ASSESSMENT · FORMATIVE VS SUMMATIVE
Formative vs Summative Assessment: Where Does AI Add the Most Value?
AI adds value in both, but the case is strongest for formative assessment. Why timeliness matters, what the EU AI Act requires, and how to use AI responsibly.
By Eduface · July 2026 · 9 min read
You mark a set of formative essays. The feedback takes three weeks. By the time students receive it, the summative deadline has passed and the comments are irrelevant. This is not a failure of effort. It is a structural problem: formative assessment only works when the feedback arrives while students can still act on it. AI changes that calculus. But the question of where AI adds most value, and how much human oversight each use case requires, deserves a clear answer.
Where does AI add the most value?
AI adds value in both contexts, but the case is strongest for formative assessment. Formative feedback must be timely to be useful, and timeliness is exactly where scale creates the biggest bottleneck for lecturers. AI can generate per-criterion written comments at any scale within minutes. For summative assessment, AI can support consistency and rubric application, but every grade must be confirmed by a lecturer before release. The human-in-the-loop principle is not optional.
What is the difference between formative and summative assessment?
The distinction is commonly misunderstood as a difference in type (essays versus exams) when it is actually a difference in purpose.
Formative assessment is ongoing assessment designed to support learning. The grade is secondary to the feedback. The purpose is to help students understand where they are, where they need to be, and how to close the gap before a high-stakes submission. A practice essay, a draft submission, a mid-module test: all formative when the feedback is the point.
Summative assessment records achievement at the end of a unit or course. The grade is primary. It contributes to a degree classification or a progression decision. The feedback, when it arrives, is largely retrospective.
Sadler (1989) established the foundational principle: students need to understand the gap between their current performance and the desired standard. Summative grades alone do not close that gap. Formative feedback, delivered while students can act on it, does. This distinction matters for AI because the risks, governance requirements, and student expectations differ significantly between the two. Getting the framing wrong creates institutional exposure.
Why does AI feedback work particularly well for formative assessment?
The answer comes down to one word: timeliness. Black and Wiliam’s (1998) landmark review found formative assessment effect sizes of 0.4 to 0.7 on student achievement. That effect depends on feedback arriving when students can act on it. Feedback three weeks after a formative submission is not formative in any meaningful sense. It is retrospective commentary on work the student has already moved past.
Hattie and Timperley (2007) placed feedback at an effect size of d=0.73, among the highest of any educational intervention. The conditions for that effect size matter: feedback is most powerful when it is timely, criterion-referenced, and forward-looking. Nicol and Macfarlane-Dick (2006) reinforced this in their seven principles of good feedback, placing timeliness and forward-looking guidance at the core of formative value.
This is where AI has a direct structural advantage. A lecturer marking thirty formative essays alongside their teaching and research commitments cannot reasonably return criterion-referenced written feedback within 24 hours. AI can. For formative use, the draft feedback can be released to students quickly after a brief lecturer review: the stakes of a small error are lower because the student is still in the learning process and can continue to improve. Scale compounds the advantage. In large cohorts, the bottleneck is not the quality of any individual lecturer’s feedback. It is the sheer number of submissions. AI removes the bottleneck without removing the feedback.
Can AI be used for summative grading, and is it safe to do so?
Yes, with the appropriate governance in place. AI can support summative marking by applying rubrics consistently, reducing inter-marker variability, and producing a first-draft grade with written justification. These are genuine benefits, particularly in large cohorts where inconsistency across markers is a real quality risk.
The key word is first-draft. For summative assessment, AI should be treated as a first-pass tool, not a final arbiter. The institutional risk of an unchecked AI error is materially higher in summative contexts: a misjudged grade can affect degree classification or progression. The human review before any summative grade is released must be thorough, not a brief scan.
The regulatory picture reinforces this. The EU AI Act (Regulation 2024/1689) classifies AI systems used for student assessment as high-risk under Annex III. Article 14 requires meaningful human oversight over high-risk AI decisions. This applies to both formative and summative AI assessment, but it applies most critically to summative grading, where the stakes are highest and the impact most direct.
What governance do institutions need before using AI for summative marking?
A clear policy on AI’s role. AI should be positioned as a support tool, not a decision-maker. Grades should only be released once a lecturer has reviewed and approved them. This should be written into assessment policy, not left to individual departmental practice.
Defined review procedures. Formative review can be lighter: a brief lecturer check before feedback is released. Summative review should require the lecturer to engage substantively with each grade, checking for rubric misapplication, context the AI could not access, and cases that require professional judgement.
Audit trails. The EU AI Act requires that high-risk AI systems maintain logs sufficient to allow post-hoc review. Institutions should ensure their AI assessment tool produces records of the AI draft, any lecturer amendments, and the final released grade.
