The Executive Office’s draft Artificial Intelligence (AI) strategy proposes that AI should help teachers grade and analyse pupil performance data. The Education Authority (EA), announcing its rollout of Copilot and Gemini, points to one supplier-funded study of Copilot. That study recorded what participants believed but did not measure what pupils learnt. While time saved on administration is a real gain, a better-educated pupil and a grade a teacher can stand over are different achievements. The strategy requires every AI initiative to show measurable benefit and calls for testing and evaluation before scaling, but it does not say how to determine whether an AI-assisted grade is sound or whether a pupil has learnt more.
The strategy lists minute-taking among the routine public-sector uses it has in mind. If an AI tool produces accurate meeting minutes more quickly, we can check the record against what was said, correct omissions, and compare the time taken with the previous process. A pupil who produces an answer more quickly has not necessarily learnt more. Sometimes the effort involved in reaching the answer is precisely what the lesson is intended to develop.
The strategy also envisages assistance with grading and with analysing pupil performance, and places these ambitions within principles covering human oversight, accountability, fairness and public benefit. Shorter administration, defensible assessments and improved pupil understanding each need their own test.[1]
A test from Belfast City Council
Belfast City Council’s (BCC) draft consultation response offers a useful test. It argues that success should be assessed through demonstrable improvements for citizens and communities. Its proposed measures include service quality, accessibility, inclusion, sustainability, and productivity. It also recognises the practical constraints: governance, data management and specialist capacity require resources.[2]
This is a stronger starting point than counting licences or completed pilots. Those figures tell us something about implementation. The public is entitled to ask what difference implementation makes.
Education is already part of the strategy’s ambition. Section 6 says the skills agenda must begin at school level. It calls for consistent regional guidelines for teachers and pupils covering responsible use, approved tools, e-safety and data use. Developing guidance for schools is one of its recommended actions. It also advocates extending initiatives such as the Trailblazing NI project.[1] The question is how that guidance will translate into school practice, and who will hold schools to it.
What the EA is offering
The EA’s published programme, announced by the Education Minister on 6 May 2026 with £10.7 million for licences, provides for Microsoft Copilot and Google Gemini within managed school systems, accompanied by training and guidance.[8] Its stated aims include reducing teachers’ administrative workload and improving planning and preparation. The offer targets teachers, with phased access and support during the 2026/27 academic year.[3]
There is a reasonable case for helping teachers with routine work. Drafting a standard letter, preparing alternative versions of a worksheet or organising meeting notes can consume time that might be used more productively. Teachers should not have to preserve every laborious process simply because it predates the technology.
However, the resulting material still needs professional scrutiny. A worksheet must be accurate and appropriate to the pupils. A letter must reflect what the school actually intends to say. A summary must preserve the information on which a decision depends. The saving that matters is the time remaining after checking and correcting the output.
Improved staff wellbeing should not be treated merely as an intermediate step towards higher attainment. Reducing unnecessary pressure on teachers is a worthwhile outcome in itself. The case weakens when it combines every benefit into a general assertion that AI improves education, without specifying whose experience has improved or how.
From paperwork to grades
The strategy’s references to grading and performance analytics matter more than its references to paperwork. Organising assessment records, recommending a mark for a teacher to review, and determining a grade automatically involve different degrees of delegated judgement. The strategy’s wording does not say which of these it has in mind, and the guidance will need to. An unjustified grade or misleading assessment of a pupil’s progress could shape feedback, expectations and subsequent teaching. Checking such outputs requires subject knowledge, familiarity with the pupil and an understanding of what the assessment is intended to establish. A teacher’s approval is meaningful only if the teacher can examine the basis of the result and has time to challenge it.
