Artificial intelligence needs a common ruler Authors: Yang Zhao, Zhenfei Song, Pengyu Tan, Xingchuang Xiong, Xiang Fang Published: 2026 Main text Artificial intelligence (AI) is rapidly transitioning from technological breakthroughs into large-scale deployment across society, engineering, and scientific discovery. As AI increasingly participates in decision-making, knowledge generation, and critical infrastructure operation, a central challenge is becoming evident: the key limitation to future AI scaling is no longer capability improvement alone but whether trustworthiness can remain sustainable. Substantial progress has been made toward trustworthy AI. Ethical frameworks define normative principles, regulatory systems establish governance requirements, and technical methods improve evaluation and monitoring capabilities. Yet most efforts focus on what requirements trustworthy AI should satisfy? A more fundamental question remains insufficiently addressed: How can trust remain objective, comparable, and verifiable as AI continuously evolves across changing datasets, deployment environments, and application domains? Persistent AI challenges, such as hallucination, weak reproducibility, accountability gaps, and limited cross-context comparability, are not isolated technical failures. They reveal a deeper structural limitation: current AI governance lacks shared measurement foundations capable of supporting objective comparison, consistent validation and sustained oversight across contexts. Capability determines how broadly AI can scale. Trust foundations determine whether that scale remains sustainable. We suggest that trustworthy AI requires four foundational properties: comparability (C), traceability (T), reliability (R), and governability (G). Comparability enables objective evaluation through shared references; traceability establishes reconstructable evidence chains across data, models, and validation processes; reliability ensures stable, reproducible, and uncertainty-aware performance; and governability requires AI systems to remain measurable, verifiable, and controllable throughout their life cycle. Together, the CTRG framework defines the essential conditions under which trustworthy AI can be established (Figure 1). However, defining trust requirements alone does not establish trust. What is missing is a shared basis for comparison and verification. We therefore introduce the concept of a common ruler. The common ruler is not a benchmark, performance metric, or regulatory requirement. Rather, it is a measurement infrastructure designed to preserve CTRG across heterogeneous systems and deployment contexts, thereby making trust objectively measurable.Figure 1A common ruler framework for trustworthy AIThe absence of shared references creates a trust gap that constrains the scaling and deployment of trustworthy AI. Inspired by the role of the International System of Units (SI) in enabling globally consistent measurement, the framework introduces SI-like references as common trust anchors for comparability, traceability, reliability, and governability (CTRG). Through a two-layer architecture that links universal trust foundations (L1) with domain-specific reference systems (L2), these properties are further translated into reference, measurement, validation, and governance functions, collectively forming a common ruler for trust-preserving AI across domains. A common ruler framework for trustworthy AI The absence of shared references creates a trust gap that constrains the scaling and deployment of trustworthy AI. Inspired by the role of the International System of Units (SI) in enabling globally consistent measurement, the framework introduces SI-like references as common trust anchors for comparability, traceability, reliability, and governability (CTRG). Through a two-layer architecture that links universal trust foundations (L1) with domain-specific reference systems (L2), these properties are further translated into reference, measurement, validation, and governance functions, collectively forming a common ruler for trust-preserving AI across domains. To operationalize the common ruler, we propose a two-layer architecture (Figure 1). The first layer establishes universal trust foundations shared across AI systems, including evaluation standards, validation procedures, uncertainty representation, monitoring mechanisms, and auditing requirements. Their purpose is to preserve trust requirements across heterogeneous contexts and provide stable conditions supporting trustworthy AI. However, universal trust foundations alone cannot determine whether educational AI genuinely improves learning outcomes, whether engineering AI remains within physical constraints, or whether scientific AI captures signal rather than noise. Trust ultimately requires objective validation within domain-specific contexts. The second layer therefore consists of domain-specific reference systems. Built upon universal foundations, these systems adapt references, validation logic, and constraints according to domain requirements, translating shared trust requirements into context-specific measurement conditions. Together, these two layers operationalize the common ruler: universal trust foundations preserve shared requirements across AI systems, whereas domain-specific reference systems translate them into measurable validation conditions. To illustrate this principle, we consider three representative AI domains: society, engineering, and science (Figure 1). In society, trust relies on normative references grounded in human values and institutional accountability. Social AI cannot be evaluated solely through efficiency, accuracy, or task completion; it must also be assessed against societal values, fairness, accountability, and long-term outcomes. Normative references translate these principles into governance practice through evaluation criteria, institutional oversight, and accountability mechanisms. They define what should be assessed, including inclusivity, contestability, privacy protection, cognitive integrity, and human development. They further establish governance pathways supporting oversight, auditability, contestation, and redress. In educational AI, for example, trustworthy evaluation extends beyond narrow