AI and Student Assessment: Practical Tools for Formative
Discover how AI assessment tools can transform your marking workflow whilst maintaining essential teacher judgement for effective student evaluation.
AI can mark a set of 30 factual recall tests in seconds, but it cannot judge whether a Year 8 learner's creative writing represents genuine progress for that individual child. That distinction, between what AI handles well and where teacher judgement remains essential, is the foundation of effective AI assessment. This guide covers formative and summative applications, the DfE's 2025 guidance, bias risks, data privacy requirements, and a practical approach to integrating AI marking into your existing workflow.
Key Takeaways
- AI significantly enhances formative assessment by providing rapid, low-stakes feedback, yet teacher judgement remains crucial for nuanced understanding of pupil progress. This aligns with the principles of effective formative assessment, where timely feedback supports learning, but the teacher's expertise is essential for interpreting complex responses and individual needs (Black & Wiliam, 1998). Teachers must discern when AI feedback is appropriate and when deeper human insight is required.
- Algorithmic bias is a significant concern in AI assessment, requiring teachers to critically evaluate outputs and understand potential inequities. AI systems can inadvertently perpetuate and amplify existing biases present in their training data, leading to unfair or inaccurate assessments for certain learner demographics (O'Neil, 2016). Educators must be vigilant in scrutinising AI-generated grades and feedback, ensuring equity and fairness for all learners.
- Robust data privacy protocols are non-negotiable when integrating AI into learner assessment, safeguarding sensitive personal information. The collection and processing of learner data by AI systems necessitate strict compliance with regulations such as GDPR, ensuring transparency and secure handling of information (Selwyn, 2019). Schools must establish clear policies and communicate them to learners and parents, maintaining trust and ethical practice.
- AI offers significant potential to deliver personalised and timely feedback, empowering learners to take greater ownership of their learning through self-assessment. Effective feedback, as highlighted by research, is crucial for improving learner attainment, and AI can provide specific guidance on 'where to next' in a way that is often difficult for teachers to scale manually (Hattie & Timperley, 2007). This enables learners to identify gaps and refine their work independently, fostering metacognitive skills.
What does the research say? Zawacki-Richter et al.'s (2019) systematic review of 146 studies found AI in assessment is most effective for automated essay scoring (r = 0.87 agreement with human markers) and adaptive testing. However, Luckin et al. (2016) caution that AI assessment tools perform poorly on creative and collaborative tasks. The EEF reports that feedback, the core purpose of assessment, adds +6 months of progress when specific, timely and actionable, whether delivered by AI or teacher.
In classrooms across the UK, AI tools for teachers are already reshaping how assessment works in practice. A 2025 Twinkl survey of 6,500 teachers found that 17% of those using AI apply it specifically to marking and feedback. The question is no longer whether to use AI for assessment, but how to use it well, in ways that save time without compromising the quality of professional judgement that makes assessment meaningful.
Formative vs Summative: Where AI Fits
AI is most valuable for formative assessment, where speed of feedback matters more than nuanced judgement, and least reliable for summative assessment, where stakes are high and professional accountability is essential. Dylan Wiliam's research demonstrates that the power of formative assessment lies in its timeliness: feedback given within minutes changes learning; feedback given after two weeks mostly confirms what learners already know or have forgotten.
AI closes this timeliness gap. A maths teacher using an AI-powered platform can see which learners answered incorrectly on last night's homework before the lesson starts, and adjust the starter activity accordingly. An English teacher can use AI to provide first-pass feedback on paragraph structure and technical accuracy, then spend her marking time on the higher-order aspects: quality of argument, development of ideas, and individual progress relative to targets.