Student transparency. Students should know that AI has contributed to their assessment and that a human has reviewed the outcome. Concealing AI involvement where it is material to the grade is both an ethical problem and a regulatory risk.
Supplier compliance. The tool itself must meet EU AI Act requirements. Institutions should verify this in procurement, not assume it.
How does Eduface support both formative and summative use cases?
Eduface is built for both, on a single platform with the human-in-the-loop workflow built in as standard. For formative use: when a student submits an essay or written assignment, Eduface generates per-criterion written feedback immediately. The lecturer reviews the draft and releases it. The turnaround can be hours rather than weeks, keeping feedback genuinely formative.
For summative use: the same workflow applies, but the review step is designed to support thorough lecturer engagement rather than a rapid scan. Every grade remains a draft until the lecturer approves it. Nothing reaches the student gradebook without that confirmation.
95% alignment in UK pilots
In UK pilots, Eduface has achieved 95% alignment with lecturer assessments. It is EU AI Act compliant, Jisc/CHEST approved, and currently operating with pilot partners including Bath Spa University, De Haagse Hogeschool, Tilburg University, and Hogeschool Rotterdam. The platform supports written assignments, exam grading, and oral exams. Lecturers can operate in blind mode (no AI score visible before their own assessment) or AI-visible mode, depending on institutional preference.
Formative vs summative AI: a comparison
Dimension
Formative AI
Summative AI
Primary purpose
Support learning
Record achievement
Grade importance
Secondary
Primary
Timeliness requirement
Critical
Important but less acute
Risk if AI errs
Lower (student can still improve)
Higher (affects progression)
Human oversight level
Review before release
Thorough review mandatory
EU AI Act implications
High-risk; oversight required
High-risk; oversight required
Where AI adds most value
Speed and scale of feedback
Consistency and rubric application
Frequently asked questions
Is AI-generated formative feedback reliable enough to share with students?
Yes, provided a lecturer reviews it first. In UK pilots, Eduface has achieved 95% alignment with lecturer assessments. For formative use, the lecturer review does not need to be exhaustive: a check for obvious errors or misapplied criteria is sufficient. The goal is rapid turnaround. Brief, reliable feedback delivered within hours is more formative than detailed feedback delivered after three weeks.
Does the EU AI Act apply to formative as well as summative assessment?
Yes. The EU AI Act (Annex III) classifies AI systems used for assessing students as high-risk regardless of whether the assessment is formative or summative. Article 14 requires human oversight in both cases. The difference is in the degree of oversight required: summative grades carry higher institutional risk and should be reviewed more thoroughly before release.
Can AI assessment tools reduce bias in summative marking?
Potentially, yes. AI applies rubric criteria consistently across all submissions, which can reduce the variability that emerges in large cohorts with multiple markers. However, AI can also encode bias from training data or rubric design. The combination of AI consistency and human review is more robust than either alone. Audit trails and regular quality checks are part of responsible deployment.
What happens if a lecturer disagrees with the AI draft grade?
The lecturer overrides it. In Eduface, the AI grade is always a draft. The lecturer’s confirmed grade is the final grade. Where lecturers regularly override AI in a particular direction, that pattern is a signal worth investigating: it may indicate a rubric that needs refinement or an AI model that needs recalibration for that assignment type.
Does using AI for assessment require student consent?
This depends on jurisdiction and institutional policy. Under the EU AI Act, transparency about AI involvement in high-risk decisions is required. Students should be informed that AI contributes to their assessment and that a human has reviewed the outcome. Institutions should take advice on whether this constitutes a data processing activity requiring specific consent under GDPR, and ensure their AI assessment supplier has a compliant data processing agreement in place.
Conclusion
Formative assessment is where the structural case for AI is strongest: speed matters, scale creates bottlenecks, and rapid criterion-referenced feedback produces measurable learning gains. Summative AI grading is also valid, but it requires thorough human review and clear institutional governance before any grade reaches a student. The two use cases are not in competition. A well-implemented AI assessment platform should handle both, with the level of human oversight calibrated to the stakes involved.
References
Black, P., & Wiliam, D. (1998). Assessment and classroom learning. Assessment in Education: Principles, Policy and Practice, 5(1), 7-74.
Hattie, J., & Timperley, H. (2007). The power of feedback. Review of Educational Research, 77(1), 81-112.
Nicol, D. J., & Macfarlane-Dick, D. (2006). Formative assessment and self-regulated learning: a model and seven principles of good feedback practice. Studies in Higher Education, 31(2), 199-218.
Sadler, D. R. (1989). Formative assessment and the design of instructional systems. Instructional Science, 18(2), 119-144.
European Parliament and Council. (2024). Regulation (EU) 2024/1689 (AI Act). Official Journal of the European Union.
See AI feedback in your own assessments
Try Eduface free or book a demo, and see how per-criterion feedback works, formative or summative.