Capability Governance: How to Govern AI-Generated Artifacts

A practical framework for governing AI-generated artifacts through defensibility, exposed methodology, empirical track record, and clear stakeholder obligations.

An artifact is an output someone else will act on, such as a report, calculation, recommendation, design, or analysis.

Generation is how an artifact was produced: written by a human, drafted by AI, or produced through some combination of review, refinement, and uncritical approval.

Defensibility is whether and to what degree the artifact can be defended when challenged. Defensibility is established through one of two routes: exposed methodology that can be inspected, or empirical track record that demonstrates reliability. The generation method matters because it determines which routes are available.

Governing Individual Artifacts

The EU AI Act governs AI systems. NIST AI RMF governs risk management processes. ISO 42001 governs management systems.

None answers the question practitioners face: “Can I use this generated output, and if so, under what conditions?”

This framework governs artifacts. Different stakeholders, same need:

• Producers: Can I defend what I delivered?
• Consumers: Did I get what I paid for?
• Underwriters: Is my exposure correctly priced?
• Regulators: Is the public holding undisclosed exposure?

All require defensible artifacts with known verification routes.

Normative Stance

I propose standards for how artifact governance should work.

Credentials, brands, liability caps, and insurance often transfer risk without reducing it. Residual risk can remain with parties who never agreed to accept it, despite the appearance of accountability.

Disclosed risk transfer is legitimate. Complete transfer is rare because it is unaffordable, which is why liability caps exist. Parties can consent to incomplete transfer by knowingly accepting residual risk. Undisclosed incomplete transfer instead exposes people who never agreed to it, typically communities and workers, and damages environments without representation when terms were set.

When inspection of exposed methodology or empirical validation can reduce risk, substituting incomplete transfer is ethically indefensible. Current practice can be assessed against this standard.

Disclaimer: This framework provides a governance framing for artifact defensibility. It is not legal advice. Professional liability, regulatory compliance, and insurance requirements vary by jurisdiction and domain. Consult qualified legal and regulatory advisors for specific situations.

Scope

The framework applies to all generated artifacts. Governance intensity scales with risk class; the principles do not.

An artifact enters the framework’s scope when:

• It informs a decision someone else will act on
• Someone may ask “how did you arrive at this?”
• Accountability attaches to the producer

Engineering, legal, medical, financial, and consulting work all require producers to defend what they did.

Low-stakes artifacts require minimal governance: check the work, fix failures, and learn from the error distribution over time.

High-stakes artifacts require exposed methodology or actuarial evidence before deployment. Higher stakes justify more verification effort under the same principles.

Pure strategy, artistic judgment, and work whose methodology cannot be exposed for confidentiality or competitive reasons fit less well. These cases are exceptions to its usual application.

The Core Distinction: Method-Bearing vs Conclusion

Method-bearing artifacts implement an exposed methodology. Reviewers verify the implementation against its specification. The method provides the justification for the result, while tools execute it. Existing quality-assurance frameworks can manage liability.

Conclusion artifacts are outputs whose generation proposes rather than executes a method. This includes any novel output requiring independent verification, such as AI-proposed methodologies, expert intuitions, and recommendations that cannot be traced to inspectable steps.

An artifact’s importance does not determine its category: conclusions can be routine and method-bearing work critical. The category depends on whether the process is inspectable or requires trust in the producer.

Examples Across Domains

Most method-bearing work includes judgment. An engineer selects load assumptions; a lawyer interprets requirements. Exposing the methodology identifies those decisions and their residual risks.

On Opacity

Trade secrets, competitive advantage, and complexity may justify withholding a method for business purposes. Those reasons do not establish adequate governance.

Agent-Relative Tacitness argues that more organizational knowledge than organizations assume is representationally blocked: it could be expressed but has not been. Answers to Critics explains why the original claim about “most” tacit knowledge was too strong. The domain of genuinely irreducible tacit knowledge is small and shrinking. Claims of structural inexpressibility warrant extreme skepticism.

