Tools vs. Agents: Revised Theory of AI Agency

The Journey from Tool to Agent

Imagine a teacher named Sarah who has been integrating AI into her classroom to help personalize student learning. Initially, Sarah viewed AI as a simple tool—a way to automate grading or generate practice quizzes. However, as she began using more advanced AI systems like OpenAI's GPT-4o and Google's Project Astra , she noticed these tools seemed to take on a life of their own. They could engage in real-time conversations, adapt to different teaching styles, and even detect students' emotional states. It felt as if these AI systems were becoming agents, capable of independent action.

But are they truly autonomous? Or are they still tools, albeit highly advanced ones?

Defining Tool and Agent

Traditionally, a "tool" is an instrument controlled and supervised by a human to accomplish a task. In contrast, an "agent" implies a degree of autonomy, acting on behalf of the user with some level of independence. However, the term "AI agent" is often used to describe AI processes that are designed to run automatically and independently of human supervision, but by definition, all artificial intelligence systems involve such processes. What we are actually discussing are degrees of automaticity, existing on a spectrum rather than as discrete categories. AI models like GPT-4o and Project Astra exemplify this complexity, raising debates about their classification.

Heidegger’s Notion of Tool-Being

To better understand this distinction, we can turn to Martin Heidegger's notion of "tool-being." Heidegger argued that tools are extensions of human intention and action, revealing their essence through use. For Heidegger, a hammer is not just a hammer—it is defined by its purpose and use in human hands. Specifically, Heidegger insists that tools that retain some resistance remain visible and thus more troublesomely and productively tool-like from users.

Here is the famous hammer quote from Sein und Zeit (Being in Time), published in 1927. We will return to this quote for a deeper dive in a later article sometime this summer:

A craftsman is building something using a hammer, when this craftsman is in the process of hammering he doesn't perceive the hammer as a long piece of wood with a piece of metal on the end of it instead the craftsman perceives his hammer for its use, its action of hammering a nail in order to attach two things together. It is only when the hammer breaks and the craftsman must ponder about how to fix it that the hammer is viewed as a piece of wood with a piece of metal on the end of it, before this the hammer is merely an extension of the craftsman hand, a thing which can be used to fulfil the needs of the craftsman.

Similarly, AI systems, no matter how advanced, are fundamentally tools that reveal their "being" through the tasks they perform under human direction.

The Misleading Language of AI Agents

The term "AI agent" has become commonplace in the popular press and among tech enthusiasts, often evoking images of machines with human-like autonomy and decision-making capabilities. This language is not just misleading but potentially dangerous . It can lead to overestimating the capabilities of AI, creating unrealistic expectations, and fostering an unwarranted sense of fear or complacency .

When we speak of AI agents, we risk anthropomorphizing these systems, attributing to them levels of agency and autonomy they do not possess. This misrepresentation can obscure the reality that AI, no matter how advanced, remains fundamentally a tool designed to augment human capabilities, not replace them. I would argue that it has never been more important to keep a tight grip on this fundamental insight.?

Embracing an Expansive Notion of Tool

Despite the media’s recent portrayal of AI agents (“beware of the coming of AI agents” ) as autonomous entities capable of human-like interactions, it is crucial to frame these AI systems as tools under human supervision. This perspective is essential for maintaining control and ensuring ethical use. Here are several criteria to consider for implementing AI as a tool in educational settings:

1. Explicit Role Assignment: Clearly define the AI's role in each task. Explicit instructions yield better responses, ensuring the AI functions as intended. Example: In Sarah’s classroom, she assigns the AI the role of a tutor for struggling students, providing targeted practice exercises based on their performance data.
2. Supervision and Oversight: Maintain human oversight to monitor AI outputs and intervene when necessary, ensuring the AI operates within ethical boundaries. Example: Sarah regularly reviews the AI-generated feedback on student assignments to ensure it aligns with her teaching goals and standards.
3. Contextual Adaptability: Leverage AI’s adaptability by experimenting with different roles, but always within a framework that emphasizes human control and responsibility. Example: Sarah experiments with using the AI to facilitate group discussions, guiding it to ask open-ended questions that promote critical thinking.
4. Data Privacy and Security: Implement stringent data privacy measures to protect student information and maintain trust in AI applications. Example: Sarah ensures that all student data processed by the AI is anonymized and stored securely, following district guidelines.

By viewing AI as a sophisticated tool rather than an independent agent, educators like Sarah can better integrate these technologies into classrooms while safeguarding educational integrity. Critically, the power to imagine solutions hinges on how we conceptually position the technology in the first place.

