Why feedback needs urgent attention

Before we discuss AI, it is important to recognise the feedback challenge educators are already facing.

Imagine marking thirty assignments over a weekend. The first few receive detailed, thoughtful comments. By the final few, fatigue has inevitably crept in. The feedback may still be professional, but it is often thinner, slower or less detailed than intended.

This is not because educators care less. It is because feedback is one of the most time-intensive and cognitively demanding parts of teaching.

Across education, three persistent challenges continue to emerge:

• Workload
• Timeliness
• Consistency
  
  

Feedback is one of the most pedagogically valuable things educators do. However, when feedback arrives weeks after submission, the learning window may already have closed.

The case for AI-supported feedback is not that AI can replace teachers. It is that, when used responsibly, it can reduce some of the mechanical burden involved in producing feedback while keeping educator judgement firmly at the centre.

The governance challenge institutions now face

As AI tools become more capable, the questions institutions ask are changing.

Early conversations often focused on whether AI-supported feedback worked. Now, the questions are more strategic:

• How do we govern AI-supported feedback?
• How do we explain it to students?
• What should we tell external examiners?
• How do we ensure fairness and consistency?
• What are our responsibilities around data protection?
  
  

These are the right questions. They show that institutions are moving beyond curiosity and towards responsible implementation.

This is why the SAFE Framework was developed.

What is the SAFE Framework?

The SAFE Framework is a structured, pedagogy-first model designed to help institutions adopt AI-supported feedback and assessment responsibly.

SAFE is built around four pillars:

These pillars are not a sequence to complete one by one. They are four lenses that should be applied together whenever institutions make decisions about AI-supported feedback.

What SAFE is not

One of the most important aspects of SAFE is its honesty.

SAFE is not:

• A product endorsement
• A replacement for professional judgement
• A guarantee that AI-generated feedback is always accurate
• A substitute for awarding body guidance
• A justification for removing educators from the process

This matters because responsible AI adoption depends on trust. Overclaiming what AI can do risks undermining the confidence institutions need to build.

Pillar 1: Safeguarding Data and Privacy

When an educator uploads student work to an AI tool, that is a data processing activity. Institutions therefore need to ask clear questions before implementation.

• Where is student work stored?
• Who processes it?
• Is student work used to train or fine-tune AI models?
• What happens to uploaded work after feedback is generated?
• Who retains ownership of submissions and outputs?

These questions should be asked of every AI provider in writing.

For institutions, safeguarding also requires internal action. Existing data protection policies may need to be updated to reflect AI-supported feedback workflows. Student consent, retention procedures, deletion processes and provider agreements all need to be considered before rollout begins.

Pillar 2: Augmenting Professional Judgement

This pillar sits at the heart of responsible AI-supported feedback.

The educator must remain the author of the feedback decision.

AI may generate a draft. The educator reviews it, edits it, challenges it where necessary and decides whether it is suitable for release.

Under SAFE, educator review is non-negotiable. The moment human oversight is removed, the process moves beyond AI-supported feedback and becomes something far more problematic.

A useful way to think about this is the hybrid marking model. AI can support the structure, consistency and first draft. The educator brings context, nuance, professional judgement and final approval.

This is not passive approval. It is active professional review.

Pillar 3: Fairness and Inclusion

A common assumption is that AI is fair because it can be consistent. However, consistency is not the same as fairness.

AI-generated feedback may reflect bias in training data, favour particular writing styles or miss important contextual information about learners.

This is especially important when supporting:

• Learners with SEND
• Neurodivergent learners
• English as an Additional Language learners
• Students from diverse cultural or educational backgrounds
  
  

Feedback is not just information. It is communication. Tone, language and framing can shape how students understand their progress and their next steps.

Educators know their learners. They can soften blunt phrasing, clarify meaning, adapt language and ensure feedback is constructive, accessible and appropriate.

Pillar 4: Ethical and Transparent Practice

Transparency is what makes AI-supported feedback sustainable.

Students should understand how AI is being used, what role the educator plays and whether AI has any role in grading decisions.

Staff also need clarity. Hidden or inconsistent AI use can undermine confidence, create confusion and weaken institutional trust.

External stakeholders, including quality teams, awarding bodies and external examiners, may also need clear documentation showing how AI use is governed, reviewed and recorded.

Ethical practice depends on openness. AI-supported feedback should not be something educators feel they have to hide.

Why framing matters

One of the most powerful insights from the webinar was the importance of how AI-supported feedback is explained to students.

The underlying process may be similar, but the level of trust created is very different.

Students are far more likely to understand and accept AI-supported feedback when they know that the educator remains responsible, AI does not decide grades and the process is being used to improve the quality and timeliness of feedback.

Putting SAFE into practice

SAFE is not only a set of principles. It is designed to support practical implementation.

Institutions should begin by considering four foundational areas.

The educator-led workflow

A responsible AI-supported feedback process should make clear where AI is involved and where human judgement remains essential.

1. Upload: The educator uploads the brief, rubric, learning outcomes or assessment criteria alongside the student work.
2. Generate: AI generates draft feedback aligned to the criteria. The output is advisory.
3. Review: The educator reviews, edits and approves every piece of feedback against their professional judgement.
4. Release: Feedback is released to the student with a declaration that AI was used in the drafting process.
5. Record: Use is logged for quality assurance, external review and audit purposes.

In this workflow, AI is involved in generation. The educator remains responsible for every consequential decision.

Start with formative feedback

Institutions do not need to begin with high-stakes summative assessment. In fact, that is often the wrong starting point.

Formative feedback provides a lower-risk, high-value entry point. It allows staff to build confidence, test workflows and develop governance processes before moving into more complex assessment contexts.

Suitable starting points may include:

• Draft submissions
• Practice assignments
• Reflective journals
• Portfolio development
• Low-stakes formative tasks
  
  

Once formative use is embedded, institutions can then consider whether and how AI-supported feedback might extend into summative contexts, always with awarding body guidance checked first.

Important limitations

Responsible implementation requires honesty about what AI cannot do.

SAFE is clear about several non-negotiable limitations:

• Educator review cannot be bypassed.
• AI should not determine final marks or grades.
• Awarding body rules vary and must be checked individually.
• AI-generated feedback may still contain bias or inaccuracies.
• AI plagiarism detection should not be treated as definitive evidence of misconduct.
  
  

These limitations do not make AI-supported feedback unusable. They make responsible governance essential.

A living framework shaped by the education community

SAFE is not intended to be a finished document.

AI in education is developing quickly. Regulatory guidance is still emerging. Awarding bodies are continuing to form their positions. Sector evidence, including work from the JISC AI Assessment Pilot, will continue to shape practice.

This is why SAFE has been designed as a living framework. Feedback from colleges, universities, schools and sector partners will help shape future versions.

Final thoughts

The conversation around AI in assessment often focuses on technology. SAFE deliberately brings the conversation back to people, practice and professional judgement.

It asks institutions to consider:

• Who needs to be involved?
• What policies need updating?
• How will students be informed?
• What safeguards need to be in place?
• How will educator judgement remain central?
  
  

These are not only technical questions. They are pedagogical, ethical and strategic questions.

As AI-supported feedback becomes more common across education, frameworks such as SAFE give institutions a practical way to move forward with confidence. Not by removing educators from the process, but by helping them use AI responsibly, transparently and with professional judgement at the centre.