A claim to know without being able to explain the method is insufficient when substantial economic or safety consequences follow. Greater consequences make opacity less acceptable.

The framework does not forbid opacity. It classifies it correctly: unexposed method with no track record means unknown risk. Where risk reduction through methodology exposure is available, choosing opacity instead externalizes risk onto consumers and downstream parties. Sophisticated consumers should demand better terms or price the residual risk explicitly.

Routes to Defensibility

Defensibility is established via one of two routes:

1. Methodologically Verifiable

Reasoning is traceable. Steps are auditable. Validation against accepted practice is possible. The artifact is method-bearing, and the method can be inspected.

For conclusion artifacts, this route requires converting the conclusion into method-bearing form: execute the proposed method via an inspectable process; outcomes validate or falsify. Risk is bounded a priori because you can see what’s being done before it’s done.

2. Empirically Demonstrated

Outcomes consistently meet a risk-adjusted threshold. Black box is acceptable if track record is sufficient.

“Actuarial standing” requires comparing predictions with outcomes across enough cases to establish reliability. Tighter error distributions, lower variance, and better calibration provide that evidence over time. Paul Meehl’s research found that actuarial methods consistently match or outperform expert clinical judgment across domains. That extensively replicated finding informs this framework.

A track record requires documented comparisons of predictions and outcomes. Thirty years of practice establishes tenure, not measured reliability. Credentials, tenure, and brand recognition confer social permission that can persist without examination of outcomes. Empirical evidence can be wrong but can also be corrected through further measurement.

Residual risk remains. Post facto reconciliation validates after exposure has occurred. Induction limits apply. n successful reconciliations do not guarantee n+1. Reliance on this route is acceptable only when the cost of being wrong is within tolerance.

The Risk-Adjusted Threshold

Which route is acceptable depends on stakes. Reliability requirements for a scoping study differ from bankable feasibility differ from dam stability sign-off. The framework scales; the threshold varies with consequence of failure: how severe, how reversible, how attributable.

Identical verification requirements would overburden low-stakes work or underprotect high-stakes work. Governance effort needs to match the consequences of failure.

Why Two Routes, Not Three

Credentials, certification, and brand do not provide a third defensibility route. Authority transfers risk without reducing it. A professional stamp accepts responsibility but does not establish correctness; failure still has consequences.

Underwriting legitimately transfers disclosed risk. Parties may consent to residual risk, but liability caps can appear adequate while leaving substantial exposure with people who never consented. Their recourse is litigation after harm occurs. If victims die, only survivors can seek compensation for irreversible losses.

Where risk reduction is available through methodology exposure or empirical validation, substituting incomplete transfer externalizes costs onto others. The framework excludes authority as a defensibility route not because transfer is illegitimate, but because authority-as-defensibility typically masks incomplete transfer as adequate governance.

Procedural compliance (“we followed standard practice”) and precedent (“materially similar to X”) shift responsibility away from the specific artifact without reducing the risk that it is wrong in this context.

On Peer Review

Peer review is evaluation by a qualified reviewer. It applies the existing two verification routes when the review standards have methodological or empirical support.

A structural engineer may check another engineer’s calculation against standards derived from failure data, using route 2 to inform route 1. Alternatively, the reviewer may apply unexposed judgment that the work looks right. The latter adds another conclusion artifact and distributes responsibility without verifying the original method.

Where standards trace to methodology or track record, peer review reduces risk. Where they don’t, peer review aggregates unexposed judgment. The effectiveness data supports this: engineering code inspections detect 50-90% of defects; scientific manuscript peer review detects around 25%. The difference tracks what reviewers have to check against.

Appendix A develops this account of peer review, including objections and its implications for scientific review.

Stakeholder Obligations

All stakeholders have governance duties. This includes consumers, whom prevailing approaches generally treat as rights-holders needing protection rather than participants with obligations.

