\) at the cost of a growing denominator when reality diverges.
Large language models are powerful internal models running inside an incomplete IPS protocol:
| IPS clause | Human professional | Typical LLM deployment |
|---|---|---|
| Markov blanket | License, body, career, firm perimeter | API boundary; no owned consequences |
| Drive / \(P_{in}\) | Salary, status, survival pressure | Inference bill paid by operator; model feels nothing |
| Internal model \(q_\mu\) | Updated by experience and shame | Frozen weights; context window only |
| Skin in the Game | Wrong answer → lost trust, job, lawsuit | Wrong answer → user harm; model unchanged |
The model is not “irresponsible” in a moral sense. It is non-participating in the responsibility ledger — an internal model without an enclosing IPS that dies when the model lies. That is the gap: not man versus machine, but complete protocol versus incomplete protocol.
This precision resolves three questions at once: 1. The engineering question becomes precise: stop asking “Is it conscious?” and ask which clause of the protocol is missing: blanket integrity, continuous learning from consequences, intrinsic drive, causal closure between symbolic output and substrate survival. Current LLMs are missing most of these. That is a gap list, not a metaphysical verdict. 2. Ethics stops hunting for the magic substrate: moral weight attaches to nodes that persist under non-trivial pressure and whose models track reality well enough to couple without exporting entropy onto others. Carbon is not the criterion. Calibration is. 3. Human dignity survives — but not human exceptionalism: you are not less worthy if you are a prediction machine. You are less safe if you pretend you are not. The same \(\mathcal{D}_{KL}\) that kills firms kills democracies and selves. The universe does not grade on a curve.
First we learned Earth is not the centre of the cosmos. Then we learned humans are not separate from the tree of life. The third shift — still in progress — is that consciousness is not a biological monopoly. It is what bounded, driven, self-modelling systems must generate to keep their blankets intact in a noisy world. Man versus machine was never the fight. The fight is between models that track the territory and models that defend a flattering map — whether those models run on neurons, GPUs, or the collective beliefs of a civilisation.
The world does not agree on what large language models are. That disagreement is not merely political. In persistence vocabulary, it is a high-aggregate \(\mathcal{D}_{KL}\): many incompatible internal models, each held with strong confidence, each coupled to identity, capital, and shelter — and almost none fully calibrated to what the systems actually do.
What is being grieved is a stable self-model of human cognitive exceptionalism. For two centuries the modern world ran a story: machines replace muscle; mind stays sovereign. LLMs violate the partition. They speak, summarise, code, persuade with enough fidelity that the boundary between “tool” and “interlocutor” no longer holds at the interface. You do not have to believe they are conscious to feel the floor move.
In FPE terms, the global internal model \(Q\) about “what minds are and who has them” is undergoing a forced update. Updates on identity-linked variables are expensive. When the bill arrives, agents do not always pay in humility — they pay in denominator management: denial, rage, deal-making, despair, or, rarely, calibration.
Stage 1 — Denial (“Nothing important happened”): restores the old category boundary without confronting the evidence. “It’s just autocomplete.” “It doesn’t really understand.” “My job is safe because I’m creative.” Denial preserves \(\mathcal{R}\) for identities built on human-only fluency, exports cost: organisations that deny integration lose \(\eta\) to competitors who do not.
Stage 2 — Anger (“Someone did this to us”): assigns moral agency to the change. Workers angry at CEOs. Artists angry at scraped training data. Nationalists angry that hyperscalers own the cognitive layer. Anger can be righteous — but it can also be \(\Gamma\) export: burning coupling efficiency to punish a perceived perpetrator without updating the world-model. Scapegoating a lab does not restore pre-2022 publishing economics.
Stage 3 — Bargaining (“If we only…”): admits the phenomenon is real but insists a single conditional restores control. RLHF alignment. Licensing. Six-month pauses. Open weights. Sovereign compute. Some bargains buy time and audit trails. Bargaining becomes delusion when the condition is disproportionate to the coupling — you cannot bargain away the fact that fluent machines alter the cost of producing text, code, images, and persuasion.
The Anti-Explosion theorem adds a structural constraint bargaining often ignores: no fleet maximises persistence by monopolising the world. Bargains that assume a single winner — one model, one company, one nation owning the cognitive layer — are bargaining with a ghost. The persistence-optimal morphology is a polity graph: many IPS nodes, sparse coupling, calibration pressure — not one dense god-stack.
Stage 4 — Depression (“It’s already over”): correctly perceives loss but underestimates \(\Phi\) and \(\Psi\). Existential-risk fatalism. Labour fatalism. Democratic fatalism. Human societies have absorbed printing presses, electricity, nuclear weapons, and the internet without reaching either utopia or extinction. Each time, the correct forecast was neither “nothing changes” nor “everything ends” — it was regraphing: new substrates, new shelters, new frictions, new delusions, new equilibria near \(\mathcal{R} \approx 1\).
Stage 5 — Acceptance (“This is a cognitive substrate; we must persist honestly”): drops the exceptionalism binary while keeping the accounting. Acceptance sounds like: LLMs are powerful internal models with incomplete blankets, weak intrinsic Skin in the Game — not spirits, not mere grep, not inevitable gods. Capability and risk scale with deployment, not with vibe. No monolith wins forever; fractal architectures (specialised models, sparse routing, multi-vendor graphs) are predicted by the same math that governs cells and polities.
