# AI in education: how schools and universities are using it

[Skip to content](#lm-inhoud)Network/[NL](/en/ai-in-het-onderwijs-hoe-scholen-en-universiteiten-het-inzetten)EN[Hubhub.llmnet.nlCompare models on task, language, cost and license.](https://hub.llmnet.nl/en/)[Communitycommunity.llmnet.nlPrompt techniques, patterns and system prompts.](https://community.llmnet.nl/en/)[APIapi.llmnet.nlLLMs in production: rate limits, routing, structured output.](https://api.llmnet.nl/en/)[Consultancyconsultancy.llmnet.nlRolling out AI in an organization, pilot to production.](https://consultancy.llmnet.nl/en/)[Newsnieuws.llmnet.nlAI developments, explained for the Netherlands.](https://nieuws.llmnet.nl/en/)[Benchmarkbenchmark.llmnet.nlMeasure AI quality yourself, on your own tasks.](https://benchmark.llmnet.nl/en/)[Careersvacatures.llmnet.nlAI roles, salaries and career paths in the Netherlands.](https://vacatures.llmnet.nl/en/)[Learnleren.llmnet.nlAI concepts in plain language, beginner to builder.](https://leren.llmnet.nl/en/)[Guidegids.llmnet.nlRun AI privately on your own Mac, PC, NAS or home server.](https://gids.llmnet.nl/en/)[Directorydirectory.llmnet.nlMapping the AI ecosystem: tools, models, companies.](https://directory.llmnet.nl/en/)[Radarradar.llmnet.nlSignals from X, research and communities for indie developers.](https://radar.llmnet.nl/en/)[Appsapps.llmnet.nlReviews of AI apps and open-source repos, with tips for builders.](https://apps.llmnet.nl/en/)[llmnet.nl — main site](https://llmnet.nl/en/)[](https://x.com/intent/post?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fai-in-het-onderwijs-hoe-scholen-en-universiteiten-het-inzetten&text=AI%20in%20education%3A%20how%20schools%20and%20universities%20are%20using%20it)[](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fai-in-het-onderwijs-hoe-scholen-en-universiteiten-het-inzetten)[](https://www.reddit.com/submit?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fai-in-het-onderwijs-hoe-scholen-en-universiteiten-het-inzetten&title=AI%20in%20education%3A%20how%20schools%20and%20universities%20are%20using%20it)[](#)[](https://x.com/intent/post?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fai-in-het-onderwijs-hoe-scholen-en-universiteiten-het-inzetten&text=AI%20in%20education%3A%20how%20schools%20and%20universities%20are%20using%20it)[](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fai-in-het-onderwijs-hoe-scholen-en-universiteiten-het-inzetten)[](https://www.reddit.com/submit?url=https%3A%2F%2Fnieuws.llmnet.nl%2Fen%2Fai-in-het-onderwijs-hoe-scholen-en-universiteiten-het-inzetten&title=AI%20in%20education%3A%20how%20schools%20and%20universities%20are%20using%20it)[](#)

 
# AI in education: how schools and universities are using it

 By Ivo Donker — compiled with AI assistance (Claude & Gemini)

 The introduction of generative language models initially triggered panic in the education sector over large-scale digital cheating and the devaluation of diplomas. In the meantime, the playing field has become considerably more mature. Educational institutions in the Netherlands — ranging from primary schools and regional training centers (mbo) to universities of applied sciences and universities — have made the shift from ad hoc bans to structural policy, tested didactic integration, and targeted infrastructural provisions. Where the earlier background article on [the opportunities and concerns of education AI](https://nieuws.llmnet.nl/en/ai-in-het-onderwijs) mainly highlighted the initial reactions and philosophical questions, here we analyze the actual implementations, technical architectures, and legal obligations that educational Netherlands operates under in 2026.

 Today's reality shows a pragmatic split. On one hand, language models function as personalized learning assistants, formative feedback tools, and administrative relief for teachers. On the other hand, the technology forces examination boards to thoroughly revise the traditional essay and take-home assignments. In addition, strict European frameworks set hard limits on what is permissible in automated student assessments.

 
## The shift from standalone chatbots to integrated learning environments

 In the early phase of generative AI, pupils and students mainly used public consumer interfaces. This approach carried significant risks around data protection, source citation, and hallucinating answers without pedagogical context. Today, use within educational institutions increasingly runs through controlled learning environments (VLEs) such as Canvas, Brightspace, and Moodle, or through specialized intermediary layers.

 Instead of an answer machine that produces a ready-made paper, models are configured via specific system instructions as Socratic tutors. A Socratic AI tutor never gives the student the solution to a math problem or a historical analysis directly, but asks targeted counter-questions to stimulate the thought process. If a pupil asks: "Why did World War II break out?", the model does not just list dates, but asks: "Which economic and political consequences of the Treaty of Versailles do you remember?"

