Responsible AI in a project is not a yes-or-no decision about a tool. It is a continuing decision about the learning purpose, the thinking students must still own, the authority educators retain, the data and access conditions around the activity, and what happens when the system is wrong or no longer appropriate.
The ten checks below are designed to be useful before a project begins and while it is under way. They do not certify a tool or replace school safeguarding, procurement, accessibility, data-protection, or legal review.
Start with the learning purpose, not the tool
UNESCO's Guidance for generative AI in education and research takes a human-centred approach to regulating generative AI. It proposes steps that include protecting data privacy and setting an age limit for independent conversations with GenAI platforms, and it proposes an age-appropriate approach to ethical validation and pedagogical design. It is global policy guidance, not proof that a particular use is safe, lawful, or educationally effective.1
That makes the first question educational rather than technical.
1. Purpose
Decision question: What learning problem is AI addressing?
Actionable check: Write down the intended learning outcome, the friction AI may reduce, and what success would look like without referring to the tool's features. Name the work students and educators must still do themselves.
Warning sign: The rationale begins with “we have access to this tool,” “students will use AI anyway,” or pressure to make the project look innovative. No specific learning need is named.
Project example: In a community-water project, an educator may use AI to generate alternative stakeholder complications during planning. Students still frame the problem, gather evidence, weigh trade-offs, and defend their response.
Protect student thinking from answer substitution
UNESCO's AI competency framework for students frames students as responsible users and co-creators of AI, emphasizing critical judgment, ethics, foundational knowledge, and inclusive design. It is a curriculum reference framework, not evidence that using AI automatically develops those capabilities.2
In one randomised experiment with 91 university students, those assigned to ChatGPT 3.5 rather than Google for an inquiry task reported lower cognitive load but showed lower-quality reasoning and argumentation in their recommendations.3 England's Department for Education product safety standards expect learner-facing products to start with hints rather than give final answers by default.4
2. Thinking
Decision question: Does the AI use protect or bypass student thinking?
Actionable check: Mark the reasoning, creation, judgment, and explanation students are expected to practise or demonstrate. Configure support around questions, hints, explanation, critique, reflection, and references back to evidence.
Warning sign: The system produces the argument, analysis, solution, code, design, or conclusion that the student was meant to develop.
Project example: A student preparing a recommendation for a local council receives a prompt to compare two pieces of evidence and explain a trade-off. The system does not write the final recommendation.
Dwyl.studio's position is that student-facing AI should not complete the thinking or learning work a student is expected to practise or demonstrate. That is more precise than saying AI should never answer a factual question. A definition or explanation may be appropriate when it is not the learning task; producing the student's reasoning is a different role.
Set disclosure, oversight, and recourse expectations
UNICEF's Guidance on AI and children includes transparency, explainability, accountability, safety, privacy, fairness, well-being, and inclusion among ten requirements for child-centred AI. Version 3.0 is broad global guidance that requires local adaptation; it does not certify a product or settle a school's legal duties.5
Australia's national schools framework includes related guiding statements on disclosure, human responsibility, monitoring, and contestability. It is aspirational, not evidence of safe or effective use.6
3. Disclosure
Decision question: Do educators and students understand when and how AI is involved?
Actionable check: Explain what the system does, what information it uses, where uncertainty remains, and when AI involvement should be disclosed in student work or an educator's decision. Make the explanation appropriate to the students' age and the consequence of the use.
Warning sign: AI materially shapes support, feedback, or evaluation but appears as a neutral system instruction, invisible background process, or unexplained score.
Project example: A team records that AI helped summarize community-interview notes. The students compare the summary with the recording, correct omissions, and explain how the tool affected their synthesis.
4. Control
Decision question: Can the educator inspect, change, stop, and override meaningful AI-supported actions?
Actionable check: Test the real workflow. Can the educator see the relevant output and evidence, edit it materially, reject it without abandoning the project, stop further action, reverse a mistake where appropriate, and make the final consequential judgment?
Warning sign: An “approve” button is the only practical alternative to leaving the workflow, or the system acts before meaningful review can occur.
Project example: AI drafts rubric feedback from selected project evidence. The educator can inspect the cited evidence, remove an irrelevant observation, rewrite the feedback, reject an unsupported inference, and make the final evaluation.
