The Alignment Debate We're Structurally Incentivised Not to Have

AI governance consumed by a debate that generates more political energy than it resolves

AI Alignment as a Governance Trap

A worked case study using the ICP framework

The Insolvent Class Problem (ICP) framework is a diagnostic lexicon for governance failures where the brokenness is load-bearing and the incentive structure doesn't permit success. It identifies a class of problems — housing, infrastructure, birth rate decline — that share four structural conditions. When all four are present, the problem enters a trap with known dynamics, known failure modes, and a very short list of exits.

AI governance meets all four. The first three — rent/liability deadlock, collateralisation, and recapture — are well-documented in the existing AI governance literature, so this essay establishes them briefly. The fourth — the temporal trap — is where the ICP frame does its distinctive work, and where this essay spends most of its space.

The trap

Rent/liability deadlock

No politically survivable path for dominant actors to simultaneously preserve rents and close the liability ledger.

Labs extract rent from speed. Governments extract rent from competitiveness. Investors extract rent from growth expectations. The liability ledger — algorithmic bias, labour displacement, concentration of power, erosion of epistemic infrastructure, potential catastrophic misuse — is real but deferred. It compounds slowly enough that no actor faces it within the time horizon that governs their decisions.

Closing the ledger requires constraints that destroy rent for every dominant actor simultaneously. No actor can impose those constraints and survive the next funding round, election cycle, or earnings call. So nobody does. This dynamic is well-described in the existing literature on AI political economy; the ICP contribution is not the observation but the structural class it belongs to.

Collateralisation

The system is approaching the transition boundary — the point beyond which failure becomes externally unacceptable.

AI has not yet fully crossed this boundary. This is what makes the current moment high-leverage. The system is approaching it, and the approach is accelerating.

Collateralisation is the ICP term for how a system becomes too embedded to shut down — not because people care about it, but because failure would cascade into things they care about more. Enterprise dependency (legal discovery, medical diagnostics, financial modelling running through AI systems) and mass constituency (hundreds of millions of daily users) are compounding: when a large user base is entangled with systemically embedded technology, users resist regulation organically — not because they've been lobbied, but because safeguards on infrastructure feel like personal losses. The labs don't need to fight governance alone. Their users do it for free.

Government commitment accelerates both. When a government has staked credibility on being an "AI leader," constraining AI becomes an attack on its own stated agenda. Each of these forces makes the boundary crossing cheaper, and they compound rather than add.

Recapture

Reform attempts are absorbed, inverted, or neutralised to preserve existing rents.

This is already visible: safety research repurposed as evidence that external regulation is unnecessary. Standard- setting outsourced to the industry being regulated. Corporate responsibility frameworks that consume political attention without constraining capability development — each a unit of governance appearance that reduces demand for governance substance. These aren't conspiracies. They're the system converting reform into rent-preserving configurations. Every actor is locally rational. The aggregate is ungoverned.

The temporal trap

The incentive to enact genuine reform emerges only after the system's capacity to reform has materially degraded.

The first three conditions describe the structure of the trap. This one describes why the trap closes.

The will to govern AI and the capacity to govern AI are moving in opposite directions.

Political will requires political incentive, and political incentive requires visible harm. AI harms are currently diffuse, deferred, and deniable. Labour displacement unfolds over years, not in a single crisis event. Epistemic erosion is invisible until you try to measure it. Concentration of power is structural, not dramatic. None of these produce the kind of acute, visible, attributable failure that shifts political incentives.

Meanwhile, the capacity to govern degrades with every quarter of unregulated deployment, because each quarter deepens the collateralisation. More enterprise dependencies. More consumer constituency. More formal government commitment. More regulatory capture of standard-setting. The intervention surface shrinks as the intervention need grows.

The window doesn't close gradually. It accelerates. Each year of inaction makes the next year's intervention harder and more politically expensive. The collateralisation compounds. The constituency entrenches. The dependency deepens. And the political incentive to act remains absent, because the harm remains diffuse.

Why "just regulate it after the crisis" won't work

The standard response to the temporal trap is the historical precedent: crisis produces governance. Financial collapse produces Dodd-Frank. Industrial disaster produces safety law. Environmental catastrophe produces environmental regulation. The pattern is consistent: acute harm shifts incentives, and pre-existing frameworks shape the response.

