Accessible school materials with AI assistance

Learn in your browser · Chapter 01 of 06

Start with the intended learning and an observable barrier. Preserve meaningful choice without inferring a diagnosis.

AI can help draft a different presentation of a lesson, but the teacher must preserve what the learner is meant to understand. Accessibility is about removing barriers to an equivalent task, not automatically making every task easier or replacing the learner’s reasoning. This course follows the fictional Linden Geography Resource Team. Its lesson uses an invented grid map, two routes and three invented rainfall totals. There are no real children, learning records, diagnoses or school locations. Adults practise adapting the resource before deciding whether any material is suitable for a real class.

Linden’s learning contract LC1 asks learners to use a stated scale, compare two supplied route lengths, describe relative position and read a rainfall chart with units. It does not assess eyesight, mouse precision, decoding a particular font or unaided memory. The teacher retains the geographical concept and the evidence required for success while offering different ways to access the same information. If the actual lesson assessed visual symbol recognition, replacing every symbol with its answer would change that task; the teacher would first need to decide which support preserves that specific objective. A generic request to make a worksheet accessible cannot resolve the instructional decision.

Scroll sideways to see every column.

Describe the barrier you can observe before selecting an adaptation. “The diagram labels become unreadable when enlarged” is an inspectable problem. “This learner must be dyslexic because they skip a line” is an unsupported inference. A person can need a different format for many reasons, including temporary conditions, device constraints, language or preference. Ask which information or action is difficult and what route the learner prefers, through the setting’s established support process. Do not turn the answer into a diagnosis, a ranking of ability or a permanent profile. A preference for audio does not establish that all future teaching should use audio only.

For this exercise, an AI drafting prompt contains LC1 and only the fictional resource. It excludes names, support plans, work samples and classroom behaviour. Deleting names from a real document is not proof that it is anonymous; combinations of context can remain identifying. The course does not authorise sending real pupil data to an AI service. If a real staff task later needs such information, the responsible school must make its own approved handling decision. A local editor and a manually written template are useful alternatives. There is no need to use an AI service to complete any task in this package.

Assign the content owner, material adapter and reviewer before drafting. The content owner establishes LC1 and the reference facts. The adapter proposes a changed presentation and records what changed. The reviewer compares the exact candidate with the sources and observes the intended tasks. Learners and assistive-technology users can reveal barriers that a checklist misses, but one person’s successful attempt cannot represent all users. Specialist communication methods, such as an individual’s braille conventions, tactile presentation or symbol system, require informed human discussion rather than a model declaring an automatic conversion suitable. This fictional course does not manufacture a specialist approval.

Worked example: a source card says “Compare the green and red route on the map.” At a monochrome printout the colour difference is lost, so the learner cannot know which route is being discussed. Linden replaces the instruction with “Compare Route A, shown with a solid line, and Route B, shown with a dashed line. Their grid-unit totals are also listed in the route table.” The goal remains scale calculation; the identifiers become available without colour. Verification checks that the route names, line styles and table rows agree. Changing the instruction to “Choose the safe route” would introduce an unsupported safety judgement and a new objective.

Source context checked 12 October 2026: the Department for Education guidance on generative AI in education applies to England and was updated on 12 August 2025. It retains professional responsibility for the appropriateness and accuracy of produced material. This course uses that bounded principle, not a claim of school approval. W3C WAI guidance on involving users supports combining user experience with evaluation. Neither source makes this invented lesson universally accessible. The technical examples below draw on selected W3C guidance; they are not an audit of every WCAG criterion or a determination of local accommodation duties.

Learn in your browser · Chapter 02 of 06

A clearer presentation must carry the same spatial relations, units and learning task. Compare claims against a frozen source.

Freeze the source before judging a rewrite. Linden’s source packet S1 defines an abstract grid, not a real walking route. East is increasing x and north is increasing y. One grid unit represents 0.5 kilometres. The School is at (0,0), the Library at (2,2) and the Park at (4,2). A river runs west to east along y=1. Route A goes north two units from the School, crossing the west bridge at (0,1), then east two units to the Library. Route B goes east four units, north two through the east bridge at (4,1) to the Park, then west two to the Library. No terrain, travel-time or safety information is supplied.

Scroll sideways to see every column.

A meaning ledger is a small table linking each essential claim to its source and proposed wording. For this lesson the critical rows cover orientation, coordinates, scale, route sequence, rainfall values and units, and the task instruction. Mark a row preserved, changed, omitted or unsupported; do not average away an important defect with many correct words. A rewrite that preserves six of seven rows but changes millimetres to centimetres cannot be accepted merely because its preservation percentage is high. Some wording changes are harmless; others alter the relationship the learner must reason about. Review the meaning, not just shared keywords.