The EA’s public programme describes administrative, planning and resource-preparation benefits and is silent on grading. That silence does not establish whether grading is permitted, prohibited or governed elsewhere, but it leaves readers of the programme without an account of how the strategy’s wider ambition would be implemented.[3]
The silence is partly filled elsewhere. The Department of Education (DE) issued Circular 2026/02 on generative AI on 11 May 2026, three months before the draft strategy. It is advisory, covers staff and pupil use, and tells schools to give learners explicit direction on when AI may be used in learning tasks and assessment, to embed the JCQ rules for examinations and to point pupils to the NI Council for the Curriculum, Examinations and Assessment’s (CCEA) advice on AI in assessments. It also tells schools to appoint a GenAI lead, name the person responsible for the risk of errors from undue reliance on AI output, and monitor and evaluate the impact of their own policies. It also recognises the risk this article raises, asking schools to ensure that undue reliance on GenAI does not limit learners’ cognitive development. What it does not do is say how any of these expectations will be tested. It sets no measure of pupil learning, and it specifies no external process for checking that schools have met them.[10]
What the pilot establishes
The evidence supporting expansion deserves equally close attention. Trailblazing NI was a proof-of-concept study of Microsoft Copilot involving 84 teachers and ten other education professionals. C2k approached Ulster University in partnership with Microsoft. Microsoft Ireland funded the study; the report states that the research was conducted independently and that Microsoft was not involved in data collection, analysis or interpretation.[4]
Funding alone does not establish bias. The concern is the weight placed on a small, supplier-funded study of that supplier’s product when making the case for expansion. The strategy names the project as something to extend, and the EA identifies the report as the culmination of its proof of concept. The study covered Copilot. The EA also ran a Gemini pilot with over 100 teachers; its newsletter does not report the methods or findings.[7] That makes further evaluation, independent of the supplier and directed at educational outcomes, necessary before this pilot is treated as sufficient evidence for wider educational claims.[1][3]
The findings are encouraging about participants’ experiences. In the post-pilot attainment question, 41.2 per cent reported that GenAI had already contributed to improvement. On a base of 68 respondents, that is 28 people. Another 38.2 per cent expected future benefit, while 14.7 per cent did not know.[4]
These are participants’ judgements about attainment. They are not measurements of it. They do not establish how much pupils learnt, whether improvement lasted or whether AI caused it. Fewer respondents said they did not know after the pilot (14.7 per cent, against 34.8 per cent before), but the two surveys had different response totals (92 and 68), and the aggregate figures do not establish that any individual’s view changed.
The report is useful evidence of perceived benefit and of greater confidence among post-pilot respondents. Its own description as a proof of concept is appropriate. Moving from that evidence to wider provision requires a clear account of what has been demonstrated and what still needs testing.
Pupils are a separate question
Direct use by pupils raises a further set of issues. It should be considered separately from the EA’s offer of tools to teachers. A teacher using AI to prepare a lesson and a pupil using it to complete the lesson’s task occupy different positions. One is exercising an established professional skill; the other may still be acquiring the underlying knowledge and judgement.
Consider a history exercise requiring pupils to read two sources and explain why their accounts differ. AI might help a pupil understand unfamiliar vocabulary, offer a question to guide attention or provide feedback on an attempted comparison. It might also produce the comparison before the pupil has read either source. The same polished paragraph could conceal very different amounts of learning.
Or consider a pupil struggling with a difficult text. A simplified version may make the subject accessible and allow meaningful participation. Used indiscriminately, simplification could remove opportunities to become a more confident reader. The educational judgement concerns the pupil, the purpose of the task and the support needed at that stage.
This makes the familiar advice to check AI’s answers more demanding than it sounds, because checking requires knowledge. Pupils must have some basis for recognising an implausible claim, a missing qualification or an invented reference. Asking them to verify information is useful only if we also teach them how to do so and provide sources they can assess.
Trailblazing records participants’ concerns about plagiarism, the undermining of independent thinking and the misreading of AI output; it does not demonstrate harm to pupils’ learning, because it did not study pupils.[4] The classroom examples show that the same output can sit on top of very different amounts of learning, and that is why a pupil’s own work should be treated differently from a teacher’s paperwork.
Studies elsewhere show what an answer would look like. In the Education Endowment Foundation’s 2024 trial of ChatGPT for lesson preparation, involving 259 science teachers across both arms, those allocated to ChatGPT spent 31 per cent less time on the specified planning tasks, with no apparent reduction in resource quality; the trial did not measure pupil learning. Where pupils have used the tools directly, a World Bank trial in Nigeria found gains in English after six weeks with Copilot, and a trial with nearly 1,000 Turkish maths pupils found large gains during practice and poorer unassisted performance afterwards in the group given an unguarded tool, a disadvantage largely avoided where the tool was designed to withhold answers.[9] The design of the assistance mattered. The Northern Ireland (NI) documents discussed here do not set out an equivalent evaluation of pupil learning.
What Wales asked for
The guidance will also need to meet the curriculum. The DE’s consultation on the NI Curriculum 2028, reviewed here in June, rests its promise of equity on clearer, sequenced content and offers schools a digital hub of free and adaptable materials. The EA’s programme gives teachers tools to adapt materials. Neither document explains how AI-adapted resources are to preserve the content and progression the new curriculum specifies, and the draft AI strategy’s references to curriculum concern further and higher education only. That is the question the school guidance must answer.[5]
Wales offers a relevant comparison. Estyn’s original report, A New Era, calls for national guidance supported by practical materials, professional learning and clear accountability for implementation. It also calls for monitoring focused on the effects on learners and the education system. Its evidence includes a self-selecting survey and visits to 21 schools chosen for their engagement with AI. It offers examples of developing practice, and it does not demonstrate improved attainment.[6]
Guidance, responsibility and evidence
The proposed regional guidance is where these questions get answered. It should distinguish staff preparation from grading and pupil use, identify permitted information and explain what must remain subject to a teacher’s independent judgement. Parents and governors need to understand the rules, while teachers need practical examples of how they apply to different subjects, ages and learning needs.