Meaningful control is not a claim that human review removes every error or bias. It is a requirement that professional judgment can have practical consequences. For a deeper treatment, read Why educational AI must remain teacher-led.
- Student work and evidence
- AI suggestion available for inspection
- Educator judgment and student recourse
This illustrates an educational decision, not a current Dwyl.studio product workflow.
Review privacy, access, bias, and data handling
UNESCO's GenAI guidance explicitly raises data privacy and age-appropriate use. UNICEF's child-centred guidance places children's data and privacy alongside safety, fairness, transparency, accountability, and inclusion. These sources identify questions institutions should address; they do not guarantee that asking the questions produces a safe or compliant result.
5. Privacy
Decision question: What data is processed, for what purpose, by whom, and for how long?
Actionable check: Minimize the information provided. Identify the processor, purpose, access, storage, retention, deletion, onward use, and rules for sensitive or identifiable information before the activity begins. Check the relevant school policy and reviewed privacy information.
Warning sign: Identifiable student work, chats, images, recordings, or community data are uploaded because “the tool needs context,” without a bounded purpose, access rule, or deletion path.
Project example: Students test a survey-analysis workflow with fictionalized sample data rather than uploading identifiable community responses. The real responses remain in the school-approved environment and are analyzed through the agreed process.
For reviewed information about Dwyl.studio's public website data practices, use the privacy information. Do not infer future authenticated-product practices from this article.
6. Access
Decision question: Who may be disadvantaged, excluded, or pressured to participate?
Actionable check: Check device, connectivity, language, disability, cost, account, age, and consent barriers. Provide an equivalent non-AI route that protects the same learning opportunity and does not penalize a student whose access or circumstances change.
Warning sign: Opting out means a lower grade, a lesser project role, weaker feedback, extra burden, or loss of meaningful participation.
Project example: A group may use a printed questioning protocol and structured peer critique instead of an AI feedback tool while completing the same milestone and receiving the same educator support.
An alternative does not need to be identical, but it should preserve the intended learning opportunity. Educators should also look for uneven effects: whose language is misread, whose ideas are repeatedly flattened, and who must do extra work to correct the system? In one 2023 evaluation, seven AI-text detectors wrongly flagged non-native English essays as AI-generated at an average rate of about 61%, while classifying US 8th-grade essays almost perfectly.7
7. Evidence
Decision question: What supports the claimed educational benefit, for which context and population, and with what limitations?
Actionable check: Separate evidence about the specific tool and intended use from research or guidance about AI in general. Record the population, setting, outcome, comparison, uncertainty, and claims the evidence does not support.
Warning sign: A vendor cites a policy framework, a benchmark, a testimonial, or a study about another tool as proof that this use improves learning.
Project example: If an AI reflection coach increases the number of submitted reflections, treat that as an activity measure. Do not call it evidence of deeper reflection unless the evaluation actually examined the quality of student thinking in a relevant context.
Official guidance can sharpen the questions educators ask. It does not provide causal proof that a tool or checklist improves learning.
Likewise, a detector score is not proof of misconduct,7 and lower effort is not deeper learning.3
8. Recourse
Decision question: What can a student or educator do when the system is wrong, unfair, or harmful?
Actionable check: Name the responsible human, the challenge route, the expected response, how missing or incorrect evidence can be addressed, and what remedy is possible when a consequential decision was affected.
Warning sign: The only options are to retry the model, accept the output, or contact a generic help channel with no accountable educational reviewer.
Project example: A student challenges an AI-supported interpretation of their progress, adds a missing design note and peer-feedback record, and asks the educator to review the evidence before the project evaluation is finalized.
England's standards expect a formal complaints mechanism,4 and Australia's framework includes contestability as a guiding statement.6
Revisit the decision as evidence and tools change
Approval at the start of a project is not permanent. Project goals evolve. A provider may change a model, terms, data use, or interface. Students may encounter harms or barriers that were not visible in planning.
NIST's voluntary AI Risk Management Framework, version 1.0, treats ongoing monitoring, periodic review, and safe decommissioning as ordinary risk management. It is cross-sector, not education-specific.8
9. Review
Decision question: Who will revisit the decision, when, and using what evidence?