Compare two supplied reading adaptations. Draft A says: “The Library is north of the School. Take the red path, which is the safe route. Each square is 0.5 metres. May has 35 centimetres of rain. Choose the best route.” Draft B says: “The Library is two grid units east and two north of the School. Each grid unit stands for 0.5 kilometres. Route A has four grid units; Route B has eight. Use the scale to calculate both distances. Compare your results. The fictional rainfall totals are April 20 mm, May 35 mm and June 25 mm.” B uses shorter statements without changing the essential comparison or supplying the calculated distance answers.

Draft A has several different errors. North alone loses the east component. Safe has no source support. A colour-only path name is unusable when the colour distinction is unavailable. Metres replace kilometres, a thousand-fold unit change. Centimetres replace millimetres, a ten-fold unit change. “Best” leaves the success criterion unspecified. A responsible repair does not simply ask the model to simplify again. Linden supplies the specific ledger defects, requests a new candidate that preserves the units and objective, then compares every essential row afresh. Repeating a model’s assurance that it retained meaning is not verification.

Shorter is not automatically clearer. Keep indispensable terms and explain them when they carry the concept: a grid unit is a distance step on this map, while a kilometre is the distance unit obtained using the scale. Replacing both with “step” would conceal the conversion. Chunk instructions into a sequence, use a consistent name for each place, put the scale near the relevant data and avoid unexplained pronouns. A lower reading-age label from an automated tool is at most one rough signal; it does not establish comprehension, age suitability or preservation of the task. No learner is assigned a simplified version automatically.

Worked answer: A is 4 × 0.5 = 2 km. B is 8 × 0.5 = 4 km. The distance difference is 4 - 2 = 2 km. B has twice A’s distance in this grid model. These results compare stated lengths only; they do not prove which route would be safer or quicker. Rainfall totals sum to 20 + 35 + 25 = 80 mm. May exceeds April by 15 mm, not by 15 percent. If a percentage were requested relative to April, 15/20 × 100 would be 75 percent, but LC1 does not require it. Distinguishing an absolute difference from a percentage is part of preserving the wording of a question.

Label the original S1, the adaptation version and the reviewer’s disposition. If Draft B v2 is checked and v3 changes the scale to 0.8 km, the earlier review does not cover v3, even if the layout is unchanged. A hash can bind the reviewed bytes to a file, but it cannot prove that S1 is true or that a reviewer understood it. Record unresolved matters separately rather than quietly filling them in. The release note for this fictional resource might say “LC1 facts preserved; browser and assistive-technology observations pending.” That is more useful than a green tick without a stated scope.

Learn in your browser · Chapter 03 of 06

Make task-relevant information available across formats. Describe evidence and relationships without adding unsupported judgements.

Write an image alternative for the information that matters in its context. The same picture can be informative in one lesson and decorative in another. An ornamental divider conveys no lesson content and should not interrupt reading with a long description; a route chart carries essential values. A functional image, such as an icon used as a download link, needs a name that explains the action rather than its appearance. Linden has four instructional images, I1 to I4. Each therefore needs a useful short identification and access to the relevant details, not the same generic phrase “geography image”.

I1 is the schematic route map. Its short identification is “Fictional Linden grid map with two routes from School to Library; full route and coordinate description follows.” The adjacent long description supplies orientation, scale, point coordinates, river direction, bridge positions and ordered route segments from S1. A learner can reconstruct the relationships without relying on colour. Keep the source limitation that this is a grid exercise, not real navigation. The description should not invent a steep bank, busy road or tactile feature merely because the image generator or map style suggests it. It also need not list irrelevant paper texture or decorative shadows.

I2 compares route lengths in grid units. Its short identification names the comparison and unit. The equivalent table has Route A = 4 grid units and Route B = 8 grid units. The visual bars start at zero and carry numeric labels, with distinct names available in text. State that B’s bar represents twice A’s grid length. The scale remains separately available so that every learner can calculate the kilometre values. Giving only the bar colours would omit the values; replacing the data with “A is better” would insert a judgement. Do not add the calculated answers solely to the text alternative if the visual version leaves those calculations to the learner.

I3 is a rainfall bar chart. A suitable identification is “Invented April-to-June rainfall totals in millimetres; data table below.” The table lists April 20, May 35 and June 25 mm. A concise relationship statement says May is the largest of these three totals and April the smallest. Preserve that the numbers are invented monthly totals; they are not measurements from a place or evidence that a season is getting wetter. “Rain increases” would be misleading because the series rises from April to May and falls in June. If a version has no visible unit, both the image and its alternative need repair against S1.