 In these environments, teachers can follow along via anonymized interaction logs. This lets them see exactly where classroom-wide misconceptions arise without violating individual pupil privacy. This shift transforms language technology from a tool for cheating into a measurable instrument for formative evaluation.

 
## Adaptive learning systems in primary and secondary education

 In primary education (po) and secondary education (vo), the focus of AI use lies mainly on adaptive learning tools for language, math, and modern foreign languages. Established educational publishers and digital learning platforms have enriched their deterministic algorithms with compact language models. The goal is not to replace teachers, but to accommodate differences in level within full classrooms.

 A pupil who struggles with reading comprehension is offered a text that has been dynamically rewritten to a more accessible language level (for example from B1 to A2), while retaining the core content concepts. As vocabulary and text comprehension improve, the system automatically scales up the complexity. In subjects such as English and French, models generate interactive role-plays in which pupils practice speaking and writing skills in realistic scenarios, such as ordering a meal or conducting a job interview.

 For teachers, the biggest time savings are in the preparation phase. Differentiating teaching materials — traditionally a time-consuming task — now happens within minutes. A physics teacher enters a basic concept and has three versions generated: one with visual metaphors for visual thinkers, one with extra mathematical depth for vwo pupils, and a practice-oriented variant for vmbo.

 
## Higher education and universities: synthesis and code generation

 At universities of applied sciences and universities, the nature of AI use is fundamentally different. Students and researchers use language models for large-scale literature synthesis, data analysis, and accelerating software development within technical programs. Manually sifting through hundreds of scientific PDF files has shifted to targeted semantic querying via Retrieval-Augmented Generation (RAG).

 In computer science and data science programs, writing routine source code with AI support has by now become standard practice. As a result, the focus of teaching has shifted from syntax and elementary programming structures to architectural design, code auditing, security analysis, and debugging. Students must demonstrate that they thoroughly understand model-generated code, can optimize it, and can check it for vulnerabilities.

 At the same time, academic education faces intellectual challenges. When students use a model to summarize papers, there is a risk of superficial knowledge absorption. Critically weighing methodological shortcomings in scientific sources requires deep cognitive effort that can be lost when one blindly relies on automated summaries. Many faculties therefore apply strict guidelines: AI may serve as a sparring partner, but the synthesis and source accountability must demonstrably have been carried out by the student themselves.

 
## Assessment, examination, and the bankruptcy of traditional writing assignments

 The biggest institutional shock from generative AI is playing out within examination boards. Virtually all Dutch higher education institutions have determined that so-called AI detection tools are unreliable. These statistical detectors show too many false positives — where authentic student work is wrongly flagged as generated — and are easily circumvented by lightly paraphrasing text.

 The result is a fundamental revision of the assessment matrix. The traditional, unsupervised take-home essay is rapidly losing its status as primary evidence of academic competence. Instead, institutions apply a combination of alternative assessment forms:

 
 
 
 
 Assessment form | 
 Didactic focus | 
 Role of AI | 
 Vulnerability / Point of attention | 
 

 
 
 
 Oral defense (criterion-based) | 
 In-depth understanding and reflection | 
 No access during the assessment | 
 High burden on teacher capacity | 
 

 
 Process-based writing (version control) | 
 Development of the line of argument | 
 Permitted with mandatory logbook audit | 
 Requires detailed assessment rubrics | 
 

 
 Scheduled in-person testing (on-campus) | 
 Ready knowledge and direct application | 
 Blocked via secure browser environment | 
 Limited connection to later professional practice | 
 

 
 AI-critical review assignment | 
 Error detection and academic synthesis | 
 Model output serves as raw input | 
 Requires a high level of source knowledge | 
 

 
 
 

 With process-based writing, students submit not only the final product but also the intermediate versions, including the prompts used and the reasoning behind editorial choices. This allows the student to demonstrate that the thought process and final editing remained under their own control.

 
## Privacy, data sovereignty, and the AI Act

 Using commercial AI services in a public education context brings strict legal frameworks along with it. Under the GDPR (Algemene Verordening Gegevensbescherming, AVG), it is not permitted to enter personal data, pupil performance, or confidential graduation projects into public models that reuse this data for training. Dutch education umbrella organizations such as Kennisnet and SURF have therefore established firm frameworks for data processing.

 In addition, European legislation imposes strict restrictions on educational applications. To understand which obligations and prohibitions have applied to high-risk systems since this year, consult the [AI Act timeline showing what actually took effect in 2026](https://nieuws.llmnet.nl/en/ai-act-tijdlijn-wat-er-in-2026-daadwerkelijk-in-werking-is-getreden). Under the AI Act, AI systems used for admission, selection, evaluation of learning outcomes, or monitoring behavior during exams explicitly fall into the category high-risk.

 This means that institutions deploying such software must meet strict requirements regarding transparency, human oversight, data quality, and non-discrimination. Systems that attempt to analyze pupils' emotions in the classroom are even prohibited under the regulation. Educational institutions must therefore conduct a rigorous Data Protection Impact Assessment (DPIA) before a new AI tool is rolled out more broadly.

