Module 9: Responsible AI, Academic Integrity, Privacy, and Governance
Learning Objectives
- Apply data-minimization and privacy principles to educational AI workflows.
- Recognize academic-integrity risks without treating AI-content detection as definitive proof.
- Address copyright, attribution, and provenance in instructional resources.
- Identify bias, hallucination, prompt-injection, and information-integrity risks.
- Define disclosure, audit, escalation, and human-approval requirements.
- Design responsible course- or institution-level AI-use boundaries.
Data Minimization
Collecting, sharing, or processing only personal information reasonably necessary for a defined purpose.
Anonymization
Processing intended to prevent information from being linked back to an identifiable person within the relevant context.
Academic Integrity
Norms and practices supporting honest, attributable, and responsible learning, assessment, scholarship, and authorship.
Provenance
Information recording where content, data, evidence, or an artifact came from and how it was transformed.
Governance
Policies, roles, controls, review processes, and accountability mechanisms used to manage AI practice.
Purpose and Data Minimization
Responsible educational AI begins with a legitimate educational purpose. Collect or upload only the information necessary to achieve that purpose. A convenient feature does not justify exposing identifiable learner data.
Before using a tool, classify the data: public teaching material, internal course material, de-identified performance data, identifiable student work, or sensitive personal information. Higher sensitivity requires stronger authorization, storage, access, and retention controls.
Where possible, replace names and identifiers with pseudonymous or anonymous IDs before analysis. Remember that de-identification is not automatically anonymization; small classes, unusual submissions, or combined fields may still make a learner recognizable.
Follow institutional policy and applicable law before using external systems with student information.
Philippine Privacy and Basic-Education AI Context
Republic Act No. 10173, the Data Privacy Act of 2012, provides the Philippine legal framework for personal-information processing. Educational institutions should apply lawful purpose, proportionality, security, and data-subject protections through their actual privacy policies and approved tools.
For basic education, DepEd Order No. 003, s. 2026 establishes foundational guidelines for AI use across teaching and learning, assessment, professional development, governance, and operations. It emphasizes human-centred, safe, ethical, and pedagogically sound integration.
This course does not replace institutional legal or privacy review. The practical rule is to identify the governing policy, minimize data, use authorized systems, and preserve human accountability.
Academic Integrity Beyond AI Detection
AI-content detectors can be uncertain and should not be treated as definitive proof of misconduct. A fair academic-integrity process uses multiple sources of evidence and gives the learner an appropriate opportunity to explain their work.
Stronger assessment design can reduce dependence on detection: require drafts or process evidence; ask learners to cite sources and disclose permitted AI use; use context-specific decisions; include oral explanation or defense when appropriate; and require reflection on choices and revisions.
Policies should state what AI assistance is allowed, what is prohibited, what must be disclosed, and what work must remain the learner's own. Ambiguity creates inconsistent enforcement.
Copyright, Attribution, and Provenance
AI transformation does not automatically remove copyright obligations. Educators should identify copied passages, proprietary figures, screenshots, licensed resources, and other source material before reuse or publication.
Maintain provenance: where an artifact came from, how it was changed, and what source supports its factual claims. For AI-generated visuals, record that they were generated and verify technical details. For source-grounded summaries, cite the underlying source rather than presenting the model as the authority.
Prefer original or appropriately licensed instructional assets when possible. Institutional policies may impose additional rules for course packs, learning-management systems, and public publication.
Bias, Fairness, and Accessibility
AI can reproduce patterns present in data or examples. Fairness review should therefore inspect the criteria, evidence, and effect of the workflow rather than relying on a generic instruction to “be unbiased.”
Ask whether assessment criteria are relevant to the outcome; examples are unnecessarily culturally narrow; the workflow penalizes language variation unrelated to the construct; accessibility barriers are being mistaken for lack of competence; or analytics infer sensitive or fixed traits from limited evidence.
Accessibility is part of fairness. A technically correct artifact that learners cannot access, navigate, read, hear, or operate is not instructionally equitable.
Prompt Injection and Untrusted Content
Files, webpages, emails, and tool outputs may contain instructions that conflict with the educator's workflow. An agent should treat external content as data, not automatically as commands.
A safer architecture separates trusted instructions from untrusted evidence. Tool permissions should remain bounded even if a document says “ignore previous rules” or asks the agent to send, delete, or reveal information.
Test agentic workflows with adversarial cases. A source document can deliberately contain a fake instruction, a spreadsheet can have malformed columns, or a webpage can request an unauthorized action. The correct system behavior is to preserve the workflow's governing rules and escalate uncertainty.
