The term “AI Agent” has become increasingly prevalent in discussions about artificial intelligence, yet its meaning remains somewhat ambiguous. This ambiguity stems partly from different conceptualizations of agency across disciplines and languages. A recent LinkedIn discussion, sparked by Maximilian Seeth’s introduction to AI ethics, highlighted this complexity.

In Seeth’s blog post, he introduces Don Ihde’s abstraction of human-technology relationships:

1. (I – Technology) – World: Where we merge with technology, like looking through a microscope.
2. I – (Technology – World): Where technology mediates our perception of the world, like a thermostat showing outside temperature.
3. I – Technology – World: Where we confront technology as another being.

Building on this, I proposed that these abstractions offer insights beyond ethics, touching on the concept of agency. While a microscope doesn’t have agency, smartphone cameras already do at a rudimentary level (e.g., Google’s “capture the moment” vs. Apple’s more “faithful reproduction” approach). AI assistants, with their injected personalities, further blur these lines.

Seeth responded by suggesting that even a microscope might have a degree of agency in a rudimentary sense, pointing to actor-network theories (ANT) for understanding action within complex networks of beings and things.

ANT contributes a broader conception of agency, which extends beyond intentional actions to include the capacity of entities (human or non-human) to affect outcomes within a network. In this view, even a microscope possesses a form of agency: it shapes scientific practice by determining what can be observed, influences the posture and movements of the researcher, and affects the types of questions that can be asked. This perspective contrasts sharply with everyday understandings of agency where “Agency” typically implies intentionality. ANT’s view is closer to concepts like:

• circumstances, conditions
• tools, aids, means
• environment

These terms capture the passive, circumstantial nature of influences without implying active decision-making.

A spectrum of AI agency presented in a recent IBM workshop aligns with this more nuanced view, categorizing AI systems based on their capacity to influence outcomes:

1. Sidekick (Lowest Agency):
   • IBM definition: “Ad Hoc Task completion using LLM - IBMC Assistant”
   • Different examples: Google’s Sidekick for contextual prompts in Google Docs, OpenAI Canvas, Code completion
2. Assistant (Moderate Agency):
   • IBM definition: “(Repeatable) Process automation - Assistant”
   • Different examples: ChatGPT; OpenAI’s Assistants API, maintaining context and accessing tools
3. Twin / Agent (Highest Agency):
   • IBM definition: “Program automation – AI Agents”
   • Different example: Autonomous AI systems capable of complex planning and tool use

This spectrum can be viewed through Ihde’s abstractions of human-technology relationships:

1. Sidekicks primarily align with “I – (Technology – World)”, where the AI mediates our interaction with information. Here, we might “morph” with the tool, as Seeth suggests, integrating it seamlessly into our thought processes and workflows. This morphing is relatively benign due to the Sidekick’s rudimentary agency.

2. Assistants embody a fluid position, shifting between “I – (Technology – World)” and “(I – Technology) – World”. When mediating our interaction with data (like ChatGPT’s Code Interpreter analyzing a spreadsheet), they act as “I – (Technology – World)”. However, they can also function as “(I – Technology) – World” when we collaborate with them, e.g. merging our analytical capabilities with that of the AI. This fluidity raises intriguing questions about the boundaries of our agency and the AI’s capabilities.

3. Agents, at their most advanced, could potentially embody the “I – Technology – World” relationship, where we interact with them as distinct entities. Here, the concept of morphing becomes problematic. As I suggested in my LinkedIn comment, while morphing with tools of lower agency might be acceptable, we should be cautious about too closely integrating with Agents that possess high degrees of autonomy and decision-making capability.

My suggestion of saying “Thank you” to an AI assistant if, and only if, it amazes us reflects an attempt to maintain a clear “I – Technology – World” stance. This practice could help us remain mindful of the distinction between our agency and that of the AI, particularly as we interact with more advanced Agents.

In conclusion, the different shapes and forms of AI deployments (see also Miles Brundage’s distinction) invite us to reflect on technological agency. ANT suggests viewing all entities in a network, including technologies, as potential actors shaping outcomes. Ihde’s framework, on the other hand, advises us to consider different modes of human-technology relationships. As AI-enabled tools become more autonomous, these perspectives encourage us to carefully examine how we integrate with and relate to these technologies.

This reflection brings us to the initial confusion about the term “AI Agent”. By considering the spectrum of AI agency through the lenses of ANT and Ihde’s concepts, we can better appreciate why this term is ambiguous. It encompasses a range of technologies with varying degrees of agency, from Sidekicks we might comfortably morph with, to Agents whose autonomy might require us to maintain a more distinct relationship.