AI helps educational publishers most by describing, organising and adapting the resources they already have. It can map each page, activity, poster or audio track to the curriculum learning outcomes it supports, produce consistent metadata, help reuse content in new formats and speed up accessibility work. Subject editors and teachers still review the results, because a wrong curriculum tag sends a resource to the wrong classroom.

This guide covers how AI-assisted curriculum alignment works, the metadata standards worth knowing (LRMI and Schema.org, and IEEE LOM), reuse and accessibility, and what guidance from UNESCO and national education departments says about quality and human review.

Key takeaways

  • Curriculum alignment is the most valuable metadata an educational publisher holds, and it has to be redone when the curriculum changes.
  • AI mapping is reliable only when the model chooses from the official list of learning outcomes, never when it writes its own.
  • LRMI and Schema.org properties such as teaches, assesses and educationalLevel give alignment a standard, machine-readable form.
  • Reusing content in new formats works best when generated material keeps a link to its source pages and their curriculum mapping.
  • Guidance from UNESCO and from education departments in Ireland and England keeps people in charge, and the Irish and English guidance both require AI output to be checked before use.

Why curriculum alignment is the core problem

Teachers look for resources by subject, class or stage and learning outcome. Publishers' catalogues are usually organised by series, title and page. Curriculum alignment metadata bridges the two: it records which outcomes each part of a resource teaches or assesses, so teachers can find it and platforms can recommend it.

The difficulty is that curricula change, and the back catalogue does not update itself. Ireland is a clear example. The Primary Language Curriculum was introduced in 2019 for all four stages, from junior infants to sixth class. The Primary Curriculum Framework followed in March 2023 and the Primary Mathematics Curriculum in September 2023. Circular 0067/2025, dated 22 September 2025, marked the approval of five curriculum specifications, completing a redevelopment that builds on the 1999 Primary School Curriculum. Schools are introducing the redeveloped curriculum gradually over at least six years: from 2026/2027 schools get focused support in one new curriculum area each year, with each area taking two years to enact, so all areas are due to be fully in place by 2031/2032, or 2032/2033 for schools that take an optional consolidation year.

The new structure also regroups subjects. Social and Environmental Education brings together History, Geography and learning about religions, beliefs and worldviews; Wellbeing incorporates Physical Education and Social, Personal and Health Education; and STEM Education combines a Science, Technology and Engineering specification with the Primary Mathematics Curriculum. A resource tagged "Geography" under the old curriculum needs a new home, and for several years schools will work across old and new structures at once.

General-purpose AI does not know your curriculum by default. Guidance from England's Department for Education warns that AI returns results based on its training data, which may not be specific to the curriculum, and may not produce results comparable with a human-designed resource developed for it.

How AI curriculum mapping works

A dependable mapping process constrains the model to real curriculum entries and keeps evidence for every decision.

  1. Load the official framework as structured data. Each row holds subject, language, stage or class, strand, unit and the learning outcome text, plus a stable ID. Remove duplicate rows that appear across tabs or files.
  2. Split resources into the units teachers use: a page, an activity, a chapter, a poster or an audio track. Long files need sensible boundaries; PDF, XML, IDML, InDesign and EPUB: What Publishing Files Contain explains what each format gives you to work with.
  3. Retrieve the candidate outcomes for the resource's subject, stage and language, so the model only ever chooses among real rows.
  4. Ask the model to choose the best-matching outcomes, usually one to three per unit, ordered by relevance, with a short quote from the resource as evidence.
  5. Validate every returned ID against the lookup and reject anything that is not in it.
  6. Leave non-teaching pages unmapped. Title pages, contents pages and introductions should carry no outcomes rather than a forced match.
  7. Route uncertain mappings to a subject editor, and sample-check the rest.
  8. Roll up unit-level mappings into title-level metadata, such as the most frequent outcomes across a book.

A finished mapping record might look like this (a generic illustration):

Field Example Why it matters
Resource unit Workbook, page 14 Teachers need the exact page or activity
Subject and stage Maths, Stage 3 (3rd and 4th classes) Filters out outcomes for other classes
Strand and unit Measures: Measuring Groups related resources
Outcome ID Taken from the official lookup Proves the outcome exists
Evidence "Measure the length of your desk in centimetres." Lets a reviewer check quickly
Review status Accepted by subject editor Separates checked from unchecked

Common mistakes

  • Letting the model write outcome text. Plausible but invented outcomes are the most damaging error, because they look correct.
  • Matching the right topic at the wrong stage. Fractions appear across several stages; the outcome must match the class the resource is for.
  • Ignoring language. Where a curriculum has separate outcomes for different languages or language-learning contexts, filter by them.
  • Over-tagging. Ten outcomes on one page makes search results noisy and teachers stop trusting them.
  • Mapping once. When a curriculum is revised, the whole catalogue needs remapping, so keep the pipeline repeatable.

