In brief

From publicly documented workflows, we formulated twelve needs as working hypotheses. They help us prepare interviews, use cases and technical experiments.

The direction of reasoning matters: an observed problem can support a plausible needs hypothesis, but it does not make that hypothesis a validated requirement. These points now need to be tested with the people who actually perform and use the work.

A. Preprocessing and subject semantics

1. Accelerate structural preparation

Document structure, reading order, tables, formulae, task references and figures should be machine-prepared so that specialists spend less time on mechanical reconstruction.

2. Capture the semantics of visual STEM content

Diagrams, circuits, plots and other STEM graphics should not merely be described. Objects, relations and subject structure need to be represented explicitly.

3. Enable efficient expert review of AI output

Professional converters need traceable proposals, visible uncertainty and targeted correction tools rather than black-box output.

4. Generate several outputs from one reviewed basis

Once subject meaning has been reviewed, it should ideally form a shared basis for text, mathematical, audio, tactile or visual representations.

B. Production economics and reuse

5. Make large workloads manageable

Textbooks and material packages need job control, prioritisation, progress tracking and recovery, not only single-file upload.

6. Reuse existing conversions

Before new work begins, organisations should be able to determine whether a work, section or similar content has already been converted or is currently being converted.

7. Detect relevant changes between editions

A new version should not automatically trigger complete reconversion. The important parts are those that changed semantically.

C. Quality, standards and secure processes

8. Process examinations securely and controllably

Confidential material needs restricted access, controlled storage, auditability and binding human approval.

9. Apply standards and institutional rules reproducibly

Braille, mathematics and other conversion rules, together with local conventions, should be applied consistently and traceably.

10. Preserve provenance and quality status

For each output, it should be possible to see which source and version it came from, which automated steps were applied and who reviewed it.

D. Personalisation and shared learning

11. Support individual usage needs

Especially in higher education, assistive technologies, notation, level of detail and accommodations vary. A common subject representation should support different output profiles.

12. Support accessible collaboration between blind and sighted learners

Longer term, accessibility should not mean separate special material. Visual and non-visual representations could instead be derived from the same reviewed semantics.

Why these twelve points belong together

Together, these hypotheses describe something quite different from a simple “upload a file and get alt text” service. Professional accessible-media production is a production and quality workflow: material arrives in different forms, must be understood, reviewed, converted into several media, distributed and later possibly reused or updated.

IncluLearn.AI’s semantic-first approach fits this model because it treats subject meaning as an explicit, reviewable intermediate layer.

What needs validation next

In interviews, we do not simply want to ask whether these twelve ideas sound useful. We need to learn which of them matter in daily work, how often they occur, and which quality or security conditions determine whether an organisation would use such support at all.

A valuable outcome may therefore be that some hypotheses are rejected, merged or assigned a lower priority.