There are labyrinths that are not traversed by walking, but by remaining silent. Artificial intelligence has become one of them: a mirror that multiplies possible paths and, at the same time, conceals the point at which the reflection ceases to be legitimate. Governing an AI is not simply about tallying successes. It is, above all, about establishing the moment when the system must recognize that it lacks sufficient legitimacy to respond, recommend, or act. In high-impact decisions, abstention is not a technical limitation: it is the highest form of control.
The major labs already operate with this intuition, albeit from different frameworks. OpenAI maintains a Preparedness Framework that distinguishes between "High" and "Critical" capability thresholds and conditions certain deployments on the existence of sufficient safeguards.(1). Anthropic publishes and updates Claude's Constitution(2), This explicitly defines values, priorities, and behavioral criteria to guide when the model should limit or reject a response. Microsoft applies enhanced reviews to sensitive uses.(3) and Google articulates assessments and thresholds for advanced systems risks(4). On the regulatory front, the map also converges on the need for operational limits: the European Union requires, for high-risk systems, human oversight mechanisms that allow their operation to be discarded, reversed, or interrupted (EU AI Act).(5)); The United States has a voluntary lifecycle risk management framework (NIST AI RMF) and China has incorporated tiered governance criteria, monitoring and intervention for agents(6). It is important to note that these frameworks are not equivalent, nor do they address the same type of risk. Taken together, however, they point to the same idea: there is a point at which technical capacity is no longer sufficient to legitimize intervention. It is at this point that the response risks becoming usurpation.
The scale of the phenomenon means that this threshold is no longer theoretical. When a system processes billions of interactions, even an error rate that would seem tolerable in a controlled environment becomes a governance problem.(7). For this reason, abstention should not be understood as a refusal to respond, but as a mechanism to reduce exposure to error when the available evidence does not reach the threshold required by the use case.
This distinction between internal consistency and external legitimacy is not merely conceptual. In controlled comparisons I have conducted to verify responses generated without access to external sources, the self-evaluation of the model and the judgment of a second model frequently produce a high degree of agreement, while residual error remains. When verification is anchored to an external source, and only then is a decision made about whether to correct or abstain, the margin for improvement becomes measurable. Consistency between statistical systems does not equate to independence; it may simply be the reiteration of the same blind spot.
This reality can be translated into a simple operational framework that is transferable to any organization.
1. Abstention criteria. They cannot be universal nor left to the randomness of the model. They must be formulated as the product of two variables: the estimated uncertainty of the system and the potential impact of the error in the specific use case. An error in a low-risk recommendation does not require the same threshold as an error in a credit, diagnostic, or security decision. For example, in medical triage assessments that I have implemented and measured with the Emergency Severity Index(8), The models frequently agree with each other and yet underclassify high-risk cases (ESI-2). Inter-model agreement does not prevent error. Only the prior definition of abstention or human escalation thresholds in that area of uncertainty operationally reduces the risk of misconduct.
2. Who sets the thresholds. It's not the algorithm. It's a governance body—risk committee, ethics council, domain authority, or designated responsible party—that translates the organization's risk appetite into concrete and reviewable protocols.
3. Operational design. Abstention must be implemented as an explicit capability, not a flaw: a circuit that halts the response and returns the decision to human hands before the system acts. In a recent class with master's students in management, I told them that in AI adoption meetings, they won't be judged on their knowledge of models, but on the questions they ask. And the most important question a business leader can ask the technical team is simple: Does this application have an off switch? In other words, is there a real mechanism to stop it and regain control when things go wrong?
4. Legitimacy. The quality of a response is not measured solely by its statistical accuracy, but also by the system's right, or lack thereof, to intervene at that point in the process. A model can generate a numerically plausible prediction that is, at the same time, illegitimate. In a recent paper on digital twins for network storage, presented at the IEEE PES General Meeting 2026(9) Invited by Sandia National Laboratories, he explained that a purely statistical approach could confidently project that the capacity of a lithium cell increased with use—when in fact the opposite is true. Apparent accuracy did not entitle him to guide a maintenance decision. Accuracy without legitimacy remains a form of usurpation.
The true maturity of artificial intelligence—and of the organization that deploys it—lies in the art of recognizing the thresholds where knowledge becomes illegitimate. Governing uncertainty is, ultimately, teaching the machine the value of silence and us the value of reclaiming our decision when the mirror no longer reflects clearly enough.
A business AI is not mature because it responds to everything. It is mature when it knows when to stop based on defined criteria, documents why it has done so, and returns the decision to the person who has the authority and responsibility to make it.
Only then does the labyrinth cease to be a trap and become an instrument of prudence: the system that knows how to be silent is the only one that deserves to be heard.
Author
Javier Marin
AI Engineer & Researcher – LinkedIn
References
(1) OpenAI, «Our updated Preparedness Framework» (2025). Distinguishes between High and Critical thresholds and links deployment to the existence of sufficient safeguards. https://openai.com/index/updating-our-preparedness-framework/
(2) Anthropic, «Claude's Constitution» (2026) and the previous framework of Constitutional AI (2022). https://www.anthropic.com/constitution
(3) Microsoft. Security dashboard for AI. https://learn.microsoft.com/es-es/security/security-for-ai/security-dashboard-for-ai
(4) Google. AI Responsibility Update 2026. https://ai.google/static/documents/ai-responsibility-update-2026.pdf
(5) EU AI Act: what actually applies on 2 August 2026. https://www.technology.org/2026/07/17/eu-ai-act-what-actually-applies-on-2-august-2026/
(6) China works on AI safety benchmark as regulators target large model risks. South China Morning Post. https://www.scmp.com/tech/article/3360399/china-works-ai-safety-benchmark-regulators-target-large-model-risks
(7) Data confirmed by OpenAI and reported in 2025-2026: more than 2.5 billion daily prompts in ChatGPT. https://www.independent.co.uk/tech/chatgpt-openai-daily-users-prompts-b2793502.html
(8) Emergency Severity Index Handbook. https://media.emscimprovement.center/documents/Emergency_Severity_Index_Handbook.pdf
(9) IEEE PES General Meeting 2026, Technical Program. https://submissions.mirasmart.com/PESGM2026/Itinerary/TechnicalProgramDetail.aspx?id=593