Accuracy should be measured against a hand-checked sample for each subject and stage; Evaluating AI Tagging: Precision, Recall and Human Review sets out the method.

Metadata standards for learning resources

Two families of standards are most relevant to publishers.

LRMI and Schema.org. The Learning Resource Metadata Initiative was created by the Association of Educational Publishers and Creative Commons through a project that ran from 2011 to 2014. Schema.org adopted LRMI in April 2013, and stewardship moved to the Dublin Core Metadata Initiative (DCMI) in 2014. LRMI introduced the AlignmentObject type, which describes an alignment between a learning resource and a node in an educational framework using alignmentType, educationalFramework, targetName, targetDescription and targetUrl. Schema.org's LearningResource type, used alongside a primary type such as Book or VideoObject, now offers direct properties including teaches and assesses, and Schema.org advises against using educationalAlignment where one of those simpler properties can express the alignment.

What you need to record Schema.org / LRMI property Example value
Outcome the resource teaches teaches A defined term for the outcome
Outcome it assesses assesses A defined term for the outcome
Level or stage educationalLevel 'Stage 3'
Typical learner age typicalAgeRange '8-10'
Kind of resource learningResourceType 'worksheet'
Purpose in teaching educationalUse 'group work'
Time needed timeRequired 'PT30M' (ISO 8601)
Other alignments, such as reading level educationalAlignment with an AlignmentObject alignmentType: 'readingLevel'

A short JSON-LD example for a hypothetical workbook page:

{
  "@context": "https://schema.org",
  "@type": ["CreativeWork", "LearningResource"],
  "name": "Measuring length in centimetres",
  "isPartOf": { "@type": "Book", "name": "Example Maths Workbook 3" },
  "educationalLevel": "Stage 3",
  "learningResourceType": "worksheet",
  "teaches": {
    "@type": "DefinedTerm",
    "name": "Example outcome text copied from the official framework",
    "termCode": "MATHS-S3-MEA-04",
    "inDefinedTermSet": "https://example.org/curriculum/maths"
  }
}

IEEE LOM. The IEEE Standard for Learning Object Metadata, IEEE 1484.12.1-2020, is the active revision of the 2002 standard and was published in November 2020. It groups a learning object's characteristics into nine categories: general, life cycle, meta-metadata, technical, educational, rights, relation, annotation and classification. Schema.org markup suits web pages and catalogue feeds; if a platform or repository you supply asks for LOM records, the same information can be exported into LOM's categories. The practical approach is to keep one internal record and export to whichever form each platform needs. Metadata for Publishers: Descriptive, Structural and Administrative covers the wider metadata picture.

Reusing existing content in new formats

A mapped archive is a reusable one. Once every page and activity carries subject, stage and outcome labels, the same material can be assembled into practice sets for one outcome, revision summaries, teacher planning notes, differentiated versions of an exercise, or a search tool that lets teachers find every resource for an outcome across the whole list.

Government investment in England follows the same logic. The Department for Education describes a content store pilot, funded by the Department for Science, Innovation and Technology, which aims to make available the underpinning content and data needed for great AI tools. The department has also funded Oak National Academy to develop AI tools for teachers, and Oak has launched an AI lesson assistant, Aila. The premise is that AI tools for teachers work better when grounded in curriculum-specific source material.

Guardrails for reuse:

  • Keep provenance. Every generated item should link back to the source pages it was built from.
  • Inherit, then re-check, alignment. A quiz built from a page usually shares its outcomes, but an editor should confirm it.
  • Hold house standards. Reading level, terminology, spelling conventions and tone should match the series.
  • Check answers and worked solutions by hand; errors in answer keys travel quickly.
  • Track status. Record whether each item is draft, reviewed or published.

Accessibility

Accessibility is both a quality requirement and a metadata task. The W3C's EPUB Accessibility 1.1 Recommendation requires every EPUB publication to include Schema.org accessibility metadata describing its access modes, accessibility features and hazards (accessMode, accessibilityFeature and accessibilityHazard), and recommends a human-readable accessibilitySummary. To conform, a publication must meet at least WCAG 2.0 Level A; the specification strongly recommends the latest WCAG 2 version and Level AA, and conformance is judged across the whole publication rather than file by file.