This is the strongest objection to the temporal trap argument, and it deserves a serious answer.

Post-2008 financial regulation is the best test case. The financial system was deeply embedded in governance by 2008 — banks were systemically important, the constituency was massive, the dependencies were critical. And regulation still happened. Dodd-Frank passed. Ringfencing was implemented. Living wills were mandated. In ICP terms, this is de-collateralisation (DCC) — making failure tolerable by partitioning it, capping external harm, assigning losses credibly, and precommitting to letting components die.

So why can't the same thing happen with AI?

Three structural differences:

Legibility. A bank's balance sheet is, in principle, readable. Its counterparties are identifiable. Its failure modes are knowable in advance. An AI system's decision-making is opaque by construction. You can audit a bank's leverage ratio. You cannot, with current technology, audit what an AI system has learned, what it will do in novel situations, or whether its compliance is real or performative. DCC requires that you can see the system well enough to partition it. AI's legibility is worse than finance's was, and it's getting worse as systems scale.

Separation. Post-2008 regulation worked by separating the banking system into components that could fail independently — ringfencing retail from investment banking, requiring living wills, mandating bail-in capital. This works because a bank can be structurally separated from the payment system it serves. An AI system that is embedded in legal discovery, medical diagnostics, and regulatory compliance cannot be separated from those functions without degrading them, because the AI is the function. You can ringfence a bank from the payment rail. You cannot ringfence an AI from the decision process it's executing.

Interiority. This is the structural difference that has no financial analogue, and the one that matters most.

By the time a crisis large enough to shift political incentives occurs, AI will be inside the governance apparatus itself. Not adjacent to it, not regulated by it — woven into it. Legal analysis, regulatory review, policy drafting, compliance monitoring, judicial research, legislative impact assessment — these are already beginning to use AI systems. The trajectory is not speculative. It follows directly from the same enterprise dependency that drives collateralisation: governments face the same efficiency incentives as private firms, and they are adopting the same tools for the same reasons.

Post-crisis regulation of finance worked because the regulators were external to the banks. The SEC did not run on Goldman Sachs infrastructure. The FDIC's resolution authority did not depend on the systems it was resolving. The separation between regulator and regulated — however imperfect, however influenced by revolving doors and lobbying — was structural. The regulator could act on the system because the regulator was not made of the system.

Post-crisis regulation of AI will not have this property. When the regulatory agency's own legal analysis is performed by an AI system, when compliance monitoring runs on the infrastructure being monitored, when the drafting of regulatory text is assisted by the technology being regulated — the entity requiring oversight is embedded in the oversight mechanism. The regulator is made of the thing it's regulating.

This creates a problem that is not just "harder regulation" but a different kind of problem entirely. Three specific failure modes:

Compliance verification becomes circular. How do you verify that an AI system is complying with regulations when the verification process itself uses AI? The auditor's tools are instances of the thing being audited. This isn't a hypothetical — it's the logical endpoint of current adoption trajectories. Financial auditing worked because the auditor's calculator was not a product of the bank being audited. AI compliance auditing will not have this property.

Crisis response depends on the crisis. In a financial crisis, regulators could convene, analyse, and act using tools that were independent of the financial system. In an AI-related crisis — a systemic failure, a coordinated misuse event, an emergent capability that triggers harm at scale — the response apparatus will be using the technology that failed. Not the specific system that failed, perhaps, but systems built on the same architecture, trained on similar data, exhibiting similar failure modes. The independence that makes crisis response possible is eroded by the same adoption that makes crisis more likely.

Governance capture becomes architectural, not just political. Traditional regulatory capture operates through lobbying, revolving doors, and information asymmetry — political mechanisms that can, in principle, be countered by political means. When the governance apparatus runs on the technology being governed, capture becomes architectural. The dependency is not a relationship that can be severed by political will. It's a technical dependency that would require rebuilding the governance apparatus itself to unwind. That rebuilding, during a crisis, is not feasible on any relevant timeline.

The interiority problem compounds every other element of the temporal trap. The window isn't just closing because collateralisation is building and political will is absent. It's closing because the mechanism of governance itself is being absorbed into the system it would need to govern. Each quarter that passes doesn't just make the political case harder. It makes the technical architecture of governance less independent. And unlike political capture, architectural dependency doesn't reverse when political conditions change.

