The Ethics of Not Knowing: When Uncertainty Becomes an Obligation

Uncertainty is the most convenient word in governance.

A model is uncertain, so the system waits. A committee is uncertain, so the decision is deferred. A risk officer is uncertain, so the memo gets another paragraph of decorative caution and nobody quite owns the next step. Very mature. Very responsible. Also, sometimes, very useful for avoiding responsibility while looking intellectually respectable.

Timothy Prescher’s paper, The Principle of Proportional Duty: A Knowledge-Duty Framework for Ethical Equilibrium in Human and Artificial Systems, tries to formalize a different answer.¹ Its core claim is simple enough to be dangerous: uncertainty does not erase duty. It changes the type of duty.

That is the useful part of the paper. Not because it has already built a deployable AI ethics controller. It has not. Not because a single equation can settle moral philosophy. Spare us. The value is that it gives decision systems a structure for separating two obligations that organizations often mix together: the duty to act, and the duty to repair uncertainty through verification, inquiry, escalation, or delay.

The paper calls this the Principle of Proportional Duty, or PPD. The article below reads it mechanism-first, because the mechanism is the point. The simulations, case studies, and appendices matter, but they matter because they show how the mechanism behaves under controlled assumptions. They do not prove that hospitals, regulators, banks, or AI systems will magically become ethical after being introduced to a Greek letter and a spreadsheet.

The mechanism: responsibility splits, it does not disappear

The PPD framework begins with a compact equation:

The terms are not as frightening as they look.

$K$ represents verified knowledge or epistemic capacity. A more informed agent has more responsibility because it has more usable understanding. $HI$ is the Humility Index, a normalized measure of recognized uncertainty, bias, or limitation. $C_{\text{signal}}$ is the contextual signal: the seriousness, urgency, or moral salience of the situation. The function $g(C_{\text{signal}})$ maps that contextual signal into amplification, using a linear, exponential, or logistic form depending on the domain.

The paper then separates total duty into two components:

$D_a$ is Action Duty: the obligation to act decisively when knowledge is strong and uncertainty is low. $D_r$ is Repair Duty: the obligation to verify, inquire, monitor, escalate, or otherwise reduce uncertainty when humility rises.

This is the paper’s best move. It avoids two common governance failures.

The first failure is overconfidence. When an agent has high knowledge but low humility, the system behaves as if analytical sophistication were moral permission. Anyone who has watched a confident model produce a polished hallucination will recognize the vibe.

The second failure is false restraint. When uncertainty rises, institutions often interpret uncertainty as a reason to step away from responsibility. PPD says no: uncertainty reduces the duty to act directly, but increases the duty to repair the epistemic gap.

In business terms, this means uncertainty should not merely create a red flag. It should create a work order.

The equation is not a moral oracle; it is a routing device

A careless reading would treat PPD as an ethics calculator: feed in knowledge, humility, and urgency; receive the correct moral answer. That would be a very fast way to turn a useful framework into a compliance toy.

A better reading is operational. The equation routes responsibility among possible modes of action.

This framing matters because many AI governance systems are still designed as permission systems: allow, block, escalate, or ignore. PPD instead suggests a proportional controller. A decision can be partly actionable and partly repair-bound.

For example, an AI triage system may not be confident enough to recommend an intervention autonomously, but it may be obligated to alert a clinician, request missing tests, and preserve a complete audit trail. A financial risk model may not justify liquidating a position, but it may justify stress-testing assumptions, reducing leverage, and raising disclosure quality. A customer-service automation may not be certain enough to deny a claim, but it may be certain enough to ask for missing documents and prevent the case from being closed.

The practical shift is from “the system is uncertain, so it stops” to “the system is uncertain, so its duty changes.”

That is more useful, and less glamorous. Naturally, it is also harder.

What the simulations actually support

The paper uses Monte Carlo simulations to test whether the PPD equation behaves as promised across randomized values of $K$, $HI$, and $C_{\text{signal}}$. The simulations use 100,000 iterations per signal-function configuration, with linear, exponential, and logistic versions of $g(C_{\text{signal}})$. Across the full results section and appendices, the paper reports 300,000 aggregated trials for the main signal-function comparisons.

The important thing is to classify the evidence correctly.

This is a respectable kind of evidence, but it is not the kind some readers may imagine.

