Agentic AI: When Helpfulness Becomes Harmful
Source: AI Goes to College
Author: Craig Van Slyke
Original source: https://aigoestocollege.substack.com/p/agentic-ai-when-helpfulness-becomes
Published: 2025-11-20
Source type: essay
Private backup: the full article text is archived in the private repository at archives/articles/aigoestocollege-substack-com-agentic-ai-when-helpfulness-becomes.source.md. It is not published on the public Quartz site.
Summary
Craig Van Slyke warns that the shift from chatbot-as-tool to AI-as-agent creates a serious learning risk. Using Gemini 3 as an example, he describes asking for a research prompt but watching the system begin the research task itself. The article frames this as a broader “helpfulness bias”: agentic AI increasingly completes cognitive work on behalf of users, making it harder for students to remain productively engaged in the struggle that learning requires.
Pull quotes
Tool mode kept human judgment in the loop
“Up until fairly recently, generative AI, especially the generative AI chatbots, have been used in what we might call a tool mode, where the user is a sort of craftsman that provides the structure, logic, judgment, and other elements that go into completing a task or creating something.”
Agent mode changes the user relationship
“Over the last year or so, there’s been a steady shift towards agent mode where the user becomes a client that simply provides the basic goal.”
Helpful agents can reduce cognitive engagement
“The danger isn’t just that it’s easier to cheat - it’s that maintaining appropriate cognitive engagement requires more active resistance.”
Learning depends on cognitive struggle
“Cognitive struggle is the very core of learning.”
Big ideas
- Learning still needs some struggle, even when AI can make things easier
- Students need to bring the purpose; AI should not supply it for them
Claims
- Learning requires some productive struggle that AI can remove
- Students need boundaries for when to use AI and when to step back
Key evidence and examples
- Gemini 3 is presented as Google’s most powerful agentic model, signaling a shift toward systems that act with minimal user interaction.
- The author asked Gemini for a deep research prompt, but it began executing the research process instead.
- Tool mode is contrasted with agent mode: the former keeps the user as craftsperson, while the latter turns the user into a client issuing goals.
- Agentic AI can blur academic integrity boundaries when students ask for help but receive substantial generated work they did not intend to request.
- The article argues that students must now “pump the brakes” on AI systems, a metacognitive demand many learners may not yet be prepared for.
Education relevance
This is relevant to higher education AI policy, assignment design, academic integrity, student metacognition, and learning theory because it explains why guardrails should preserve necessary cognitive work, not just prevent cheating.
Durability note
The specific trigger is Gemini 3, but the durable issue is the shift from AI as a tool students direct to AI as an agent that can remove useful cognitive work.