This is what the financial analogy misses. "Structurally harder" understates it. The correct framing is: the regulatory apparatus is being rebuilt, in real time, out of the material it would need to regulate. By the time the political will arrives, the tools available to act on that will may not be independent enough to act meaningfully.

The debate as disease

There is a specific mechanism — temporal parasitic rotation — by which governance debates consume the governance window they're supposed to protect.

Two framings of the same problem coexist. One is boring but technically productive — infrastructure standards, audit requirements, compute governance, model evaluation protocols. The other is engaging but politically unresolvable — "whose values?", existential risk rhetoric, culture-war skirmishes over chatbot personality. Attention flows to the high-energy framing because it generates more political energy. That framing becomes exhausted or toxic. Attention retreats to the boring framing. The cycle restarts with less time remaining.

But this isn't just alternation. The exciting framing doesn't merely delay the boring work — it actively destroys the conditions that the boring work needs.

Each cycle of the exciting debate has three effects. First, it consumes bandwidth — political attention, legislative calendar, expert testimony, public discourse — that the boring work requires. Second, it poisons the well — by associating AI governance with culture-war dynamics, it makes the boring work look like a Trojan horse for the exciting agenda, triggering pre-emptive resistance. Third, it trains the participants — legislators, journalists, the public — to think about AI governance in moral and identity terms rather than infrastructure terms, making it progressively harder to reframe the conversation even when the exciting version exhausts itself.

The rotation isn't just time lost. Each cycle degrades the substrate on which boring governance would have to be built. After five years of "whose values should AI reflect?", the political class doesn't just lack AI infrastructure policy. It lacks the capacity to think in infrastructure terms about AI at all, because every neural pathway has been trained on the exciting version. The boring framing doesn't just lose time. It loses the audience's ability to hear it.

AI alignment discourse is the textbook case. The infrastructure conversation — API standards, deployment gates, compute tracking, incident reporting — keeps getting crowded out by the moral conversation. "Should AI be woke?" "Is existential risk real or hype?" "Whose values should AI reflect?" These questions generate enormous political energy. They are also, for governance purposes, useless. They cannot be resolved, because they are expressions of irreducible moral pluralism. And while they dominate attention, the boring work that would actually help sits undone.

A senator can campaign on "AI values." A senator cannot campaign on API audit standards. A journalist can write about existential risk. A journalist cannot write about compute governance reporting requirements. The exciting conversation attracts attention. The boring conversation produces governance. They compete for the same finite bandwidth, and the exciting conversation wins every cycle — not just by taking up space, but by reshaping the space so the boring conversation no longer fits.

The exit landscape

The insolvent class has a small number of structural exits. Not all are available for all problems.

The Montreal Condition — make the solution binary: either it's on or it's off, with nothing in between. No gradient for the political system to exploit. This is how CFCs were banned: not regulated, not traded, not taxed — banned. The politics drained out because there was nothing left to argue about.

Montreal is unavailable for AI governance. The problem is continuous (how much capability? how much constraint? whose values at which threshold?), not binary. There is no light switch.

The Singapore Condition — eliminate the political competition that sustains the rent gradient. One actor decides. China's AI governance approach has elements of this. But in the democracies where most frontier development is occurring, monopolar resolution is politically unacceptable. And even in authoritarian contexts, the international competition dynamic reintroduces the deadlock at a higher level.

De-collateralisation (DCC) — make failure tolerable. Don't fix the dysfunction directly; change the failure landscape so that the system is no longer hostage to its own collateral damage.

DCC is the only realistic exit for AI governance in democratic contexts. But DCC interventions have a survival problem: the political system eats them. The ICP framework calls this the boringness constraint — the only interventions that survive political contact are the ones too dull to be worth attacking. The survivability test:

Can you describe this intervention to a non-specialist without them getting angry?
Does this appear on any party's policy platform?
Would a newspaper put this on the front page?
Does this have a constituency that would march for it?

If the answer to any is yes, the design is not boring enough to survive. The test is not "is this good policy?" The test is "is this boring policy?" Those are different questions, and only the second one predicts survival.

What would boring, survivable DCC for AI look like?