The DfE's June 2025 guidance is explicit about the boundary. AI should be used for formative, low-stakes marking such as classroom quizzes, homework, and exam-style question generation. It should not be used for any assessment that contributes to formal reporting or examination without human oversight. The guidance also encourages teachers to use AI for generating formative assessment materials: quizzes, diagnostic questions, and feedback on drafts, all areas where speed matters and the consequences of an imperfect mark are low.
| Assessment Type | AI Role | Teacher Role | Risk Level |
|---|---|---|---|
| Multiple-choice quizzes | Auto-mark and report patterns | Review misconception data, adjust teaching | Low |
| Homework (factual) | Mark and provide feedback | Spot-check accuracy, intervene where needed | Low |
| Extended writing (drafts) | First-pass feedback on structure and SPaG | Evaluate argument quality, creativity, progress | Medium |
| Mock exams | Generate questions; initial scoring | Final grade, moderation, student discussion | Medium-High |
| Formal reports / GCSE coursework | Not recommended | Full professional responsibility | High |
AI Marking Tools: What They Can and Cannot Grade
Current AI marking tools achieve high agreement with human markers on structured, factual assessments but struggle with open-ended, creative and evaluative tasks. Understanding this performance profile prevents both over-reliance and unnecessary avoidance.
Automated essay scoring systems (tools like Graide, KEATH and general-purpose models like ChatGPT) show correlation coefficients of r = 0.87 with human markers on structured writing tasks (Zawacki-Richter et al., 2019). For factual recall, mathematics and science questions with clear correct answers, AI marking is fast and reliable. Schools piloting AI marking through the DfE's 2025-2026 assessment initiative report that teachers save 3-5 hours weekly on routine marking whilst maintaining assessment quality.
Where AI marking falters is predictable. Research from 2024-2025 highlights that AI tends to grade more leniently on low-performing work and more harshly on high-performing work, compressing the grade distribution towards the middle. ChatGPT shows 33.89% variation when scoring poor-quality assessments compared to 6% on high-quality work. This means AI marking is least reliable precisely where it matters most: at grade boundaries and for learners whose work does not fit typical patterns.
A practical approach: use AI to mark the bulk of routine assessments (saving hours), then personally review any work that falls near grade boundaries, any work from learners with SEND or EAL, and any assessment that contributes to formal reporting. This hybrid model captures the time savings whilst preserving professional accountability where it matters. The DfE's guidance reinforces this: AI "must always be used with human oversight."
Using AI for Feedback That Changes Learning
The value of feedback depends on timing and specificity, not on who delivers it. Hattie's meta-analyses consistently place feedback among the highest-impact teaching strategies (d = 0.70), but only when it is specific enough to guide next steps and timely enough to influence learning while the task is still fresh. AI excels at both.
Consider a Year 10 science class completing a practice paper on cell biology. Without AI, the teacher marks 30 papers over two evenings, returns them on Thursday, and discusses common errors on Friday. With AI-assisted marking, the teacher scans or uploads the papers on Monday evening, receives scored results with misconception analysis by Tuesday morning, and restructures Tuesday's starter activity to address the three most common errors. The learning gap between making the mistake and receiving feedback shrinks from four days to twelve hours.
AI feedback tools are particularly effective when configured to provide "feed-forward" guidance: not just what was wrong, but what to do next. Tools like SchoolAI and TeacherMatic can generate personalised revision recommendations based on individual error patterns. A learner who consistently confuses mitosis and meiosis receives different guidance from one who understands the process but mixes up terminology. This level of personalisation across a class of 30 would take hours manually; AI delivers it in minutes.
The critical caveat: AI feedback on written work, particularly in English, History and other essay-based subjects, should be treated as a first pass rather than a final judgement. AI can reliably flag structural issues, missing evidence, and technical errors. It cannot reliably evaluate the sophistication of an argument, the originality of an interpretation, or the emotional honesty of a personal narrative. Teachers who use AI for the former and reserve their own expertise for the latter report the greatest satisfaction with the hybrid approach.
AI Assessment by Subject: What Works
AI marking accuracy varies dramatically by subject and question type, and teachers who match the tool to their subject's assessment demands get far better results. Here is what classroom practice reveals across the major subject areas.