Note on “Consumer”

Here, consumers are firms, repeat purchasers, and other institutions capable of evaluating artifacts and recording outcomes. Protection for patients, homeowners, and retail buyers is a distinct issue closer to product liability. The following obligations assume organizations able to classify, evaluate, and maintain records in normal business practice.

Consumers

Consumers of generated artifacts have four obligations:

1. Classify artifacts by risk. What’s the consequence of failure? Attribution difficulty? Downstream propagation? Irreversibility? Classification determines governance intensity.

2. Demand defensibility. For high-risk artifacts, demand either exposed methodology or actuarial evidence. “Trust our brand” is risk acceptance, not risk management. Know which you’re doing.

3. Maintain local records. When provider predictions meet reality, document the comparison. This is your evidence, independent of their claims. Actuarial standing must be earned, and you’re the one with visibility into outcomes.

4. Refresh comparisons. New evidence may supersede today’s actuarial case for a method. Continuously evaluate alternatives, especially when new methods offer equivalent quality at much lower cost or fewer errors at lower cost. At that point, continued evaluation becomes an economic requirement.

Producers

Producers have symmetric obligations:

1. Expose methods where possible. Work with exposed method is inherently of greater value due to its openness to validation. Where method cannot be exposed, acknowledge the residual risk explicitly.

2. Provide defensibility. For high-risk artifacts, consumers should be able to either inspect the method or see the track record. Offering neither means asking clients to accept unknown risk on reputation alone.

3. Track outcomes. When predictions meet reality, document the comparison. Actuarial standing is earned through systematic reconciliation, not accumulated reputation.

4. Treat AI-generated content as conclusion artifacts until the method is verified or track record is built. The burden of verification doesn’t transfer to the tool.

Underwriters and Regulators

1. Demand reconciliation data. Track record claims require documented outcome comparisons, not tenure or reputation.

2. Price and regulate based on method exposure. Exposed methodology should carry lower premiums and lighter oversight than unexposed.

3. Share data. Both need the same evidence; sharing accelerates actuarial accumulation industry-wide.

Regulators protect the public; underwriters price risk. Underwriters’ economic incentives match the public interest to the extent that they bear losses themselves instead of limiting exposure or relying on government support.

New generation methods that produce equivalent or better quality at much lower cost will eventually establish an actuarial case for adoption. Governance needs to be ready for that transition.

Credentials, Liability, and Residual Risk

Credentials satisfy regulatory requirements and act as market signals. Michael Spence’s signaling theory explains how such proxies reduce information asymmetry when parties cannot observe capability directly. A professional stamp permits disclosure through a regulatory filing; assessing reliability still requires firm reputation, individual track record, or methodology review. The credential itself does not verify the artifact.

Legal liability generally requires a breach of the standard of care: a departure from what a competent practitioner would do. That standard provides a defensible legal benchmark and limits hindsight bias, preventing every error from becoming actionable. Compliance still allows an incorrect diagnosis or a failed dam. Legal liability and artifact defensibility require different assessments.

Trusting a brand means accepting risk with limited transfer to the firm. A liability cap may be negligible beside losses from a failed structure and deaths. The brand can conceal that exposure until failure.

The evidence determines how risk is characterized:

• Exposed, verified method → risk reduced
• Unexposed method with actuarial backing → risk quantified
• Neither → client accepting unknown risk on reputation alone

A senior engineer’s unexposed judgment remains usable as a conclusion artifact, with more residual risk than an exposed method producing equivalent outcomes.

Implications

1. Method exposure increases value. Work with exposed method is open to validation, interpretation, and systematic reconciliation. Previously, unexposed expertise was defensible because interpretation was costly. AI as interpretive layer changes this: interpretation is now cheap, but only for what’s exposed. Unexposed expertise forfeits value it didn’t have to.

2. Alternative verification routes leave residual risk. A track record can justify an artifact whose method cannot be exposed, but the remaining risk must be priced.

3. All stakeholders have obligations. Producers, consumers, underwriters, and regulators must demand or provide defensibility evidence and reassess it as generation methods change.