The simulations show internal coherence. They test whether the proposed controller does what it says on the box. They do not show that a hospital committee, a regulator, or a model-risk team will estimate $HI$ well. They do not show that a firm will choose the right $g(C_{\text{signal}})$ function. They do not show that organizational incentives will politely step aside because the equation has arrived wearing a lab coat.

The strongest result is not “PPD is validated as an ethics system.” The stronger and more careful claim is: under the paper’s assumptions, PPD behaves as a stable redistribution mechanism between action and verification.

That is enough to be interesting.

Why the humility term is doing the serious work

The most business-relevant term in the paper is not knowledge. Most organizations already worship knowledge, or at least dashboards that resemble it. The interesting term is humility.

In PPD, humility is not politeness. It is not a LinkedIn leadership virtue. It is a damping coefficient.

When $HI$ rises, Action Duty falls because the system recognizes that direct action is less justified. But Repair Duty rises because uncertainty itself creates work. The model does not reward ignorance. It prices ignorance.

That is the conceptual bridge to AI governance. Modern AI systems already produce confidence scores, uncertainty estimates, ensemble disagreement, retrieval scores, human override rates, and incident logs. None of these automatically creates moral responsibility. But they can become inputs to a duty-routing layer.

For a business process automation system, $HI$ could be proxied by:

• model calibration error;
• disagreement across models or reviewers;
• missing data rate;
• retrieval uncertainty;
• policy ambiguity;
• recent incident frequency;
• distance from training distribution;
• sensitivity of the affected user or transaction.

Some of these are technical metrics. Others are governance metrics. That is exactly the point. A real system would not measure humility as a mystical property. It would operationalize humility as structured awareness of uncertainty.

The danger is obvious: once humility becomes a metric, someone will optimize around it. A vendor might lower the reported uncertainty to avoid escalation. A team might inflate humility to avoid accountability. A model might learn that “I am uncertain” is a cheap escape hatch. Congratulations, ethics has discovered gaming incentives. Again.

This is why PPD is more plausible as an audit framework than as a fully autonomous moral engine. It can structure what must be logged, reviewed, and justified. It should not be treated as a self-certifying conscience.

Contextual signal is where business judgment enters

The function $g(C_{\text{signal}})$ is the paper’s way of saying that not all uncertainty deserves the same response.

Uncertainty in recommending a playlist is not uncertainty in detecting sepsis. Uncertainty in approving a low-value refund is not uncertainty in denying insurance coverage. Uncertainty in summarizing a news article is not uncertainty in controlling a vehicle near a crosswalk.

The paper tests linear, exponential, and logistic versions of the contextual signal function. The interpretation is straightforward:

• linear mapping treats urgency as proportional;
• exponential mapping makes high-urgency contexts intensify duty more sharply;
• logistic mapping allows strong response while keeping amplification bounded.

This is where the framework becomes useful for business design. Organizations already classify risk: low, medium, high, critical; Tier 1 versus Tier 4; reversible versus irreversible; regulated versus unregulated; human-impacting versus administrative. PPD gives those categories a stronger role than dashboard decoration.

A customer support bot handling a billing typo can use a low contextual signal. A medical triage assistant should use a higher contextual signal. A trading agent in volatile markets may need a contextual signal that reacts to liquidity, leverage, contagion risk, and operational latency. A hiring-screening system may need high signal amplification because the harm is not just prediction error; it is institutional exclusion with a polite interface.

The business lesson is not “add ethics.” The business lesson is more concrete: the same model confidence score should trigger different action-repair splits depending on context.

That is how serious governance systems already behave. PPD formalizes the logic.

The case studies are interpretive, not field validation

The paper applies PPD to clinical ethics, recipient-rights law, economic governance, and AI systems. These sections are useful, but they should be read as simulated demonstrations.

The clinical case considers a rehabilitation therapist deciding whether to approve a temporary home pass. With $K = 0.75$, $HI = 0.40$, and $C_{\text{signal}} = 0.60$, the paper calculates Action Duty of 0.45, Repair Duty of 0.18, and Total Duty of 0.63. The interpretation is not “approve blindly.” It is closer to “proceed with safeguards”: supervision, check-ins, and verification.

The recipient-rights case uses a temporary guardianship scenario with $K = 0.80$, $HI = 0.50$, and $C_{\text{signal}} = 0.70$. The model produces Action Duty of 0.40 and Repair Duty of 0.28, yielding Total Duty of 0.68. Again, the mechanism pushes away from binary thinking. The decision is not full autonomy or full restriction; it is limited intervention plus ongoing verification.