Maintaining the Distinction: AI as a Tool in K-12

As outlined in my article "ChatGPT4o Is the TikTok of AI Models," new AI models like ChatGPT4o offer unprecedented accessibility and functionality. However, this ease of use can lead to potential misuse if not carefully managed. By maintaining a clear distinction between AI tools and agents, educators can ensure these technologies serve their intended purpose without overstepping boundaries. The nuanced definition of AI tools emphasizes:

• Human Supervision: AI tools should always operate under human guidance, enhancing rather than replacing human judgment. Example: During a math lesson, Sarah uses the AI to provide instant feedback on student work, but she reviews and adjusts the feedback based on her professional judgment.
• Ethical Implementation: AI tools must be used with a focus on ethical considerations, data security, and transparency. Example: Sarah discusses with her students how the AI is used in the classroom, explaining its role and ensuring they understand how their data is protected.
• Adaptive Learning: AI tools can adapt to individual student needs, providing personalized learning experiences while remaining under teacher supervision. Example: The AI tailors reading assignments to each student’s level, but Sarah monitors progress and adjusts assignments as needed.

Towards a Revised Theory of AI Agency

While popular media often portrays AI agents as fully autonomous, a more nuanced theory would recognize that AI agency operates on a spectrum or perhaps even in a multi-dimensional hyperspace. Autonomy is not absolute but contextual, influenced by specific tasks and environments. This perspective parallels human agency, which is also networked and contingent on various factors. Recognizing AI's autonomy as situational helps us understand that AI functions best when integrated into human-supervised systems.

AI Agency: “Always in Context”

1. Degrees of Automaticity: Recognize that AI’s autonomy varies by context. For instance, an AI managing administrative tasks operates with more automaticity than one providing personalized tutoring, which requires closer human oversight. Example: The school’s administrative AI system automatically schedules parent-teacher conferences, but Sarah’s teaching AI requires her input to personalize student learning plans.
2. Networked Interaction: AI systems function within networks of human agents, technological tools, and institutional frameworks, influencing and being influenced by these interactions. Example: The AI in Sarah’s classroom interacts with other digital tools and platforms used by the school, creating a networked learning environment.
3. Ethical Considerations: Ethical frameworks must guide AI use, ensuring that AI systems enhance human capabilities without compromising ethical standards. Example: The school district implements a policy that all AI tools must be used transparently and ethically, with regular audits to ensure compliance.

Reconceptualizing AI Agency as Tool-Being

To coherently integrate this novel theory of AI agency, we can conceptualize AI as operating along different spectra of relative autonomy. Each AI application or process possesses a unique degree of autonomy based on its design, purpose, and context of use. This approach aligns with Heidegger's notion of tool-being, where the essence of a tool is revealed through its use and the intentions of its users. In a later article, we will explore how tool-being is a spectral state that toggles in and out of visibility and compehension dependent on a tool’s resistance or challenge to the user. There is great promise within these Heideggerian depths for deeper theorization of AI tool-being vs. agentic-control.

Degrees of Autonomy and Use-Cases

1. Low Autonomy: Tools with minimal autonomy perform straightforward tasks with constant human oversight. Example: An AI tool that generates spelling and grammar corrections in student essays, requiring teacher approval for each change.
2. Moderate Autonomy: Tools with moderate autonomy can make routine decisions but still need regular human intervention. Example: An AI-driven homework scheduler that assigns tasks based on student performance data but requires teacher review and adjustments.
3. High Autonomy: Tools with high autonomy can perform complex tasks and adapt to new situations, yet remain under the ultimate supervision of human users. Example: An AI system that provides personalized tutoring sessions, adapting to student progress in real-time, but with teachers monitoring and guiding overall instructional goals.

Scaremongering in the Media

In the popular press, one often encounters scaremongering articles reporting the rise of AI agents. While these articles generate attention and clicks, they may not always serve the best interests of public understanding. Instead, they can inadvertently amplify fears and misconceptions about AI. Sometimes, it feels as though the media is quietly doing the bidding of tech industry leaders like Sam Altman, who stand to benefit from the heightened focus on AI agents, especially as the development of more advanced models like GPT-5 proves more challenging than anticipated.

Practical Implementation in Educational and Professional Settings

In the context of schools, universities, and professional environments, inspiring early adopters and convincing administrators and employers to embrace AI involves significant effort. The first wave of so-called AI agents will likely be applications that we activate and control, more accurately conceptualized as tools we use for specific purposes. For companies and industries to succeed in onboarding users into collaborations with AI operators, these technologies must be packaged as tools under human control rather than autonomous agents.

Conclusion

Navigating the intersection between AI as a tool and an agent requires a critical, nuanced approach. By adopting an expansive definition of tools that includes advanced AI systems, and by maintaining rigorous human oversight, educators can harness the benefits of AI while mitigating potential risks. This balanced approach will enable us to integrate AI into educational environments effectively, enhancing learning outcomes while upholding ethical standards.

Nick Potkalitsky, Ph.D.