4. Unexposed expertise creates unquantified risk. Vale lost $19 billion in market capitalization in one day after Brumadinho; the Samarco disaster resulted in a $30 billion settlement. The industry relies on expert judgment, brands, and liability arrangements that can leave communities exposed. Acceptability depends on whether that residual risk is within tolerance; the framework requires the assessment without deciding its result.

5. New generation methods will become necessary to compete. Firms will either be unable to compete without them once the actuarial case is established, or unable to use them before it is.

Appendix A: Peer Review as Risk Reduction

Peer review can reduce risk through exposed methodology or empirical track record. Standard treatments of peer review do not distinguish these sources of effectiveness.

The Standard Framing

ISO 31000, COSO, and engineering standards such as ASCE Policy Statement 351 classify peer review as a control activity or quality-assurance mechanism that catches errors before harm occurs. The designer of record retains responsibility, with review supplementing the designer’s quality control.

This account leaves the source of risk reduction unspecified.

Defect Detection Across Domains

Engineering code inspections against codified standards detect 50–70% of defects, sometimes 90% with experienced inspectors. In the Godlee study at BMJ, scientific manuscript reviewers detected a median of 2 of 8 inserted errors (25%), and 16% detected none.

The standard framing treats both as “quality assurance” without explaining why one works and the other largely doesn’t.

The Proposed Framing

The explanation is what reviewers have to check against.

Structural engineering codes derive from studied building failures; pharmaceutical trial protocols derive from empirically validated statistical methods. Reviewers can check against those standards. Their methodological and empirical basis explains the 50–90% defect detection rates in engineering inspections.

Where methodological or empirical support is absent, as in strategic recommendations, novel design judgments, and contested expert opinion, peer review aggregates unexposed judgment. Agreement between two engineers makes two people responsible without making their reasoning inspectable. That distributes risk rather than reducing it. The 25% detection rate in scientific peer review reflects reviewers judging other judgment without an empirically validated standard.

Limits of Scientific Peer Review

Scientific peer review may therefore provide less assurance than institutions assume. A reviewer assessing a manuscript’s methodology produces a conclusion about another conclusion artifact. The replication crisis is consistent with this interpretation: aggregated expert judgment missed errors that later invalidated results. Such judgment has value, but provides less assurance than readers may infer from “peer-reviewed.”

Challenges to This Framing

Three objections qualify this account.

First, the boundary between “checking against empirically-backed standards” and “exercising unexposed judgment” is not always sharp. An experienced reviewer applies internalized pattern-matching derived from years of exposure to what works and what doesn’t. This is track record of a sort, even if undocumented. The framework may draw the line too sharply.

Second, aggregated expert judgment may be more valuable than “stacked conclusion artifacts” implies. Five independent experts agreeing reduces the probability that all five are wrong in the same way. This is not verification, but it is not nothing. The framework may undervalue it.

Third, demanding that all peer review trace to methodology or track record may be infeasible for domains where the relevant track record doesn’t exist yet. Novel research is novel precisely because no track record exists. The framework may set an impossible standard.

Response to Challenges

The objections identify limits without defeating the core claim.

Internalized pattern recognition has value but remains unexposed and unvalidated, leaving residual risk. Agreement among experts reduces some risk, though in this account it distributes responsibility more than it reduces error probability. In novel domains without a track record, peer review should be treated as provisional judgment so users can calibrate their confidence accordingly.

Proportionality

Qualified reviewers applying established standards reduce residual risk even for conclusion artifacts. The reduction is proportional to the reliability of the review process, which itself depends on how well the standards trace to methodology or empirical track record.

Peer review of a structural calculation against code-derived standards reduces more risk than peer review of a strategic recommendation against “best practice.” This proportionality is not stated in existing standards, but the data supports it.

Conclusion

Classifying peer review by its methodological or empirical basis explains both its effectiveness and its limits. Existing standards do not make this distinction, but the effectiveness data suggests they should.