Responsibility also needs to be settled. How will the Executive Office, DE and EA divide policy, implementation and oversight? Principals and boards of governors need to know what they are expected to check, and inspection needs a view of what good practice looks like. The EA’s proposed school GenAI leads may support professional learning, but that role should not be confused with independent assurance. Guidance will have limited force if nobody knows who must respond when practice falls short.
The strategy’s own requirement of measurable benefit supplies the test. In schools, independent evaluation should test understanding, retention and application to unfamiliar problems, including what pupils can subsequently do without assistance. It should examine whether benefits and risks differ by age and by household disadvantage, assess continuing costs and control of information, and publish disappointing results as well as successes. This would connect the scrutiny of supplier-funded research to the decisions that follow from it.
The rule that follows is one of proportion. A teacher should not need an attainment trial before using AI to draft a letter or vary a worksheet, and the guidance should say so. The standard is not one this article invents; it is the strategy’s own requirement of measurable benefit, applied to the uses the strategy itself names. AI-assisted assessment, and any use that may substitute for a pupil’s own reasoning, should be permitted only where the guidance names it, a teacher can examine and overrule the output, and evidence about learning is being gathered in proportion to the consequences of the use. Carried into schools, BCC’s emphasis on demonstrable outcomes points to the same distinction. Faster paperwork can release time for teaching. The educational test is whether pupils become more knowledgeable and better able to think for themselves.
Sources:
1. The Executive Office, Northern Ireland’s Draft Artificial Intelligence Strategy: consultation, and Draft Northern Ireland Artificial Intelligence Strategy, August 2026, section 6, pp. 32–35, and section 7. The proposed school guidance and the reference to extending Trailblazing appear on pp. 32 and 35; the requirement for measurable benefit and the outcome metrics appear on p. 26; testing and evaluation before scaling appears on p. 31; and guidance on quantifiable measurement of success or failure appears on p. 37.
2. Belfast City Council, Draft response to consultation on the Northern Ireland Public Sector AI Strategy, appendix to the Strategic Policy and Resources Committee report dated 18 September 2026, responses on societal benefit, oversight and data governance. The report dated 18 September sought the committee’s approval.
3. Education Authority, GenAI for Teaching and Learning, accessed 26 September 2026: “Research and Feasibility” states that the proof of concept culminated in the report; “Next Steps” specifies a GenAI lead in each school.
4. Sammy Taggart and Stephen Roulston (2025), Trailblazing NI: GenAI in Education: A Proof-of-Concept Study in Schools in Northern Ireland with MS Copilot, Ulster University: project origins, p. 8; funding and independence statement, p. 9; attainment perceptions, tables 25–26, pp. 54–55; participants’ concerns about learners, executive summary, p. 2. The pre-pilot attainment counts total 92; the post-pilot counts (10, 1, 3, 26 and 28) continue at the top of p. 55 and total 68.
5. El Cavador, Excellence and equity deferred: reviewing the draft Northern Ireland Curriculum 2028, Slugger O’Toole, 19 June 2026; Department of Education, Consultation on the Northern Ireland Curriculum 2028, 16 June 2026.
6. Estyn, A New Era: How Artificial Intelligence (AI) is Supporting Teaching and Learning, October 2025, recommendations pp. 9–11, implementation paragraph p. 11, methods pp. 4 and 50–51.
7. Education Authority, EdIS Newsletter, June 2026, “Google for Education Research Partnership”: “Building on our recent Gemini pilot with over 100 teachers”. No method, findings or report are identified.
8. Department of Education, Givan announces roll out of AI tools and training for Northern Ireland teachers, 6 May 2026.
9. Education Endowment Foundation, ChatGPT in lesson preparation: a Teacher Choices trial, evaluated by NFER, report December 2024 (259 teachers, 68 schools, 31 per cent less planning time; pupil outcomes not measured). De Simone, M.E. et al. (2025), From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria, World Bank Policy Research Working Paper 11125 (six weeks, Microsoft Copilot, first-year senior secondary pupils; 0.23 standard deviations on English). Bastani, H. et al. (2025), Generative AI without guardrails can harm learning: evidence from high school mathematics, Proceedings of the National Academy of Sciences, 122(26), e2422633122 (unassisted scores 17 per cent lower for the unguarded arm).
10. Department of Education, DE Circular 2026/02: Guidance for education professionals on Generative AI, 11 May 2026. Status of contents: information and advice for teachers and schools.
El Cavador is a Slugger reader from Belfast with a particular interest in education
Discover more from Slugger O'Toole
Subscribe to get the latest posts sent to your email.