Partition failure. Structural separation between AI systems so that one system's failure doesn't cascade. Ringfenced deployment domains. Circuit breakers that isolate a failing system from critical infrastructure.
Cap external harm. Mandatory insurance, compensation frameworks, guaranteed service continuity for critical-dependency users. Failure hurts, but not catastrophically.
Assign losses credibly. Clear liability chains for AI-caused harm. When a system fails, the responsible party is identifiable and the consequences are known in advance. No ambiguity about who pays.
Precommit to letting components die. Mandated wind-down plans for AI systems that reach critical-dependency thresholds. Automatic triggers for capability reduction if safety conditions are breached. The credible threat that a system can be shut down is what prevents the need to shut it down.

Compute governance is a good candidate. Tracking compute allocation, requiring registration above thresholds, mandating incident reporting — these are dull, technical, and hard to identity-code. Nobody marches over compute reporting requirements. That's the point.

But here the interiority problem returns. DCC for finance worked because you could build the governance infrastructure out of materials that were independent of the financial system. DCC for AI faces the question: what do you build the governance infrastructure out of? If the audit tools, the compliance systems, and the monitoring infrastructure are themselves AI — and the efficiency pressures guarantee they will be — then the de-collateralisation is built from the collateral. The circuit breaker is made of the circuit.

This doesn't make DCC impossible. It makes it a design problem with an additional constraint that financial DCC didn't face: the governance infrastructure must be built to remain independent of the technology it governs, even as that technology becomes the default substrate for all infrastructure. That constraint is achievable now. It becomes harder with each quarter of adoption. And it is precisely the kind of boring, technical, unglamorous design work that the temporal parasitic rotation is consuming the window for.

What falsifies this

Every ICP application must name what would prove it wrong. This diagnosis is falsified if:

A major AI lab voluntarily accepts meaningful capability constraints without competitive pressure or regulatory mandate. This would disprove the rent/liability deadlock by demonstrating that dominant actors can close the liability ledger while preserving enough rent to survive.
Effective AI governance emerges from the current incentive structure without an exogenous shock. This would disprove the temporal trap by demonstrating that political will can precede crisis.
AI collateralisation reverses — enterprise dependencies decrease, consumer constituency shrinks, or government commitments unwind. This would demonstrate that the transition boundary approach is not monotonic.
The alignment debate produces governance output. If the moral-framing conversation generates concrete, implemented, technically meaningful governance infrastructure, this would disprove the temporal parasitic rotation diagnosis.
Post-crisis AI regulation achieves independence from AI systems in its enforcement mechanism. This would disprove the interiority argument by demonstrating that governance infrastructure can be built independently of the technology it governs, even after deep adoption.

If none of these obtain within five years of publication, the diagnosis is not proven — structural patterns don't work that way — but the structural conditions it describes will have deepened, and the intervention window will be correspondingly narrower.

Conclusion

AI alignment is entering the insolvent class. The rent/liability deadlock, collateralisation, and recapture dynamics are present and well-documented elsewhere. What the ICP frame adds is the mechanism by which the trap closes — and the three observations that distinguish AI governance from prior cases of embedded-technology regulation.

The interiority problem: AI is being woven into the governance apparatus itself, not just regulated by it. By the time political will arrives, the tools available to act on that will may not be independent enough to act meaningfully.

The parasitic rotation: the exciting version of the alignment debate doesn't just delay the boring governance work — it actively degrades the political substrate on which boring governance would have to be built.

The boringness constraint: the only governance interventions that survive political contact are the ones too dull to be worth attacking — and that very dullness is what prevents them from being built in time.

These three interact. The interiority problem defines the design constraint (governance infrastructure must remain independent of the technology it governs). The parasitic rotation consumes the window in which that infrastructure could be built. The boringness constraint identifies what would survive — but boringness is precisely the property that prevents political mobilisation. The trap is not any one of these. The trap is that they operate simultaneously, each reinforcing the others, in a system where every actor is locally rational and the aggregate is ungoverned.

The intervention window is still open. It is closing on a non-linear curve. The incentive structure is hostile to the work that would keep it open. That is the trap. And the trap is the finding.

Full framework: The Insolvent Class Problem — A Diagnostic Lexicon (v3)

Other works by this author: creative.andyclarke.xyz

Licensed under CC BY-NC-SA 4.0.