I4 is an instructional process image with four stages: read the stated grid-unit total, multiply by 0.5 km per unit, write the kilometre result with units, then compare the two results and state the source limitation. The adjacent ordered text list preserves this sequence. Arrows indicate order, not a claim that a computer validates the method. A description that merely lists four box labels without their relationship could hide the order needed by the task. The print version must retain the ordered list as well as the drawing. The steps describe the lesson method; they do not prove that a learner has followed it correctly.

W3C’s complex-image tutorial recommends making the detailed information available as structured text where appropriate. For these examples, use a short image name and a visible adjacent description or table. A long unstructured accessible-name string is a poor replacement for navigable data. Do not hide the only long description inside hover content, or require a pointer to reveal it. If the same facts appear nearby, identify where the equivalent can be found and avoid pointless repeated paragraphs. Test what is actually announced in the target browser and assistive technology rather than assuming that the presence of an alt attribute establishes a useful experience.

Text, captions, transcripts and specialist formats are different deliverables. If Linden later records a spoken explanation, it should provide a checked text version of the spoken content and identify essential non-speech information when relevant. A model-generated transcript can corrupt a place name or a unit, so compare it with the recording and source. This package supplies text and diagrams, not a professionally prepared braille, tactile or sign-language edition. Staff can use the source facts to commission or discuss an appropriate alternative, but the model cannot certify that an individual’s communication method has been respected. Do not simulate that review by labelling a conversion “specialist approved”.

Worked review of a proposed I1 description: “A colourful town map with a red route to a library” names the object but omits the scale, positions, sequence and task-relevant distances. Linden replaces it with the short I1 identification plus the S1 coordinate and route tables. A second checker finds that the long text says the river flows north. That defect remains blocking for this candidate despite correct HTML and clear prose, because S1 says west to east. Repair the relation, then recheck the affected description and any derivative audio or print text. The goal is equivalent evidence, not the largest number of descriptive words.

Learn in your browser · Chapter 04 of 06

Structure and controls must support the same task through different input and reading routes. Inspect actual behaviour as well as markup.

Readable HTML is more than large text. The lesson needs a useful document title, declared language, one clear main heading, meaningful section headings, a main region, named links and data tables with proper header cells. These features let people understand and navigate the structure using different tools. Visual boldness alone does not make text a heading. In the local example, a skip link leads to the lesson’s main content; headings name the source, four figures, questions and worked answers. There is no account, upload or remote connection. The page can be read locally with scripting disabled.

Use a table for the rainfall and route data because row/column relationships are meaningful. Supply a caption and explicit column headers, including the unit. In this small exercise, th elements with scope="col" identify the columns and the first cell of each data row is a th with scope="row" naming the route or month. A screen-reader user should be able to associate 35 with May and millimetres. Replacing the table with spaced text inside an image would remove that structure. Conversely, do not use a data table merely to arrange decorative layout. The W3C tables tutorial explains these header relationships; implementation still needs observed use.

The sample lesson uses a native details/summary control to reveal worked answers after the learner attempts the questions. Its label states what it opens. Native semantics reduce the amount of custom keyboard behaviour an author must invent, but they do not remove the need to test the result. A clickable div with an onclick handler is not automatically a keyboard button. The page does not require dragging a route, selecting a tiny map marker or answering within a timer. A learner can calculate on paper, then reach the same worked guide with the keyboard. Completion of that interaction is not a graded result.

For a keyboard pass, start at the page beginning without using a pointer. Use Tab and Shift+Tab to follow interactive controls, verify visible focus, activate each link or disclosure with its native keyboard convention, and ensure you can leave it. Confirm that the skip link reaches the main content and that revealing answers does not trap focus or erase the question. Record the browser and platform, exact page version, task, actual behaviour and any unexpected stop. A successful tab sequence checks this task in that setup; it does not prove all assistive technology or every criterion. A pointer screenshot alone cannot establish keyboard operation.

Check legibility and reflow with the intended content. W3C’s WCAG 2.2 Understanding guidance describes reflow for ordinary vertical content at a width equivalent to 320 CSS pixels, corresponding to a 1280-pixel starting viewport at 400 percent zoom. There are exceptions for content that genuinely requires two dimensions, such as a map or some tables; that is not a blanket exemption for surrounding paragraphs or controls. In this exercise, text wraps and the map has a full text equivalent. For ordinary text, the selected contrast criterion uses at least 4.5:1; large-scale text has a 3:1 threshold with defined size/weight conditions. Do not treat 3:1 as the rule for every label.