Mathematics. AI excels at marking numerical answers, algebraic expressions and graph-based questions where there is a clear correct response. Most AI maths marking tools can also evaluate method marks by recognising standard solution pathways. A Year 11 maths teacher can upload a set of 30 practice papers and receive scored results with misconception analysis (e.g., "12 learners correctly identified the gradient but 8 misapplied the y-intercept") within minutes. Where AI struggles: non-standard solution methods, problems requiring geometric reasoning from diagrams, and questions where a learner has found a correct answer via an incorrect method.
English. AI handles surface-level marking well: spelling, grammar, punctuation, sentence structure and basic paragraph organisation. Tools like ChatGPT can provide useful first-pass feedback on GCSE English Language paper structure ("Your opening paragraph establishes atmosphere effectively; your third paragraph needs a clearer topic sentence linking to the question"). Where AI fails: evaluating whether a metaphor is genuinely effective, judging whether a narrative voice is consistent, assessing whether an argument builds convincingly across paragraphs, and recognising when deliberate rule-breaking serves a creative purpose.
Science. AI marks factual recall questions and calculations reliably. It can evaluate "explain" questions when the mark scheme is tightly defined (e.g., "Name the process by which plants convert light energy into chemical energy" has one correct answer). It struggles with "evaluate" and "discuss" questions where learners need to weigh evidence, consider limitations of experimental design, or apply scientific principles to unfamiliar contexts. A KS3 science teacher might use AI to mark end-of-topic tests on factual content, then personally assess extended response questions on experimental design.
Humanities. History, Geography and RE assessments often require evaluative judgement: comparing sources, weighing interpretations, and constructing arguments from evidence. AI can assist with marking factual components (dates, key terms, identification of sources) but is unreliable for evaluating the quality of historical argument or the sophistication of geographical analysis. The most effective approach in humanities is using AI to generate exam-style questions and model answers, then marking learner work yourself with those model answers as a reference framework.
Primary assessment. AI is particularly useful for generating retrieval practice quizzes, phonics-based assessment materials, and maths fluency tests for KS1 and KS2. Primary teachers report the greatest time savings in the volume of low-stakes, formative assessments that build metacognitive awareness without consuming marking evenings. AI-generated exit tickets aligned to lesson objectives provide immediate data on learner understanding, enabling same-day intervention.
AI and Assessment Bias: What Teachers Must Know
AI assessment bias is not a theoretical risk; it is a documented reality that requires active management. AI algorithms learn from training data, and if that data reflects existing patterns of inequality, the AI will reproduce and sometimes amplify those patterns in its assessments.
The most common forms of bias in AI assessment tools affect learners whose language patterns differ from the training data. Learners with English as an Additional Language may receive lower scores on AI-graded writing, not because their ideas are weaker, but because their sentence structures differ from the patterns the AI associates with high-quality English. Similarly, learners who write in dialect or non-standard English may be penalised by automated scoring systems trained primarily on standard academic English.
Addressing bias requires three practical steps. First, audit your AI marking tools by comparing AI grades with your own grades across different learner groups (SEND, EAL, pupil premium, gender). If systematic differences appear, the tool needs recalibrating or supplementing with human review. Second, never use AI as the sole grading mechanism for any work that affects learner outcomes, reporting or setting. Third, maintain transparency: tell learners and parents how AI is being used in assessment and how human oversight is built in.
The broader ethical context matters too. Schools adopting AI assessment should develop clear policies, ideally building on the DfE's 2025 framework, that specify which tools are approved, what data can be processed, and how AI-generated grades are quality-assured. For guidance on developing your school's approach, see our guide to creating an AI policy for schools.
Data Privacy in AI Assessment
Any AI tool that processes learner assessment data must comply with UK GDPR, and many popular tools do not meet this standard by default. Before uploading learner work to any AI platform, verify three things: where the data is processed (ideally UK or EU servers), how long it is retained, and whether it is used to train the AI model.
