Traditional regulatory frameworks often rely on binary triggers: an activity is either permitted or prohibited, and a system is either "high-risk" or "low-risk." However, artificial intelligence challenges this binary paradigm. Because modern AI models are general-purpose and highly dynamic, a single model can be deployed across vastly different contexts—ranging from a benign educational recommendation engine to a high-stakes clinical diagnostic tool.
Treating AI with rigid, categorical buckets creates what is known as the "profiling paradox": it subjects harmless personalization to the same crushing compliance burdens as malicious societal manipulation. To resolve this, a new paradigm has emerged around graduated, multi-axis risk classification. Instead of regulating the underlying technology itself, this approach dynamically scales compliance obligations by mapping an AI system's technical capabilities against its specific context of deployment.
The Mechanics of Multi-Axis Risk Mapping
Why does a risk-based classification approach work where static technology regulations fail? The core mechanism relies on decoupling the AI model from its application environment, evaluating risk as an intersection of multiple distinct dimensions rather than a single score.
High Criticality (e.g., Healthcare, Infrastructure)
│
│ ▲ Highly Regulated
│ │ (Strict Documentation & Auditing)
│
├────────────────────────►
│
│ ▲ Minimally Regulated
│ │ (Basic Transparency)
│
└───────────────────────────────────────────►
Low Capability High Capability
(Static Rules) │ (Adaptive GenAI)
▼ Low Criticality (e.g., Entertainment)
Rather than looking at a model in isolation, effective governance frameworks utilize a dual-axis or multi-axis matrix:
- The Capability Axis: Evaluates the internal reasoning, autonomy, and adaptability of the model. For instance, Wasi et al. demonstrate that in highly sensitive domains like biodesign, governance must differentiate tools based on their specific generation capabilities to prevent dangerous proliferation without stifling open research.
- The Criticality Axis: Evaluates the external domain and the potential for physical or societal harm. As Imperial et al. demonstrate, mapping internal reasoning capabilities against external domain criticality allows regulators to categorize systems into precise compliance tiers.
This intersection determines the final safety requirements. For example, Haduong et al. show that an AI system's safety profile is shaped by the intersection of physical harm potential and the user's level of expertise. A high-risk task executed by a non-expert requires significantly more robust error handling and transparency than the same task managed by a specialist. By shifting the focus to the criticality of the decision process rather than the specific software architecture, this strategy ensures that oversight remains technology-agnostic, as argued by Mökander et al..
From Binary Tiers to Graduated Influence
A key insight of this research direction is that risk is a spectrum, meaning regulatory responses must be similarly graduated. Nannini proposes a "graduated influence" doctrine to replace binary "high-risk" labels. Instead of treating all recommender systems identically, a graduated framework scales oversight based on the specific nature of the algorithm's influence over the user—distinguishing low-stakes personal convenience from high-stakes algorithmic gatekeeping.
When an AI system transitions up this risk spectrum, its technical and operational obligations scale proportionally:
- Low-Risk Tiers: Require minimal constraints, primarily focusing on basic user transparency (Hupont et al.).
- Moderate-Risk Tiers: Trigger structured internal risk assessments—akin to institutional review boards—to catch systemic harms before deployment, a mechanism advocated by Eaton et al..
- High-Risk Tiers: Mandate extensive technical documentation, rigorous bias mitigation, and external explainability audits (Balayn et al.).
This architectural shift mirrors historical legal evolutions outside of AI. As Bollweg et al. observe, modern tort law frequently replaces binary fault requirements with a spectrum that transitions to strict liability for high-risk operations. In AI governance, this graduated approach achieves significantly higher coverage of actual operational risks than vague, high-level ethical principles (Luna et al.).
Preventing Evasion and Ensuring Proportionality
For a risk-based classification system to work in practice, it must overcome two critical challenges: regulatory evasion and developer burnout.
Oversight-Independent Classification
A common loophole in early AI governance was the "human-in-the-loop" exemption, where developers claimed a system was low-risk simply because a human operator was present to sign off on decisions. However, this ignores the reality of automation bias—the human tendency to blindly trust algorithmic outputs.
To close this loophole, Da Silva et al. introduce oversight-independent risk classification. This mechanism determines an AI tool's risk category purely by its clinical or operational function, preventing developers from artificially lowering their risk scores by adding superficial human checkpoints.
The deployment architecture itself also dictates the risk tier. As Wernick et al. point out, providing a raw model as a research artifact carries different compliance weights than integrating that same model into a functional web demo or API, which instantly escalates its regulatory status.
Operationalizing Proportionality
If every AI application faces the same exhaustive auditing requirements, innovation stalls under the weight of "death by a thousand benchmarks." To prevent this, Mougan et al. propose a technical framework that decomposes the legal concept of proportionality into three operational pillars:
$$\text{Proportionality} = f(\text{Suitability}, \, \text{Necessity}, \, \text{Balance})$$
- Suitability: Does the evaluation provide meaningful safety information?
- Necessity: Is there a less intrusive or less compute-heavy testing alternative?
- Balance: Does the informational gain outweigh the developer's resource burden?
By optimizing along this Pareto-frontier, the intensity of validation is strictly calibrated to the application's risk profile.
We see this tailored calibration successfully applied in domain-specific settings. In healthcare, the authors of a prominent medical device study highlight how horizontal AI regulations can successfully inherit strict vertical standards based on clinical seriousness (Qingfeng et al.; Busch et al.). Within these high-stakes domains, granular grading systems are used to classify specific model behaviors—such as distinguishing "Major" hallucinations that threaten patient care from "Minor" documentation errors (Asgari).
Similarly, in autonomous cybersecurity applications, Jones et al. demonstrate that agents can use a discrete tri-color taxonomy (Red, Yellow, Green) to dynamically govern their own refusal behaviors based on the critical infrastructure status of their target.
Ultimately, risk-based regulatory classification works because it acknowledges that AI safety is not an intrinsic property of a model, but a function of how and where that model is used. By formalizing this context into multi-axis matrices and graduated compliance tiers, the framework provides a predictable, scalable path for trustworthy AI deployment.
Key Foundational Work
Review of the Current State of UAV Regulations (2017) — contributed 2 concept elements to this direction.
Das neue chinesische Haftpflichtgesetz (2011) — contributed 1 concept elements to this direction.