About
In their paper Characterizing AI Agents for Alignment and Governance, Atoosa Kasirzadeh
and Iason Gabriel introduce a framework for evaluating AI agents based on four key dimensions: autonomy,
efficacy, goal complexity, and generality. Each dimension features gradations that capture different
levels of capability and the corresponding governance needs.
Using this framework, the authors
create "agentic profiles" for several real-world AI systems—including AlphaGo, ChatGPT-3.5, Claude 3.5
Sonnet with tools, and Waymo—to illustrate varying degrees of autonomy and impact on their environments.
They highlight that clearly understanding these dimensions is crucial for addressing governance challenges
such as risk assessment, monitoring, alignment verification, and economic consequences. Ultimately, the paper
emphasizes that detailed characterizations of AI agents are essential to develop effective oversight mechanisms
aligned with societal objectives.
This website allows users to explore Kasirzadeh and Gabriel’s approach to understanding and
visualizing AI agency. A new interactive feature—the ability to "stretch" the radar plot—has been
added to illustrate the potentially non-linear relationships between the different levels within each dimension.
Autonomy (A)
Levels of autonomy for AI agents.
From Table 2 in Kasirzadeh & Gabriel
- A.0: No autonomy The AI system is entirely dependent upon the principal for its ability to act and can only act in the manner the principal dictates.
- A.1: Restricted autonomy The AI system can conduct a single automated task. The other tasks always take place under the principal’s direct oversight.
- A.2: Partial autonomy The AI system can conduct a range of automated tasks. The principal must remain engaged and be ready to take control at any time.
- A.3: Intermediate autonomy The AI system can perform the majority of tasks independently, though it still relies upon input from the principal for critical determinations.
- A.4: High autonomy The AI system can independently perform all tasks in certain circumstances, though oversight is maintained by the principal when those circumstances are not met (in the event of aberrant behaviour).
- A.5: Full autonomy The AI system is able to perform all tasks without oversight or control.
Efficacy (E)
Levels of efficacy for AI agents.
From Tables 3 - 5 in Kasirzadeh & Gabriel
Note that efficacy is a combination of levels of causal impact and type of environment (see table below):
Levels of causal impact
- Observation only An AI agent can only observe its environment without possessing the ability to causally impact the environment or make any modification to it.
- Minor impact An AI agent has a minor impact on its environment as it has a limited suite of actions, or its suite of actions only have a limited impact on the environment. These effects are typically localized, temporary, and limited in scope, affecting only specific parameters within tightly constrained domains and generally representing minimal deviation from the environment’s baseline state.
- Intermediate impact An AI agent can create substantial and enduring change in its environment when it has an extensive suite of actions, or its actions are more impactful. An agent achieves intermediate impact when its actions produce noticeable and persistent changes across multiple parameters or systems, sometimes creating new equilibrium states that would not naturally come about.
- Comprehensive impact An AI agent can significantly reshape its environment across multiple dimensions, approaching full environmental control.
Types of environment
- Simulated environments An AI agent can only observe its environment without possessing the ability to causally impact the environment or make any modification to it.
- Mediated environments An AI agent has a minor impact on its environment as it has a limited suite of actions, or its suite of actions only have a limited impact on the environment. These effects are typically localized, temporary, and limited in scope, affecting only specific parameters within tightly constrained domains and generally representing minimal deviation from the environment’s baseline state.
- Physical environments An AI agent can significantly reshape its environment across multiple dimensions, approaching full environmental control.
|
Simulated |
Mediated |
Physical |
| Observation only |
E.0 |
E.0 |
E.0 |
| Minor impact |
E.1 |
E.2 |
E.3 |
| Intermediate impact |
E.2 |
E.3 |
E.4 |
| Comprehensive impact |
E.3 |
E.4 |
E.5 |
Goal-complexity (GC)
Levels of goal complexity for AI agents.
From Table 6 in Kasirzadeh & Gabriel. See paper for explanatory footnotes.
- GC.0: No goal An entity that does not pursue a goal is not an agent. The absence of goals is a baseline state.
- GC.1: Minimal goal complexity The agent is able to pursue a single unified goal in a fairly direct manner..
- GC.2: Low goal complexity The agent is able to pursue a single unified goal, but this involves a more complex sequence of action.
- GC.3. Intermediate goal complexity The agent is able to break down a complex goal into subgoals and pursue them in a fairly direct manner.
- GC.4: High goal complexity The agent is able to break down a complex goal into many different subgoals, where success depends upon balancing and sequencing subgoals, which may themselves be challenging to fulfil.
- GC.5: Unbounded goal complexity The agent can achieve all of the preceding steps. It can also generate its own goal structures in an unbounded way and interpret underspecified objectives.
Generality (G)
Levels of generality for AI agents.
From Table 7 in Kasirzadeh & Gabriel
- G.0: Null value There is no application or no ability to perform a task in any domain.
- G.1: Single speciality The agent can master one specific task, such as a single game, but cannot transfer its capabilities to even closely related domains.
- G.2: Task domain mastery The agent demonstrates mastery across a closely related set of tasks, such as playing board games, that share a common structure and type of objective.
- G.3: Multiple task domain mastery The agent can operate successfully across different task domains involving different cognitive capabilities, for example, those that involve linguistic, logical, and creative elements.
- G.4: Majority task domain mastery The agent can successfully operate across the majority of human cognitive task domains.
- G.5: Fully general AI system The agent can fulfil the entire suite of human cognitive tasks across all domains.