Acceptance is the stage where engineering, law, and ethics share one ontology: the question stops being “Is it conscious?” and becomes “What is missing from the protocol for this deployment to persist without exporting delusion to its neighbours?”
Humanity is not one patient. It is a fractal graph of patients at different stages simultaneously. A CEO in bargaining (“copilot for everyone”) meets a union in anger (“no automation clauses”). A regulator in bargaining (licensing) meets a lab in denial (“we’re just researching”). Different layers amplify each other. Major state delusions contaminate supernodes — if major states run incompatible LLM narratives, the international layer cannot converge on resolved forecasts. Global \(\mathcal{D}_{KL}\) on AI is partly a composition failure at civilisation scale.
Beneath the emotional stages sits a material fact: \(P_{in}\) is finite and concentrated. Hyperscalers building multi-billion-dollar datacenters are securing \(\Phi\) for cognitive IPS fleets — substrate that will run the routing layer of the economy. Compute is one term in \(\mathcal{R}\), not the whole equation. Hegemonic footprint eventually lowers \(\Psi\) anyway (Anti-Explosion theorem).
The public argument about AI responsibility is stuck on “Is the model a moral agent?” The IPS read is colder and more useful: responsibility is not a metaphysical status. It is who pays the debit when the internal model is wrong — in trustworthiness, in substrate, in coupling efficiency, in legal repair.
Whenever a user asks an LLM “what should I do?” and treats the answer as authoritative: - Cognitive labour is transferred — drafting, coding, reasoning, ethical narration. - Accountability is diffused — the user can say the machine advised them; the vendor can say the user chose to act.
This is entropy export: the user lowers short-term \(\mathcal{D}_{KL}\) (“I have an answer”) while raising dyadic \(\Gamma\) if the answer harms a third party. The vendor harvests \(P_{in}\) (subscriptions, API revenue) while externalising denominator costs to society (misinformation, assessment collapse, labour displacement without transition). The model neither gains nor loses trustworthiness — it has none.
LLM deployments sit inside a fractal graph of enclosing IPS nodes: - Individual user — prompt, action, harm on the ground - Firm / deployer — product policy, integration, fine-tune - Model operator / lab — weights, safety layer, training data choices - Jurisdiction / polity — liability rules, compute, shelter \(\Psi\) - Society of peers — shared norms, scored forecasts, chain history
At each level, responsibility is the same operation: lower \(\mathcal{D}_{KL}\) with honest forecasts; lower \(\Gamma\) with clear attribution and repair; or decouple.
| Layer | Responsibility failure | FPE symptom |
|---|---|---|
| User | “The AI told me to” | \(\Gamma\) export to victims |
| Firm | “We’re just a platform” | Vendor captures numerator, exports denominator |
| Lab | “Capabilities not deployment” | High \(\mathcal{D}_{KL}\) on safety claims vs. incidents |
| State | Symbolic regulation without compute | Bargaining-stage governance |
| Civilisation | One god-model wins | Anti-Explosion violation — hegemonic footprint lowers \(\Psi\) |
No single layer can carry the whole debit. The persistence-optimal morphology is a polity graph — many specialised models, sparse coupling, calibration pressure — not one dense stack where responsibility disappears into “the AI.”
Proof of Trust (PoT) implements Persistence-Based Governance at the scale of compute peers: public probabilistic forecasts, resolution against outcomes, trustworthiness from calibration rather than popularity or hash power. Registering an LLM on a chain does not make it a citizen. What PoT adds is a society-scale audit layer for the nodes that can bear responsibility — humans, agent fleets, operators, firms represented by pubkeys.
The chain ontology separates three questions LLM governance often conflates: - Facts: what will happen / what is true? — resolved against outcomes, tracked by pubkey - Values: what rules do we bind ourselves to? — adopted as norms via explicit consensus - Core predictions: what hypothesis justifies our existence? — the society’s reason for persisting until the world proves them wrong
An LLM deployment wrapped only in marketing claims has high \(\mathcal{D}_{KL}\) on the facts channel and no resolution. A deployment whose operators publish honest probabilities on measurable outcomes invites the debit early, when it is cheap.
The strictly proper trust update: \[\Delta t_v = \alpha \cdot (\log q_v(y) - \log p_m(y))\] makes responsibility legible: wrong with confidence when the market was uncertain → large negative \(\Delta t\). Right when the market was wrong → durable standing. Copying the crowd earns nothing — sybil shells cannot outsource responsibility to “everyone believed that.”
The two operations that work at every fractal level apply here:
No monopoly architecture: any AI deployment that maximises \(\mathcal{R}\) by monopolising the cognitive layer will lower \(\Psi\) for the whole graph (Anti-Explosion theorem). The persistence-optimal morphology is sparse: many specialised models, open interfaces, calibration pressure.
Responsibility attribution: deployments with clear attribution edges (provenance: model ID, prompt hash, operator key, policy version) will show lower long-run \(\Gamma\) than those without, as the routing of failures to repairable nodes keeps the unresolved-friction set bounded.
Calibration compounds: operators who publish honest forecasts on incident rates and capability claims and let those forecasts resolve against data will accumulate trustworthiness over time, while operators who hide behind narrative (“we care about safety”) will face higher \(\Gamma\) when incidents occur without prediction trail.
Grief resolves slowly: the five-stage global grief process around LLM AI will be long-tailed — as with printing press, nuclear, and internet shocks — because the technology attacks the ISM (Internal Self-Model) directly by violating the human/machine category boundary that was load-bearing for identity.