This document discusses using AI to make educational content more accessible. It notes that a textbook adoption requires making 75,000 images and 1 million equations accessible, which would take a human 694 days but an AI system just 14 hours. The AI system can automatically remediate documents to align with accessibility standards by adding contextual information and optimized alt text to images. However, AI also has limitations in fully understanding context and diversity that require human judgment. The document promotes learning more about using AI to improve accessibility.
3. Confidential & Restricted
We Have a Problem
This textbook course adoption is conditional on the content being accessible
And we need 75000 images and 1 million equations made accessible
Human
694 days
Multiple operators
Commercially
unviable
V
accessibilityInsight AI
14 hours
Consistency
Economic
4. Putting Context into Content
read-aloud
Documents contain visual cues and contextual information – not obvious to read-aloud tools
This session will focus on one area where AI (with apologies to Neil) has successfully been used to automate the production of alt text.
I will also pull the curtain back on the underlying AI tools that help us more generally as well as in this specific workflow, and where they stop, but also where they support anyone producing or striving to produce high quality compliant content.
Fragmentation: we all know this but it’s worth reminding ourselves that the production process for books and articles however generic is fragmented as a result of content type and materials (types of images etc), source, complexity.
That fragmentation is most vividly presented when dealing with making the material accessible. Earlier today Caroline talked about the challenges of writing good alt text. Alt text requirements added to complexity, genre, market requirements (short/med/long alt text) ensure workflows are fragmented.
Fragmentation introduces inconsistency into workflows designed to iron inconsistencies out. The writing of alt text introduces a potentially wide range of different people into the workflow compounding the challenges for deadlines and consistency in production cycles.
Hence the appeal of AI to establish control, consistency and efficiencies. But don’t think AI will address all of these challenges, or not yet anyway…
This case study was both the birth of our AI-enabled platform and of our most completely automated workflow.
A client with a course adoption had to have all the content compliant: 80 workbooks, 75k images, 1m equations.
Existing tools not fit for purpose.
If a human had embarked on the project: 2 years vs multiple operators with impact on consistency etc
Now 1200 equations a minute; automated for HTML, PDF, XML, ePub
The starting point: an unstructured PDF rendered incomprehensible via a reading device, just a series of non-distinguished lines of text, with the context we understand from the visual clues and cues absent.
Our AI reads this page, identifies the elements and imposes structure and context.
Resulting in a structured page with hierarchy that if consumed via an assistive device would now be comprehensible.
That automated structuring process not only identifies the elements in a document but extracts any “object” that might require alt text and creates a database for alt text to be added.
At this point in the process an alt text writer can create alt text within context. Anyone overseeing a project can set SLAs for writers, review status reports on the alt text production stage of any project. Alt text written into the software is automatically attached to the document. Alt text written on an exported Excel sheet is ingested and the text tagged. No author can upload alt text until the stage is completed.
The reading order of a text can be edited and refined and tagging checked for an optimized remediation process.
This is AI supporting a focus on alt text and text editing excellence.
And this is AI managing the process I described at the head of this presentation. Where the tool provides access for people to edit and write alt text, it autogenerates alt text for maths equations.
This is the output spreadsheet with the alt text beside the equations
What doesn’t AI do?
It can liberate operators to do their best work creating high quality AAA standard WCAG compliance by automating the heavy lifting
It can’t address the diversity of the workplace, particularly in tech, to help foster a more inclusive environment, to put inaccessibility in front of people day in, day out
It can’t educate and create a sustained culture that integrates digital accessibility into every piece of work that is produced
But it might give us the space and time to do those things
What else might it do?
Caroline talked about the use of image metadata in the production of alt text.
The image on the screen – 4 well known faces reconfigured by AI – illustrates how synthetic media, which literally creates an image from vocal or textual direction, might show us one direction of travel for AI, with images created defined by the alt text (to some degree) and digitally available with a core piece of alt text a human can then refine
CONCLUDING REMARKS