Never make colour the only way to identify a route or a problem. Use the visible Route A/Route B labels, distinct line patterns and the source tables. Keep focus styles visible rather than removing outlines for a cleaner screenshot. Enlarge text and inspect long words, table headers and link labels; a container can look tidy at one width while hiding content at another. Print in monochrome and verify that the route names, line differences, scale and tables remain usable. The print stylesheet should not hide the long descriptions or the only worked answers simply to make the output shorter.

A screen-reader pass is a real task, not a search for ARIA attributes. Use an available screen reader with its documented navigation commands to inspect the heading list, reach the main lesson, find an image’s name and its associated detail, navigate both tables with headers, and open and close the worked answers. Read the calculation question and obtain all the same values through this route. Record the screen reader and browser versions and what was actually announced. If the software or an experienced tester is unavailable, mark the task not run and provide the text route; do not fabricate a passing observation from source inspection.

Worked repair: a draft table contains visually bold td cells for “Month” and “Rainfall” and a mouse-only span labelled “Answer”. Linden changes the header cells to th with scope, adds the caption “Invented rainfall totals, mm”, and replaces the span with a native details/summary named “Show worked answers after your attempt”. The candidate is then checked for actual keyboard entry, announcement and return to the task. The repair makes the intended structure explicit, but only the subsequent observation can show how that browser and assistive technology expose it. The author records “structure corrected, interaction observations pending” until that happens.

Learn in your browser · Chapter 05 of 06

Use automation to find bounded defects, then test meaning and real tasks. Keep failed, assisted and unrun observations visible.

Separate three kinds of evidence: structural checks, semantic review and observed use. A script can find an absent language attribute or a missing table header. A content reviewer can detect that the river description contradicts S1. An observed task can reveal that a control is difficult to locate or that a table is announced ambiguously. None substitutes for the others. A complete checklist with no recorded observations is not a completed user test. W3C’s Easy Checks page explicitly describes a first review, not comprehensive accessibility evaluation; it is labelled a draft at the checked date. The course uses that limitation rather than selling a universal pass.

The local check_materials.py reads the supplied lesson and returns named findings. It checks explicit structural requirements for this fixture: document language and title, a main heading and region, a skip target, unique IDs, non-empty alternatives for four image files, captions and header scopes for two tables, a named native answer disclosure, and a fictional-source notice. It also compares a machine-readable reference-data block with the canonical S1 values. It performs no network request and writes no learner data. A pass means only those selected checks passed. The embedded reference block may disagree with visible prose; a declared alternative may be misleading; CSS can hide useful text. Those require other evidence.

A deliberate counterexample changes the visible sentence “The river flows west to east” to “The river flows east to west” while leaving the reference-data block unchanged. The structural checker still passes. That is not a reason to trust the sentence; it demonstrates the checker’s limited scope. A second counterexample changes only an image alternative to “Map”. The non-empty check may pass even though the description is poor. Human review asks whether the short name and adjacent details convey LC1’s information. Test outputs must be read as evidence about a particular predicate, not as a badge that the material is accurate or accessible.

Use one row per task in the supplied manual-checks.csv. Fields are check ID, file version, setup, task, expected observation, observed result, disposition and follow-up. The template deliberately leaves observed result blank. Not run, blocked and failed are different states. If a device is unavailable, record not run; if a control prevents completing the task, record failed or blocked with the reproduction steps. Do not replace a failed observation with a model-generated description of how the page ought to work. Keep any actual participant information out of this fictional exercise; record the technical setup and task without personal details.

Suppose a reviewer records four fictional observations: K1 skip link works in the tested browser; K2 answer disclosure opens and closes with the keyboard; M1 rainfall unit reads “cm”; M2 the river direction says east to west. The two keyboard observations do not compensate for the two semantic failures. Linden holds the candidate, repairs the unit and relation against S1, then repeats the affected content checks and any derivative-format checks. It retains the original observation record instead of silently marking every row passed. The repaired version gets a new identifier, so another teacher can tell which evidence belongs to which file.

Ask for task evidence and preference, not a verdict that someone is “an accessible user”. A useful prompt is “Using your chosen reading route, find the scale, calculate Route A and show where you found each value.” Observe the obstacle and the workaround with the person’s agreement through the setting’s normal process. Do not infer a diagnosis from the workaround. If the person can answer only because the facilitator supplied the missing unit, the material did not independently provide all required information. Record the assistance rather than declaring the material successful. Two people completing the task supports only those observed cases, not a conclusion about the whole class.