AI can speed up several parts of this work:

  • drafting alt text and longer descriptions for diagrams, charts, maps and photographs;
  • producing transcripts and captions for audio and video;
  • checking heading structure and reading order in converted files;
  • proposing accessibility metadata based on what a file actually contains.

Each needs expert review. In educational material a description has a teaching purpose: alt text for a maths diagram should give learners what they need to attempt the task without giving away the answer, and a map description should match the skill being taught.

Quality, teacher review and guidance on AI in education

Published guidance on AI in education points the same way. UNESCO's Guidance for Generative AI in Education and Research, launched in September 2023 as the first global guidance of its kind, calls for a human-centred approach. UNESCO noted that fewer than 10% of more than 450 schools and universities in its survey had institutional policies or formal guidance on generative AI, and that publishing a new textbook requires more authorisations than using generative AI tools in the classroom. Publishers' editorial review processes are part of what makes their resources trustworthy.

Ireland's Department of Education and Youth published Guidance on Artificial Intelligence in Schools in October 2025. It says teachers and school leaders should always review, verify and adapt AI-generated content, warns that AI can reflect bias, and keeps teachers and school leaders at the core of educational decisions. England's Department for Education guidance, updated in August 2025, warns that generative AI content can be inaccurate, biased, out of date or low quality, and says any content produced requires critical judgement to check for appropriateness and accuracy.

For publishers, that translates into a review plan with named owners:

Check Who When
Curriculum mapping accuracy Subject editor Sample per subject and stage, every batch
Generated text and answer keys Subject editor Every item before release
Bias and representation Editorial team Sample, plus anything flagged
Accessibility Accessibility specialist Every title
Classroom fit Practising teachers Pilot before wide release

Keep a log of disagreements between reviewers and the model. Repeated corrections usually point to a gap in the outcome lookup, an unclear prompt or a resource type the pipeline handles badly, and fixing the cause is cheaper than correcting the same error a thousand times.

How Druid Learning helps

Druid Learning is an AI archival processing platform, based in Dublin and New York, that turns content archives into AI-ready datasets. For CJ Fallon, the Irish educational publisher, it processed 50,000 resources and mapped them to the curriculum. Druid processes formats including PDF, XML, IDML, INDD, EPUB, MP3, WAV and MP4, and Archive Optimization turns an archive in any format into an indexed, labelled, AI-ready dataset. The Suggestion Engine generates content and suggestions in a publisher's own voice and curriculum. Read the CJ Fallon case study.

Frequently asked questions

What is curriculum alignment in educational publishing?

Curriculum alignment is metadata that records which official learning outcomes a resource, or part of a resource, teaches or assesses, along with its subject and class or stage. It lets teachers find material by outcome and lets platforms recommend it. Good alignment points to real entries in the official curriculum, uses stable identifiers, and is recorded at the level teachers use, such as a page or activity.

Can AI map educational resources to learning outcomes accurately?

AI can map resources to learning outcomes quickly and consistently when it is given the official list of outcomes for the right subject, stage and language and asked to choose among them with evidence. Accuracy should be measured against a sample checked by subject editors. Allowing a model to write its own outcome text, or to match across stages, produces errors that look plausible and are hard to spot.

Should educational publishers use LRMI or IEEE LOM?

Usually the answer is whichever your platforms ask for, generated from one internal record. LRMI, now part of Schema.org through properties such as teaches, assesses and educationalLevel, suits web pages and catalogue feeds. IEEE LOM, revised as IEEE 1484.12.1-2020, organises metadata into nine categories and may be requested by repositories or learning platforms. The hard work is the underlying alignment data; export formats can follow.

Does AI-generated learning material need teacher review?

Yes. Guidance from Ireland's Department of Education and Youth says teachers and school leaders should always review, verify and adapt AI-generated content, and England's Department for Education says any AI-produced content needs critical judgement to check accuracy and appropriateness. For publishers, that means subject-editor review of every generated item and answer key before release, plus classroom piloting for new resource types.

What should a publisher do when the curriculum changes?

Load the new curriculum as structured data with stable outcome IDs, then remap the catalogue against it, keeping the old mapping alongside while schools transition. In Ireland, for example, the redeveloped primary curriculum is being introduced over at least six years, so resources may need both old and new alignments for some time. A repeatable mapping pipeline makes later revisions much cheaper.

Sources