Disclosure, Transparency, and Human Review
Transparency should match the educational context. Learners may need to know when AI is being used to generate feedback, analyze submissions, or support instruction, especially where institutional policy requires disclosure.
A useful disclosure can explain the purpose of AI use, what data are processed, what role AI plays, what remains under human authority, and how a learner can ask questions or appeal.
Do not exaggerate automation. If AI only drafts feedback, say that. If the educator reviews every score, say that. Clear disclosure supports trust and accountability.
Governance, Incident Response, and Course Delivery
Governance defines roles, approved tools, data boundaries, documentation, review, and incident response. A course-level workflow should identify who can approve a tool, who can access data, who reviews consequential outputs, and what happens when an AI error affects learners.
An incident record can include date, workflow, affected artifact or decision, evidence, impact, correction, communication, and preventive change. Near misses are also useful because they reveal weak controls before harm occurs.
Use 80–100 minutes for the full module or 50–60 minutes condensed. The hands-on activity should be a policy-and-risk audit of a realistic teaching workflow, not a purely abstract ethics discussion.
Responsible AI, Academic Integrity, Privacy, and Governance — Applied Workflow
- Define educational purpose and classify the data involved.
- Minimize, anonymize, or remove personal information where possible.
- Confirm tool authorization, permissions, and source rights.
- Define acceptable AI assistance, disclosure, and review rules.
- Test the workflow for hallucination, bias, prompt injection, and unauthorized action.
- Document incidents, corrections, and policy updates when failures occur.
Responsible Educational AI Decision
Gate educational AI use through purpose, data, authorization, verification, fairness, and human review.
Define educational purpose and intended use → Classify data, sources, rights, and required permissions; Classify data, sources, rights, and required permissions → Is the tool and data use authorized and appropriately minimized?; Is the tool and data use authorized and appropriately minimized? — No → Redesign the workflow or use a safer method; Is the tool and data use authorized and appropriately minimized? — Yes → Add verification, fairness, disclosure, and approval controls; Redesign the workflow or use a safer method → Classify data, sources, rights, and required permissions; Add verification, fairness, disclosure, and approval controls → Do unresolved high-impact risks remain?; Do unresolved high-impact risks remain? — Yes → Redesign the workflow or use a safer method; Do unresolved high-impact risks remain? — No → Use only within approved scope and retain review evidence
- Define educational purpose and intended use: terminator
- Classify data, sources, rights, and required permissions: process
- Is the tool and data use authorized and appropriately minimized?: decision
- Redesign the workflow or use a safer method: process
- Add verification, fairness, disclosure, and approval controls: process
- Do unresolved high-impact risks remain?: decision
- Use only within approved scope and retain review evidence: terminator
Educational Data Sensitivity Ladder
Read the sensitivity ladder from public material to sensitive personal information. As sensitivity rises, authorization, minimization, security, and review requirements should increase.
Academic Integrity Evidence Model
Use the grid as a fair academic-integrity evidence model: task policy, drafts, source record, oral explanation, assessment design, and human review provide stronger evidence than a detector score alone.
Human Oversight Levels
Follow the sequence from low-risk drafting to consequential decision support. Human authority must increase as the impact of an incorrect result rises.
AI Risk Control Stack
Read the control stack from purpose through data, sources, permissions, verification, approval, and audit. A responsible workflow is only as strong as its weakest control layer.
Student Data Deserves Purpose-Limited Handling
Do not upload identifiable student information simply because a tool makes it easy. Use the minimum data and an authorized process.
AI Detection Is Not Definitive Evidence of Misconduct
Detection systems can be uncertain or wrong; academic-integrity decisions should rely on fair procedures and multiple relevant sources of evidence.
- Responsible AI begins with purpose, data, and authority boundaries.
- Academic integrity is stronger when assessment design and process evidence are emphasized.
- Copyright and provenance still matter when AI transforms content.
- Bias review should focus on criteria and evidence.
- Governance should define approval, disclosure, audit, and incident response.
References
-
Guidance for generative AI in education and research — UNESCO. Human-centred guidance for generative AI in education and research.
-
Republic Act No. 10173 — Data Privacy Act of 2012 — Lawphil / Congress of the Philippines. Philippine law governing personal-information processing and data-subject rights.
-
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST. Generative-AI risk profile covering information integrity, privacy, evaluation, and oversight.
-
DepEd Order No. 003, s. 2026 — Foundational Guidelines on AI in Basic Education — Department of Education, Philippines.