 
## Technical infrastructures: local models and API management

 To remain independent of large American cloud providers and comply with privacy legislation, universities and universities of applied sciences are increasingly experimenting with their own infrastructures. By running open weights on the institution's own servers or via SURF computing facilities, student data always stays within the institution's own network boundaries.

 For many educational tasks — such as generating grammar exercises, summarizing lecture notes, or checking code — the very largest models are, moreover, unnecessary and too expensive. By deploying [small language models (SLMs)](https://nieuws.llmnet.nl/en/kleine-taalmodellen) institutions can handle hundreds of simultaneous student sessions locally and cost-efficiently with relatively modest hardware.

 When an educational institution does use commercial APIs for advanced research tasks, this almost always happens through a centralized gateway. Anyone wanting to see how such an intermediary layer anonymizes prompts, monitors budgets per faculty, and automatically switches between providers can find technical details in [the explanation of LLM API aggregators](https://api.llmnet.nl/en/aggregator-uitleg). Such an architecture prevents teachers from having to link their own credit cards to individual suppliers, and guarantees that all data traffic complies with the data processing agreements in place.

 # Voorbeeld: Privacy-vriendelijke prompt wrapper voor onderwijstoepassingen
def sanitize_and_route_student_prompt(raw_prompt: str, user_role: str) -> dict:
 """
 Verwijdert potentiële PII (leerlingnummers, namen) en routeert
 naar het juiste lokale of goedgekeurde model.
 """
 sanitized_text = strip_student_identifiers(raw_prompt)
 
 # Socratische systeemprompt toevoegen voor educatieve context
 system_instruction = (
 "Je bent een didactische tutor. Beantwoord nooit direct "
 "met het eindantwoord, maar stel gerichte vragen die de "
 "student aanzetten tot logisch redeneren."
 )
 
 return {
 "model": "education-slm-local-q4",
 "messages": [
 {"role": "system", "content": system_instruction},
 {"role": "user", "content": sanitized_text}
 ],
 "temperature": 0.3
 }

 
## From passive assistants to agentic workflows for teachers

 The latest development in education concerns the shift from simple question-and-answer interfaces to autonomous AI processes that support entire administrative chains. Where a teacher previously entered text manually to turn it into multiple-choice questions, composite agents can now independently retrieve source material, check learning objectives, formulate exam questions according to Bloom's taxonomy, and place them directly in the digital testing system.

 This shift is closely linked to the rise of [agentic AI and autonomous systems](https://nieuws.llmnet.nl/en/agentic-ai), in which models independently call tools and carry out multi-step reasoning. For teachers with a heavy workload, this means significant relief from repetitive tasks, such as categorizing submitted homework, generating initial draft feedback on spelling and sentence structure, and scheduling individual remediation talks.

 The risk of these automated workflows is that teachers rely too heavily on the generated intermediate steps. Ultimate didactic responsibility always remains with the human teacher; the concept of 'human-in-the-loop' is not a theoretical recommendation in educational contexts, but an absolute pedagogical necessity.

 
## Digital inequality and the divide between students

 An often underexposed issue is the growing inequality of opportunity created by AI. There is a considerable quality gap between freely accessible basic models and advanced, paid subscription services with larger context windows, more advanced reasoning capabilities, and multimodal functionality. Students with sufficient financial means can afford premium models that perform considerably better on complex mathematical calculations, programming tasks, and literature analyses.

 To counter this divide, student organizations and policymakers advocate for institutional licenses. Similar to the collective availability of scientific journals and office software, more and more universities and universities of applied sciences are considering giving all enrolled students access to a standardized, secure, and advanced AI environment. This levels the playing field and prevents academic success from becoming partly dependent on a student's personal financial means.

 
## Policy frameworks: banning, regulating, or embracing

 The institutional approach to AI in Dutch education has gone through three distinct phases in recent years. Initially, institutions reacted with reflexive bans and panic over exam fraud. This was followed by a phase of passive regulation, in which institutions published vague codes of conduct urging students to use the technology 'with integrity' without offering concrete guidance.

 In 2026, we see a mature third phase: purposeful integration. Educational programs explicitly define which AI competencies students must master for their future professional practice. After all, a modern lawyer, journalist, software developer, or economist cannot function in the labor market without a deep understanding of the capabilities and limitations of generative systems.

 Curricula therefore structurally address prompting skills, critical verification of output, recognizing statistical bias, and ethically accounting for the sources used. AI has thus shifted from a threat to education into an integral part of the academic and professional literacy being assessed.

 
## Outlook: structural integration into the curriculum

 The integration of AI in education is no longer in an experimental phase. Where the discussion previously centered on detection and prohibitions, the focus is now on didactic innovation, robust assessment, and the safe setup of technical provisions. Schools and universities that succeed in this transition do not treat AI as a replacement for the learning process, but as a catalyst for deeper understanding, provided it is embedded in clear ethical and legal frameworks.