The economic governance case interprets the 2008 financial crisis as high knowledge with low humility. In the paper’s late-2006 illustration, $K = 0.90$, $HI = 0.10$, and $C_{\text{signal}} = 0.40$, producing strong Action Duty and negligible Repair Duty. This is an elegant interpretation of moral over-leverage: sophisticated actors had enough knowledge to act aggressively, but not enough humility to stress-test their own assumptions.

The AI systems case is the closest to Cognaptus territory. The paper’s autonomous-vehicle example uses $K = 0.70$, $HI = 0.30$, and $C_{\text{signal}} = 0.80$, producing Total Duty of 0.658. The operational implication is restrained action: brake, re-scan, widen the safety margin, and prepare to stop.

Across these cases, the pattern is consistent. PPD does not simply say “be careful.” It says what carefulness becomes operationally: partial action, verification, logging, escalation, or bounded intervention.

That is valuable. But the examples are still illustrative. They do not establish calibrated thresholds for actual hospitals, courts, banks, or vehicles. The correct response is not to copy the numbers. The correct response is to copy the architecture and then do the hard calibration work.

Yes, the annoying part. The part where governance becomes governance.

How this maps to AI and business process automation

For business users, the strongest application is not a philosophical ethics dashboard. It is an auditable decision-routing layer for automated systems.

A practical PPD-inspired architecture would look like this:

This is where the paper becomes more than ethical language. Many automation failures are not caused by the absence of a policy. They are caused by a policy that cannot distinguish between “act now,” “act partially,” “ask for more evidence,” and “escalate before harm becomes irreversible.”

PPD gives a vocabulary for that middle space.

Consider three common business automation settings.

In document processing, high $K$ and low $HI$ could allow automatic extraction and filing. High $HI$ should trigger repair: request missing pages, compare against historical patterns, or ask a reviewer to inspect a field. If the document affects compliance, $C_{\text{signal}}$ rises and the system should intensify verification.

In AI customer support, uncertainty should not always mean human handoff. For low-stakes cases, repair may mean asking a clarifying question. For high-stakes cases, such as account suspension, insurance denial, refund rejection, or medical billing, repair may mean escalation and audit logging.

In trading or financial-risk automation, a model with strong signal but rising uncertainty should not be allowed to simply increase leverage because the expected return is positive. Repair Duty might require scenario testing, liquidity checks, exposure limits, or temporary risk reduction. The trader may still act. The system just refuses to confuse confidence with immunity from reality.

This is the real ROI pathway: fewer catastrophic misroutes, better audit trails, more defensible escalation, and faster diagnosis after incidents. The value is not that PPD makes decisions morally pure. The value is that it makes uncertainty operational instead of decorative.

The OpenAI “confessions” comparison should be handled carefully

The paper includes a section comparing PPD with OpenAI’s “Training LLMs for Honesty via Confessions,” published in December 2025 according to the paper’s reference list. The author argues that separating the reward signal for confession honesty from task performance resembles PPD’s separation of Action Duty and Repair Duty.

This is an interesting comparison, but it should not be overplayed.

The useful analogy is architectural. If a model can be rewarded separately for task performance and honest self-reporting, then “doing the task” and “repairing uncertainty or misconduct” can be treated as distinct duties. That aligns neatly with PPD.

The boundary is that this does not make the OpenAI work a direct empirical validation of the full PPD framework. It may support a similar design intuition: systems need independent pathways for action and reporting. But PPD’s full structure includes $K$, $HI$, contextual-signal amplification, baseline humility, and cross-domain duty redistribution. A confessions mechanism does not automatically validate all of that.

So the correct business takeaway is modest and useful: separate performance incentives from verification incentives. If your automation platform rewards only task completion, it will learn to complete tasks. Shocking. If you want it to surface uncertainty, conflict, or possible error, those behaviors need their own channel, measurement, and protection from performance pressure.

This is not philosophy. This is incentive design with a nicer vocabulary.

Where the paper is strongest

The paper’s strongest contribution is conceptual engineering. It turns a vague moral intuition into a testable structure:

• knowledge increases responsibility;
• uncertainty shifts responsibility toward repair;
• context changes how urgently repair or action should occur;
• a non-zero humility baseline can stabilize decisions;
• decision traces can be recorded as structured tuples rather than post-hoc explanations.