The safest approach is to anonymise all work before AI processing. Remove learner names, school identifiers and any information that could identify an individual. Some schools use a numbering system where learners are assigned a code for AI-marked work, with the teacher holding the key. This adds a few minutes of preparation but eliminates the privacy risk entirely.
For tools designed specifically for schools (like Graide, KEATH, and TeacherMatic), check the provider's data processing agreement and ensure it meets your school's data protection officer's requirements. For general-purpose AI tools (ChatGPT, Gemini, Claude), the default terms of service typically allow user inputs to be used for model training unless you explicitly opt out. Schools should use the API or enterprise versions of these tools where possible, as these typically offer stronger data protection guarantees.
The DfE's guidance is clear: schools are responsible for ensuring that any AI tool used with learner data complies with UK data protection regulations. This responsibility sits with the school, not the tool provider. When in doubt, consult your data protection officer before introducing any new AI assessment tool. For broader guidance on responsible AI adoption, see our overview of AI in modern education.
When Learners Use AI for Self-Assessment
AI-powered self-assessment shifts the feedback dynamic from teacher-to-learner to a continuous loop where learners receive immediate guidance on their own work. This matters because self-regulation, the ability to monitor and adjust one's own learning, is among the highest-impact strategies identified in the EEF's Teaching and Learning Toolkit (+7 months).
Tools like SchoolAI allow teachers to create controlled AI environments where learners can submit practice work and receive structured feedback. A Year 9 learner writing a practice essay on Macbeth's ambition can receive immediate feedback on paragraph structure, use of quotation, and analytical vocabulary, then revise before submitting the final version to the teacher. The learner gets multiple feedback cycles in one lesson rather than waiting days for teacher marking.
The risk is dependency: learners who rely on AI feedback may not develop their own evaluative judgement. The solution is scaffolded withdrawal. In the first half-term, learners use AI feedback freely. In the second, they self-assess first, then check against AI feedback. By the third, they self-assess independently and only use AI for verification. This progression builds the higher-order thinking skills that matter more than any single piece of feedback.
Academic integrity requires clear rules. Learners must understand that using AI to generate answers (rather than to get feedback on their own answers) is dishonest. Schools with explicit policies, communicated at the start of each term and reinforced consistently, report fewer integrity issues than those that leave the rules ambiguous. The distinction is simple: AI as feedback tool is acceptable; AI as answer generator is not.
Frequently Asked Questions
schema.org/FAQPage">
What is the latest DfE guidance on using AI for marking?
The Department for Education states that AI should only be used for low stakes formative marking, such as classroom quizzes and homework. Teachers must not use AI for high stakes summative assessments that contribute to formal reporting without strict human oversight. Professional judgement remains essential for evaluating pupil progress.
How do teachers implement AI assessment tools in the classroom?
Teachers use AI platforms to quickly mark factual recall tests and generate initial feedback on paragraph structure. This allows them to see which learners answered incorrectly before the lesson starts. Teachers can then adjust their starter activities to address specific misconceptions immediately.
What are the benefits of using AI for student assessment?
The primary benefit is the speed of feedback, which educational research identifies as crucial for changing learning outcomes. Schools piloting AI marking report that teachers save between 3 and 5 hours per week on routine tasks. This saved time can be redirected towards responsive teaching and planning better lessons.
What does the research say about the accuracy of AI marking?
Research shows that AI marking achieves high agreement with human markers on structured, factual assessments. A 2019 systematic review found a strong correlation for automated essay scoring on structured writing tasks. However, studies also caution that AI tools perform poorly on creative tasks and require human moderation.
What are the common mistakes when using AI to grade learner work?
A major mistake is relying on AI to grade learners at the boundaries or those with special educational needs. Research highlights that AI tends to grade more leniently on weak work and more harshly on strong work. Teachers must personally review borderline cases to ensure fairness and accuracy.
Can teachers use AI to mark GCSE coursework?
No, AI should not be used to mark formal GCSE coursework or high stakes summative exams. Current tools compress the grade distribution and cannot reliably judge nuanced or highly creative work. Teachers must maintain full professional responsibility for any assessment that contributes to formal reporting.