Actionable check: Schedule a checkpoint. Review learning evidence, incidents, uneven effects, educator overrides, student feedback, access problems, and changes to the system or policy. Decide in advance who can narrow, pause, or stop the use.
Warning sign: Approval remains in place despite recurring overrides, changed data handling, a model update, a different project purpose, or student reports that the tool is distorting their work.
Project example: At the midpoint, a class reviews whether AI brainstorming widened the range of ideas or made the proposals more alike. The educator changes the boundary for the next sprint based on that evidence.
In an online experiment with 293 adult writers, GPT-4 story ideas raised rated creativity but made the stories more similar to one another; it was not a school task.9
10. Exit
Decision question: Can the learning continue without the system or after access, permission, or consent changes?
Actionable check: Rehearse stopping the tool. NIST's framework calls for processes to decommission AI systems safely and mechanisms to disengage or deactivate them.8 Confirm where source materials and decision records remain available, what must be exported or deleted, how students continue, and how evaluation avoids depending on inaccessible AI output.
Warning sign: Core work, evidence, feedback, or grading becomes inaccessible when an account closes, permission changes, a provider withdraws a feature, or a student no longer participates.
Project example: Teams keep source notes, decisions, and original work in a school-controlled space. If the AI service is withdrawn, the inquiry continues with those records and the non-AI feedback process.
A before-and-during-project worksheet
Use the checklist twice: once when designing the activity and again at a named project checkpoint.
Scroll sideways to read the during-project column.
| Check | Before the project | During the project |
|---|---|---|
| Purpose | Name the learning need and intended outcome. | Is the use still serving that need? |
| Thinking | Mark the work students must own. | Is AI starting to substitute for that work? |
| Disclosure | Explain AI's role and student disclosure expectations. | Do students still understand material changes? |
| Control | Test inspect, edit, reject, stop, and override paths. | Are those paths practical in real use? |
| Privacy | Minimize data and verify purpose, access, retention, and deletion. | Has the information or provider behavior changed? |
| Access | Provide an equivalent non-AI route. | Is anyone excluded, pressured, or carrying extra burden? |
| Evidence | Record the claimed benefit and its limitations. | What learning evidence actually supports or challenges the claim? |
| Recourse | Name the human challenge and correction route. | Were concerns heard and remedied? |
| Review | Schedule the checkpoint and decision owner. | Narrow, continue, pause, or stop based on evidence. |
| Exit | Preserve continuity of work and evaluation. | Can the class stop today without losing the learning? |
What the evidence does and does not show
The sources used here are of two kinds. Six are guidance or frameworks from UNESCO, UNICEF, England's Department for Education, Australia's Department of Education, and NIST. Together, they support careful attention to human-centred purpose, age appropriateness, privacy, child rights, safety, fairness, transparency, accountability, inclusion, student agency, critical judgment, contestability, review, and the ability to stop. Three are limited empirical studies of specific risks: detector bias, weaker reasoning, and more similar output.
The studies do not show general effects. Each is a single study of specific tools at one point in time, with essays, university students, or adult writers rather than school project teams. None tests this checklist.
Neither kind of source proves that this checklist improves learning. They do not certify a product, establish that human review prevents harm, provide jurisdiction-specific legal advice, or guarantee privacy, security, fairness, compliance, or educational effectiveness. A responsible decision still depends on the project, students, tool, evidence, school context, and qualified local review.
Dwyl.studio's perspective
Dwyl.studio is building with educators around a clear boundary: AI should support educators and students without quietly taking over consequential educational judgment or the learning work students need to do themselves.
Our product direction includes teacher-reviewable support, useful evidence trails, meaningful edit/reject/stop/override paths, and student recourse. These are principles and intended directions, not claims that every safeguard or workflow is currently implemented, validated, or generally available.
The practical standard is deliberately tool-independent: start with learning, keep student thinking active, make AI involvement understandable, preserve educator authority, protect rights and access, demand appropriately bounded evidence, plan recourse, revisit the decision, and remain able to leave.
Contextual next step
If you are an educator who wants to help shape practical boundaries for responsible AI in real-world projects, you can join the conversation while we are building.
If you are reviewing this at school level, register school interest.