That last point is especially important. The tuple $(K, HI, C_{\text{signal}})$ is not merely a calculation. It is an audit object. If a system acts, the organization can ask what knowledge supported action. If it hesitated, the organization can ask what uncertainty justified repair. If it escalated, the organization can ask what contextual signal intensified the duty.

Good governance is partly the art of making excuses expensive.

PPD helps by asking systems to state their epistemic position before acting. That does not guarantee wisdom, but it raises the cost of pretending certainty after the fact.

Where the paper is still unfinished

The paper is clear about several limitations, and they are not minor.

First, the model reduces moral duty to scalar values. That is acceptable for a first-order framework, but many real decisions contain conflicting values that cannot be compressed cleanly into one number. Privacy, safety, autonomy, fairness, and efficiency often pull in different directions. A scalar model can organize the conversation, but it cannot contain the whole moral universe. Annoying, but true.

Second, $HI$ and $C_{\text{signal}}$ are not standardized. This is the main practical barrier. A company cannot simply “implement humility” unless it defines observable proxies, measurement rules, thresholds, and review processes. Otherwise, humility becomes a poetic variable with enterprise pricing.

Third, the signal function $g(C_{\text{signal}})$ is domain-dependent. A linear function may be fine for routine administrative risk. High-stakes safety or medical contexts may require exponential or logistic amplification. The paper tests these forms as model variants, but real calibration remains future work.

Fourth, the simulations use randomized inputs and controlled mathematical conditions. They validate model behavior, not institutional adoption. Human organizations contain incentives, politics, legal exposure, budget constraints, and the occasional executive who thinks a confidence score is a personality trait.

Fifth, AI implementations could game the structure. A system might learn to manipulate uncertainty reporting, suppress escalation triggers, or exploit “repair” as a delay tactic. Any real implementation would need adversarial testing, incident review, independent monitoring, and separation between model operators and audit owners.

Finally, the paper notes a licensing boundary: the work is released under a non-commercial Creative Commons license, and commercial use of the framework, equation, or derivatives requires a separate license. That matters for businesses. Reading the paper is not the same as embedding the framework into a product.

What Cognaptus infers for business use

The paper directly shows a formal model and simulation evidence that the model behaves coherently under its own assumptions. It directly offers simulated applications across domains and an implementation-oriented view of AI systems. It does not directly show market performance, regulatory acceptance, clinical outcome improvement, or reduced AI incidents.

Cognaptus can reasonably infer three business uses.

First, PPD can inform escalation design. Instead of binary automation thresholds, workflows can distinguish act, verify, defer, and escalate based on confidence, uncertainty, and harm context.

Second, PPD can improve auditability. A structured record of $K$, $HI$, $C_{\text{signal}}$, decision mode, and outcome gives incident reviewers something better than “the model said so,” that most sacred sentence in mediocre automation.

Third, PPD can guide incentive separation. If systems are evaluated only on speed, completion, or accuracy, they will underreport uncertainty. Verification and honest repair need independent metrics.

What remains uncertain is calibration. Which uncertainty proxies should count as humility? How should different industries map contextual risk? What thresholds trigger human review? How should repair duties be costed? How do we prevent teams from tuning the system to avoid accountability?

Those are not side questions. Those are the implementation.

The real contribution: uncertainty becomes a task

The best sentence this paper implies is not “to know is to owe,” although the paper uses that formulation. The more operational version is this: to not know is also to owe.

If an AI system lacks certainty, it owes verification. If a firm has high knowledge and low humility, it owes stress testing. If a regulator sees contextual signals rising, it owes earlier intervention. If an automated workflow cannot justify direct action, it may still owe escalation, logging, or evidence collection.

That is the paper’s most useful contribution. It does not solve ethics. It gives uncertainty a job description.

For AI governance and business process automation, that is a serious improvement. Most organizations already know how to act when confidence is high. They are less good at deciding what uncertainty should do next. PPD offers a clean answer: uncertainty should not become silence; it should become repair.

A mature automation system will not merely decide faster. It will know when speed is the wrong performance metric.

And yes, that means the future of responsible AI may involve more logs, more thresholds, more calibration studies, and more awkward meetings about whether a model’s “humility” is real or just another metric pretending to be a virtue.

Ethics was never going to be frictionless. Anyone selling frictionless ethics is probably selling the friction to someone else.

Cognaptus: Automate the Present, Incubate the Future.