Prioritise a meaning error or blocked task ahead of cosmetic preference. A wrong unit changes the answer. A hidden control can prevent access to the worked guide. Uneven decorative spacing may matter for clarity but should not displace those failures. Some requests cannot be resolved by generic HTML edits: a specialist tactile key, individual communication vocabulary or specific accommodation may require a separate competent human decision. Record what remains open and provide the source facts for that discussion. This does not create a professional-approval gate for publishing a fictional educational course; it prevents the course from claiming that real individual provision has already been validated.

A pilot log is not evidence of learning simply because a page was opened. Consider six invented adult reviewer attempts: four independently found the scale and both route totals, one needed the missing unit supplied verbally, and one stopped before the task. Report four independent completions out of six attempts, approximately 66.7 percent, with the assisted and incomplete attempts separate. Reporting five out of five by excluding the stopped attempt and counting the assisted one would change both numerator and denominator. This tiny fictional sample proves neither improvement nor inclusion. To investigate improvement, define the comparison, task, assistance rules and failure handling before collecting actual authorised evidence.

Learn in your browser · Chapter 06 of 06

Bind the sources, alternative formats and observed tasks to one version. State the limits of the educational result.

An inclusive lesson package is a set of consistent views of one learning contract. Linden’s manifest names LC1, S1, the two reading drafts, the reference HTML, the four SVG images, the worksheet and manual-check template. The SVGs are original code-drawn fictional teaching figures. They are not photographs of a real school. The print stylesheet expands the worked-answer content and keeps the source tables and long descriptions. All assets are adjacent local files, with no external fonts, scripts, model or account. A downloaded folder is not complete if an image or referenced description was left behind.

The disposition is a statement with a version and scope, not a celebratory label. For example: “Candidate L2 retains LC1 facts in the reviewed HTML and print text. Selected structural tests passed. Keyboard tasks were observed in the recorded setup. Screen-reader observation was not run. No specialist-format or legal conformance claim is made.” If the evidence actually lacks the keyboard observations, remove that sentence rather than inheriting it from a template. A course author can hand over an educational candidate for independent review without claiming that a real school, specialist or learner has accepted it.

A small later change can affect several routes. If the teacher changes May’s invented rainfall from 35 to 30 mm, update S1, the visible chart, equivalent table, reading adaptation, worksheet answer and any recorded audio derived from them. The previous total of 80 mm becomes 75 mm and the May-April difference becomes 10 mm. That is a new version, so recheck the affected questions and outputs rather than retaining a passing label from the old one. The current exercise does not make this change; it uses 35 mm throughout. This worked counterfactual teaches traceability, not an invitation to alter the canonical fixture silently.

Worked package handover: the adapter supplies lesson.html, styles.css, four image files, S1, the drafts and the worksheet. The reviewer first confirms that the local links resolve inside the folder. They compare scale, route counts and rainfall values across all text views, inspect the four image alternatives, and review the distinction between source checking and learner answering. They then record the interaction checks they actually performed. A missing image file stops a complete-package claim even if its text alternative is present; a missing meaningful description stops an equivalent-information claim even if the image loads. The manifest supports both findings without hiding either.

Use the browser’s print preview or a separately generated PDF only when available and record which output was inspected. Verify that colour loss does not remove route identity, tables do not clip, descriptions and units remain visible, and the questions can be attempted before the worked answers. The sample print rules make details content available on paper, so a closed screen disclosure must not remove the answer section from the printed resource. A print stylesheet is an implementation proposal until a real rendered print output is checked. This authoring packet includes the HTML print route; a current native Academy workbook PDF and its independent visual acceptance are separate later deliverables.

This course has a 40-question knowledge bank and six ungraded written practices. A knowledge sitting draws 20 questions with a 45-minute limit and requires a raw result strictly above 80 percent: 16 correct is exactly 80 percent and fails; 17 correct passes. The written work and capstone are opportunities to practise and compare with worked answers, not automatically marked practical competence. Mickai Training Academy is not an accredited training provider. A knowledge result does not certify accessibility conformance, specialist teaching competence, an individual accommodation or permission to deploy AI in a school. Check the published course and assessment pages for current exam and certificate availability; this workbook alone does not establish a live service.

The next useful trial is a small, authorised task observation with a clear purpose, rather than a promise to make all materials accessible at once. Choose one real barrier through the school’s process, preserve the objective and agree what evidence would show that it improved. Record assistance, failure and feedback without unnecessary personal data. Keep a manual or alternative-format route available where appropriate. If a requested adaptation changes the objective, seek the teacher’s decision and describe the new task honestly. AI can suggest possibilities, but people remain responsible for curriculum meaning, appropriate support, source handling and the actual experience of using the material.