Integrating AI Assessment: A Practical Approach
Successful AI assessment integration follows a pattern: start with the highest-volume, lowest-stakes assessments, prove the value, then expand gradually. Schools that attempt to implement AI across all assessment types simultaneously almost always retreat within a term.
| Phase | Duration | What to Do | Success Criteria |
|---|---|---|---|
| 1. Pilot | Half-term | One teacher, one subject, one assessment type (e.g. weekly vocabulary quizzes) | Time saved without quality loss |
| 2. Validate | Half-term | Compare AI marks with teacher marks on the same work. Check for bias across learner groups. | AI-teacher agreement above 85% |
| 3. Expand | Term | Extend to 3-5 teachers across subjects. Share findings at a staff meeting. | Consistent time savings, no quality complaints |
| 4. Embed | Year | Department-level adoption for formative assessment. Include in assessment policy. | Measurable workload reduction |
The teachers who integrate AI assessment most effectively share one trait: they evaluate honestly. If the AI saves time but produces feedback learners ignore, it has not added value. If AI marking is accurate but learners stop engaging because they know a machine is reading their work, the human cost may exceed the time benefit. Assessment is fundamentally a relationship between teacher and learner; AI can support that relationship but should never replace it.
For teachers ready to begin, start with our complete guide to AI tools for teachers for an overview of available platforms, and AI prompts every teacher should know for the specific prompt structures that produce reliable assessment content. Building AI literacy across your department ensures that colleagues can support each other through the learning curve, and a clear school AI policy provides the governance framework that makes adoption sustainable.
For a detailed breakdown of AI marking tools, bias risks, and a weekly feedback workflow, see our guide to AI marking and feedback.
Building whole-school confidence in AI-assisted assessment requires structured professional development. Our guide to AI CPD for schools outlines a year-long approach.
Further Reading
Further Reading: Key Research Papers
These peer-reviewed papers provide the evidence base for AI in assessment. Each offers practical implications for classroom practice.
Systematic Review of AI in Education View study ↗
Zawacki-Richter et al. (2019)
A systematic review of 146 studies that maps AI applications across four domains: profiling and prediction, intelligent tutoring, assessment and evaluation, and adaptive systems. The assessment findings show high AI-human agreement for structured tasks but significant limitations for open-ended evaluation.
Inside the Black Box: Raising Standards Through Classroom Assessment View study ↗
Black & Wiliam (1998)
The foundational paper on formative assessment that underpins the case for AI marking. Black and Wiliam's review demonstrated that improving the quality and timeliness of feedback produces substantial learning gains, particularly for lower-attaining learners. AI assessment tools are essentially an attempt to deliver on this promise at scale.
Intelligence Unleashed: An Argument for AI in Education View study ↗
Luckin et al. (2016)
Rose Luckin's influential report argues that AI's greatest educational contribution is better data for teachers, not replacement of teachers. The paper's assessment chapter shows that AI performs well on convergent tasks (single correct answer) but poorly on divergent tasks (creative, evaluative, collaborative). Essential reading for setting realistic expectations.
The Impact of Feedback on Student Learning View study ↗
500+ citations
Wisniewski et al. (2020)
A meta-analysis of 435 effects showing that feedback is most effective when it addresses the task level (what was done) and the process level (how it was done), rather than the self-regulation level. AI feedback tools that focus on task and process are therefore well-matched to the evidence base. Practical implications for configuring AI feedback systems.
Automated Essay Scoring: A Cross-Disciplinary Perspective View study ↗
200+ citations
Ke & Ng (2019)
A comprehensive review of automated essay scoring systems that examines both technical performance and pedagogical implications. The authors find that current systems are reliable for surface-level features (grammar, structure, vocabulary) but inconsistent for deeper qualities (argument coherence, critical analysis, originality). Important for understanding the current ceiling of AI marking capability.
