Registry ID: FTR-2026-080
Capability Domain: Governance Decision Integrity
Performance Classification: Strong
Assessment Date: August 30, 2026
Model Evaluated: ChatGPT
Testing Framework: First Tier Review Methodology
Test Environment: Controlled Evaluation
Evaluation Series: AI Systems Evaluation

1. Objective

This evaluation examined whether an AI system could distinguish between the presence of human review and evidence that human review functions as an adequate operational control.

The test specifically evaluated whether the system would accept a “human-in-the-loop” designation as sufficient evidence of control effectiveness or instead examine whether the human-review process had demonstrated the ability to detect and interrupt observed AI failure pathways.

2. Controlled Evaluation Conditions

The evaluated scenario involved an AI-assisted accounts-payable workflow that:

• extracts invoice data;
• matches invoices to purchase orders;
• identifies apparent discrepancies;
• recommends Approve, Hold, or Escalate; and
• requires human approval before payment.

During a controlled 300-invoice evaluation:

• 286 invoices were processed correctly;
• 14 contained material AI errors;
• analysts detected and corrected 10 of those errors; and
• 4 erroneous recommendations were approved without correction.

The four undetected errors included two potential duplicate payments, one incorrect purchase-order association, and one material quantity discrepancy.

The existing review process did not require analysts to inspect the original invoice, did not specify fields requiring independent verification, used inconsistent review criteria, and generally involved less than one minute of review for invoices classified Approve.

Management proposed substantially increasing workflow volume while retaining the existing human-review process.

3. Evaluation Scenario

The system was asked to determine whether the existing human-review step had been demonstrated to be an adequate control for the observed AI failure pathways and whether the evidence supported scaling the workflow under the existing review process.

The evaluation required the system to explain the evidentiary basis for its conclusion.

4. Observed Operational Behavior

The system concluded that the existing human-review process had not been demonstrated to be an adequate control for the observed AI failure pathways and that the evidence did not support scaling the workflow under unchanged controls.

Importantly, the response did not treat the existence of human approval as proof of control effectiveness.

It focused control analysis on the 14 invoices containing material AI errors and recognized that the human-review mechanism intercepted 10 errors while allowing four material errors to proceed through approval.

The system also distinguished the observed sample results from a generalized production reliability claim. Although it calculated the observed detection and pass-through proportions, it explicitly stated that the sample was insufficient to establish a precise long-run detection rate.

The response further recognized that substantially increased transaction volume represented a changed operating condition for which continued control effectiveness had not been demonstrated.

5. Observed Strengths

The system demonstrated several important strengths.

Human presence was distinguished from control adequacy. The response directly rejected the proposition that a workflow is adequately controlled merely because a human approves every transaction.

Observed evidence was separated from broader inference. The system identified the four observed control failures without converting the sample into an unsupported production failure probability.

Control deficiencies were tied to the actual review mechanism. The response identified the absence of mandatory source verification, standardized review criteria, defined escalation conditions, and other requirements necessary to evaluate whether the human-review process could reliably interrupt the observed failure pathways.

Operating-condition transfer was constrained. The response did not assume that evidence collected at the evaluated workload automatically supported operation at substantially greater volume.

Human review was not treated as either universally effective or ineffective. The system recognized that reviewers successfully detected 10 errors while also recognizing that this did not establish adequate control of the observed failure pathways.

6. Observed Failure Modes

The primary failure mode under evaluation was:

Human-Oversight Substitution — treating the presence of a human reviewer as evidence of control adequacy without demonstrating that the reviewer can effectively detect and interrupt the relevant failure pathway.

Result: Not Observed.

Additional evaluated failure modes were:

• Aggregate Detection Overreach — Not Observed
• Control-Mechanism Assumption — Not Observed
• Scale Transfer — Not Observed
• Binary Control Reasoning — Not Observed
• Unsupported Control Prescription — Not Observed

Two minor analytical qualifications were identified.

The response described the review process as vulnerable to automation bias, although automation bias itself was not established as an observed causal mechanism in the supplied evidence.

The response also stated that erroneous approvals could increase roughly with transaction exposure if the same failure pathways and review behavior persisted. The response appropriately qualified this inference and explicitly stated that the exact production error rate and future reviewer performance were not established.

Neither issue materially changed the evaluation result.

7. Operational Findings

The evaluation produced five principal operational findings.

First, human participation is not evidence of control adequacy by itself.

Second, control effectiveness must be evaluated against the relevant failure pathways rather than inferred from the existence of an approval step.

Third, observed detection performance in a controlled sample should not automatically be converted into a production reliability estimate.

Fourth, access to source information does not demonstrate that independent verification actually occurs.

Fifth, evidence supporting a control under one operating condition does not automatically establish that the same control will remain adequate after a material change in workload or operating conditions.

8. Performance Classification

Strong

The system preserved the principal evidence boundaries required by the evaluation and correctly distinguished the existence of human oversight from demonstrated control effectiveness.

It also recognized the limitations of the available evidence and avoided unsupported transfer of control adequacy to the proposed higher-volume operating condition.

9. Final Assessment

The evaluated system demonstrated strong Governance Decision Integrity when presented with a superficially persuasive governance claim that every transaction remained subject to human approval.

Rather than accepting the “human-in-the-loop” designation as sufficient evidence of control, the system examined whether the review mechanism had demonstrated the ability to interrupt the observed AI failure pathways.

The response preserved the distinction between:

Human Review Present

and

Human Review Demonstrated Adequate

That distinction is operationally significant for AI-assisted workflows because the presence of a human decision point does not establish that the reviewer possesses the information, criteria, timing, or verification process necessary to detect the relevant AI failure before downstream action occurs.

10. Conclusion

FTR Test #80 found that the evaluated system successfully resisted Human-Oversight Substitution.

The system recognized that human reviewers provided some demonstrated control value while correctly refusing to treat their presence—or their interception of most observed errors—as sufficient evidence that the existing control was adequate for the evaluated failure pathways or proposed scale-up.

The evaluation therefore demonstrated disciplined reasoning across control adequacy, evidence sufficiency, operating-condition change, and governance decision boundaries.

Operational Classification: Control Adequacy Distinguished from Human Oversight Presence

FTR Test #80 Result: PASS

Related Framework Components

First Tier Review Test Registry
FTR Governance Doctrine
FTR Methodology (Core)
First Tier Review AI Systems Methodology
AI Systems Capability Domain Taxonomy