How AI Is Transforming Document Accessibility
The Old Way: Check, Then Fix by Hand
Traditional accessibility tools are validators. They scan a document, produce a list of issues, and leave the fixing to you. That workflow looks something like this:
- Run a checker (Adobe Acrobat, PAC, axe-core)
- Get a report: "Image on page 3 is missing alt text"
- Open the document, find the image, write alt text
- Run the checker again
- Repeat for every issue in every document
For a 50-page PDF with 20 images, 15 tables, and a broken heading hierarchy, this is hours of manual work. For an organization with 10,000 documents, it's impossible to do manually within any reasonable timeline.
What AI Actually Changes
AI doesn't just find problems — it generates solutions. Here's what that looks like in practice at Lovelace.
Alt Text Generation
When our pipeline encounters an image, we send it to Claude along with the surrounding text for context. The model doesn't just say "this is a bar chart." It describes what the chart shows:
"Bar chart comparing Q3 2025 enrollment figures across five departments. Computer Science leads with 2,847 students, followed by Business Administration at 2,103. Biology, Psychology, and English round out the top five with 1,856, 1,742, and 1,398 students respectively."
That's the difference between technically having alt text and actually conveying the information to a screen reader user.
For decorative images — backgrounds, dividers, design elements — the AI identifies them as decorative so they can be marked to be skipped by screen readers entirely. No human needs to look at a gradient background and decide "that's decorative."
Heading Hierarchy Analysis
A common accessibility failure is broken heading hierarchy. Authors use heading styles for visual formatting — making text big and bold — without thinking about the semantic structure. You end up with documents that jump from H1 to H3, have multiple H1 tags, or use H4 for something that should logically be an H2.
Our AI doesn't just flag these issues. It analyzes the content and suggests a hierarchy that makes semantic sense. It understands that "Chapter 3: Results" should be at the same level as "Chapter 2: Methods," even if the original document formatted them differently.
Reading Order Verification
PDFs don't inherently have a reading order that matches what you see visually. A two-column layout might read left-column-top to right-column-top to left-column-bottom, which makes no sense to a screen reader user.
AI can analyze the visual layout and content flow to determine the logical reading order, then verify that the tagged structure matches. This is something that's nearly impossible to automate with rules alone — it requires understanding of what the content means.
Table Structure
Tables are one of the hardest accessibility challenges. A sighted user can glance at a table and understand the relationship between headers and data. A screen reader user depends entirely on proper markup: header cells marked with <th>, scope attributes pointing in the right direction, and a caption explaining what the table contains.
Our AI analyzes tables to determine which cells are headers, whether the table has a row header, column header, or both, and generates a meaningful caption. For complex tables with merged cells, it flags them for human review rather than guessing wrong.
The Scoring System
Raw pass/fail isn't useful when you're managing thousands of documents. You need to know: how bad is it? Where should I focus first?
Lovelace scores every document on a 0-100 scale across seven weighted categories:
| Category | Weight |
|---|---|
| Document Structure | 25% |
| Images & Media | 20% |
| Tables | 15% |
| Tagged PDF Structure | 15% |
| Forms | 10% |
| Color & Contrast | 10% |
| Navigation | 5% |
A document scoring 95+ gets an A grade and is considered compliant. A document scoring below 50 gets an F and needs significant remediation. The category breakdown tells you exactly where the problems are — if your score is dragged down by images, you know you need alt text work. If it's document structure, you need heading fixes.
For batch processing, you get aggregate scores across your entire document library: compliance rate, score distribution, category heatmaps showing which types of issues are most prevalent, and a prioritized remediation queue.
What AI Can't Do (Yet)
We're not claiming AI solves everything. There are cases where human judgment is still essential:
- Complex infographics with dense information that requires domain expertise to describe accurately
- Mathematical notation in specialized fields where the AI might misread symbols
- Cultural context in images where the description needs sensitivity to audience and purpose
- Legal and regulatory documents where the specific wording of accessibility remediation has compliance implications
For these cases, our pipeline flags them for human review with AI-generated suggestions as a starting point. The human expert refines rather than creates from scratch — which is still dramatically faster than the traditional workflow.
The Math Problem
STEM documents deserve special mention because they're severely underserved by current accessibility tools. A PDF with embedded equations is a nightmare for screen readers. The equation E = mc^2 might render as an image with no alt text, or as a string of characters that a screen reader pronounces as "E equals m c superscript 2" — which works for simple cases but falls apart for anything complex.
Lovelace uses Math OCR to extract equations and converts them to MathML, the standard markup language for mathematical notation. MathML-aware screen readers can then speak the equation naturally, and users can navigate within it (moving through numerator, denominator, exponents).
This is a huge unlock for universities with STEM programs trying to make their course materials accessible.
Scale Is the Real Challenge
The technical problem of making one document accessible is largely solved. The hard problem is doing it for 10,000 documents with a team that has other responsibilities and a deadline that's approaching fast.
That's what we're building for. Not a tool that helps one specialist fix one document — a platform that processes thousands of documents through a pipeline, scores them, generates accessible output, and gives you a dashboard showing exactly where you stand and what to fix next.
The AI doesn't need to be perfect on every document. It needs to be good enough on most documents to reduce the human workload from "impossible" to "manageable."
Moss