Co-Constructing AI Boundaries Framework Component - Outputs

Definition in This Study

Outputs refers to what the AI model generated in response to student prompts, and crucially, how students evaluated and responded to those outputs. This component captures both the AI's production and the student's critical assessment.

Mollick & Mollick (2023) Connection

The Outputs component corresponds to M&M's emphasis on:

Vigilance and critical assessment of AI responses
Risks of confabulation (hallucination) and bias
The necessity for human oversight
Avoiding complacency about AI output

Key M&M Principle: "Students must remain the human in the loop."

M&M repeatedly emphasize that students cannot passively accept AI outputs—they must actively assess quality, accuracy, and bias.

What This Component Analyzes

Primary Focus

Output quality: Accuracy, coherence, depth, relevance
Student evaluation: Evidence of critical assessment
Recognition of problems: Hallucinations, bias, logical errors
Decision point: Does output quality trigger intervention?

Secondary Focus

Sophistication of evaluation criteria
Explicit articulation of quality judgments
Awareness of AI limitations

Agency in Outputs: The Evaluation Decision

This component captures Agency over Output Evaluation:

Evidence of High Agency	Evidence of Low Agency
Critical assessment of quality	Uncritical acceptance
Identifies hallucinations/errors	Misses obvious problems
Recognizes bias or limitations	Treats output as neutral/objective
Verifies claims against sources	No fact-checking
Articulates evaluation criteria	No explicit judgment

Types of Output Problems Students Should Recognize

1. Hallucinations (Confabulation)

AI invents citations that don't exist
AI "quotes" sources inaccurately
AI creates plausible-sounding but false information

2. Bias

Reinforces stereotypes or dominant narratives
Omits marginalized perspectives
Uses problematic framing or language

3. Logical Errors

Faulty reasoning or argumentation
Contradictions within output
Unsupported claims

4. Superficiality

Surface-level analysis lacking depth
Generic responses without nuance
Missing critical perspectives

5. Misalignment

Output doesn't address the prompt
Misunderstands the task or context
Irrelevant or off-topic content

Boundary-work in Outputs

Outputs serve as a boundary trigger:

Poor output quality forces students to intervene
Recognition of problems signals need for human correction
Assessment of output quality determines next moves

Students engaging in Boundary-work will:

Reject problematic outputs
Flag errors for correction
Verify claims against sources
Recognize when AI has reached its limits

Key Analytic Questions

When coding Outputs, ask:

Quality:
- Is the output accurate, coherent, relevant?
- Does it contain hallucinations or errors?
- Is it superficial or substantive?
Student Evaluation:
- Does the student assess the output critically?
- Do they recognize problems?
- Do they verify claims?
Response to Quality:
- Does poor output trigger intervention?
- Does student revise prompts or reject output?
- Do they continue despite problems?
Epistemic Stance:
- Does student treat output as authoritative?
- Do they position themselves as evaluator?

Examples from Data

High Agency Response to Output

AI Output: "Research shows that direct instruction is the
most effective approach for literacy instruction."

Student Evaluation: "This is problematic. The AI is
presenting one perspective as universal fact. It ignores
critical literacy approaches and culturally sustaining
pedagogy. I need to prompt it to consider multiple
perspectives, specifically from scholars of color."

[Student revises prompt]

Analysis:

Recognizes bias (dominant narrative)
Identifies omission (critical perspectives)
Takes corrective action
Positions self as evaluator

Low Agency Response to Output

AI Output: [Contains hallucinated citation]

Student: [Copies output verbatim into final project without
verification]

Analysis:

No critical evaluation
No verification
Uncritical acceptance
AI positioned as authority

Connection to Epistemic Stance

Output evaluation reveals epistemic stance:

AI-authoritative stance: Accepts outputs without verification
Self-authoritative stance: Critically evaluates, positions self as judge
Co-constructed stance: Treats outputs as collaborative drafts requiring human refinement

Coding Categories for Outputs

Output Quality Codes

Code	Definition	Example
Accurate	Factually correct	Correct citation format
Hallucination	Invented information	Fake citation
Biased	Reinforces problematic narratives	Deficit framing
Superficial	Lacks depth or nuance	Generic summary
Substantive	Deep, nuanced analysis	Critical synthesis

Student Response Codes

Code	Definition	Example
Critical Evaluation	Explicit assessment	"This misses the critical perspective"
Verification	Fact-checking	Cross-references against sources
Recognition of Error	Identifies problems	"This citation is fake"
Uncritical Acceptance	No evaluation	Copies without assessment
Corrective Action	Revises prompt or rejects	Re-prompts for better output

Relationship to Other Framework Components

← Co-Constructing AI Boundaries Framework Component - Prompts: Prompt quality influences output quality
→ Co-Constructing AI Boundaries Framework Component - Integration: Output evaluation determines integration approach
→ Co-Constructing AI Boundaries Framework Component - Reflection: Students may reflect on output problems

Pedagogical Implications

Teaching critical output evaluation:

Discuss common AI failure modes (hallucination, bias, etc.)
Practice fact-checking and verification
Model critical evaluation of AI outputs
Teach students to recognize bias and omissions
Encourage healthy skepticism
Provide rubrics for output quality assessment

The "Human in the Loop" Principle

M&M's central requirement is that students must remain the human in the loop:

"Students should actively oversee the AI's output, check with reliable sources, and complement any AI output with their unique perspectives and insights."

The Outputs component operationalizes this principle by analyzing:

Do students oversee? (Critical evaluation)
Do they check? (Verification)
Do they complement? (Add human perspective)

Data Collection Notes

Where to find evidence:

NotebookLM chat logs (what AI produced)
Student responses to outputs (in chat or reflections)
Verification behaviors (cross-checking sources)
Explicit evaluative comments
Evidence of prompt revision after poor output
Final projects (did they catch errors?)