performance metrics toward reasoning capability, autonomy, and long-term intellectual development.1 Engineering AI requires references grounded in physical laws and operational constraints. Prediction accuracy, computational efficiency, and apparent plausibility alone are insufficient if AI outputs violate physical validity, safety and robustness requirements, or engineering tolerances. Physical references preserve trust through validated evaluation criteria, benchmark environments, life cycle validation, and deployment certification. These systems establish whether models remain consistent with conservation laws, robust under perturbation, reproducible across operational environments, and reliable during the simulation-to-reality transition. Such requirements become particularly critical in intelligent manufacturing, inverse design, and process optimization, where AI-generated outputs may appear plausible while remaining inconsistent with physical laws or experimental reality.2 Scientific AI requires references that support traceability, uncertainty characterization, reproducibility, and comparability. Scientific models cannot be evaluated solely through benchmark performance or predictive accuracy; they must remain grounded in experimentally verifiable evidence and reliable measurement. Measurement-based references operationalize these requirements through reference materials, calibration chains, metadata standards, uncertainty quantification, and cross-platform validation. These elements enable reproducibility across laboratories, comparability across datasets, and interpretability across analytical conditions. Emerging implementations, including shared multi-omics references, SI-anchored transcriptomic quantification, and standardized proteomic measurement, illustrate how reference systems can strengthen comparability, traceability, and reproducibility.3,4,5 Such capabilities are important in genomics, proteomics, and measurement science, where AI increasingly contributes to scientific knowledge generation. Taken together, the common ruler functions not as a universal metric but as adaptive measurement infrastructure that integrates shared trust requirements with domain-specific reference systems. Trustworthy AI cannot rely solely on principles, governance rules, or isolated evaluation procedures to remain sustainable at scale. As AI evolves across domains, deployment settings, and lifecycle transitions, preserving trust itself becomes an infrastructure challenge. CTRG defines the core properties of trustworthy AI, operational architectures translate these requirements into practice, and domain-specific reference systems establish conditions under which trust becomes objectively measurable. Overall, these components establish the common ruler as an operational infrastructure that transforms trust from abstract principles into coordinated system capability (Figure 1). At a broader level, this integration parallels mature national quality infrastructure (NQI) systems. In scientific and industrial systems, reliability rarely emerges from isolated standards, one-time validation, or independent certification procedures. Instead, objective trust depends on coordinated measurement foundations that integrate references, validation mechanisms, and governance capability across systems, environments, and time. Trustworthy AI faces a similar demand, but under more dynamic conditions: models update continuously, deployment environments shift, and domain requirements remain heterogeneous. From this perspective, the common ruler represents an initial NQI-like pathway for preserving trust throughout AI system evolution. It does not replace existing benchmarks, audits, or governance mechanisms; instead, it provides the measurement basis that allows them to operate coherently across heterogeneous AI systems. As AI becomes more powerful and more widely deployed, the central question will increasingly shift from what AI can do toward whether AI can remain trustworthy at scale. Performance without comparability remains difficult to validate, innovation without traceability remains difficult to verify, and deployment without governability remains difficult to sustain. Future AI competition will therefore extend beyond algorithms, data, and computing power toward building reference systems, validation infrastructures, and governance mechanisms capable of preserving objective trust across domains. Institutions, industries, and nations that build trustworthy infrastructures will increasingly shape AI development. The common ruler therefore represents more than an evaluation approach; it represents an infrastructure pathway toward making trust measurable, verifiable, sustainable, and scalable. In summary, as AI scales across domains, trustworthy AI increasingly depends not only on stronger models but also on trust-preserving measurement foundations. We propose the common ruler as a measurement infrastructure pathway that integrates CTRG, operational architectures, and domain-specific references to make trust objectively measurable, comparable, and sustainable across domains. Unlike conventional NQI frameworks that typically rely on relatively stable references and validation conditions, AI systems evolve continuously as models update, deployment environments shift, and domain requirements remain heterogeneous. Trust preservation therefore cannot rely solely on fixed evaluation procedures, static reference conditions, or one-time validation. Future trustworthy AI increasingly requires adaptive trust infrastructures in which operational foundations remain stable while reference systems, uncertainty evaluation, and validation mechanisms evolve continuously across scenarios and system lifecycles. Building such infrastructures may become essential for enabling trust to scale alongside intelligence itself. Funding and acknowledgments This work was supported by the 10.13039/501100012166National Key R&D Program of China (grant no. 2022YFF0608404), the Science & Technology Fundamental Resources Investigation Program (grant no. 2022FY101200), and the 10.13039/501100012164National High-Level Talent Development Program in Metrology (grant no. JLQN2602-P). Declaration of interests The authors declare no competing interests.