Check this page on WAVE
Skip Navigation
You are here: Home / Document Remediation Using AI
word.expert logo - I'd Like to have a WORD with you

Document Remediation Using AI

As your organization considers options for complying with Section 508, you may be intrigued by various AI remediation solutions which claim to automate the document remediation process at a lower cost than hiring a human remediator. Most people think about document accessibility as a technical issue, yet the information in a document can affect a person’s safety, health, legal rights, benefits, housing, or ability to participate in a government program. Section 508 and other ADA laws governing equal access protect the civil rights of millions of Americans with disabilities. Providing an equivalent experience of your content for all users is critical to complying with Section 508 laws and unless your documents are very simple, AI automation cannot take the place of a human remediator.

Legal Ramifications of Automated Remediation and “Checkers”

One of the most significant pitfalls of AI-based accessibility remediation is the false sense of compliance it can create. Automated tools may address surface-level accessibility issues while missing more fundamental problems that affect user experience. Legal compliance requires more than just checking boxes on automated accessibility scans.

From a legal perspective, the biggest mistake organizations make is assuming that using an automated document scanner – with or without AI – demonstrates compliance. It does not. Technical conformance to WCAG and PDF-UA standards are a requirement of Section 508, but legal exposure is usually a result of barriers that may be present even if your document passed a scanner check. The user’s inability to access an equivalent experience of your content is the basis of a lawsuit – an assertion that the document ‘passed’ a checker is irrelevant if the document does not provide an equivalent experience .  YOU, not your vendor will be held legally responsible for documents that may pass a checker but not be accessible or usable.

Pitfalls of AI Remediation

While AI can remediate simple documents fairly well and process large batches of similar documents at a much faster pace than a human remediator, be aware that there are also numerous pitfalls in using a service that prioritizes speed over accuracy and ‘passing the checker’ over true accessibility and usability.

Automated tools can find missing tags in existing content and flag certain technical problems. They can’t fully determine whether the tags are correct, the reading order makes sense, or the alternative text properly explains an image. Acrobat can miss entire blocks of content during the tagging process and this happens more often than you might think. If Acrobat can’t capture that content so it is made available to the user, neither can an AI. In fact, the tagging process used by Acrobat is the same tagging process used at scale by AI, and it contains many of the same weaknesses and failings.

AI scanning and tagging may differ from Acrobat’s tagging functions in the specific method used to generate the tags, but the outcomes are based on the same ruleset for all tagged documents – WCAG 2.1 or greater. There is no AI tagging functionality that makes AI special or better than Acrobat because no matter what tool is used to tag a document, the tags and rules for applying them remain the same. Crucially, AI makes all of the same mistakes that Acrobat makes because all machine scanning is unreliable for all but the most basic use cases. The key difference between Acrobat’s tagging and AI is simply the addition of automation at scale – a folder containing hundreds or thousands of files can be tagged at a much faster rate than a human opening each file in Acrobat and tagging it.

Limitations of AI-Based Document Remediation

  • Document suitability for AI Scanning: Documents must be text-heavy files, based on a template. Files may contain differentiated content but the layout must be identical across the group of documents.
    • All of the documents in the workflow should be of the same type, whether it is memos, invoices, simple letters, transcripts, etc. because AI “learning” is dependent on processing a large number of similar documents.
    • For the AI to become efficient and accurate, it must see many, many iterations of your templated documents as part of the “training” process.
  • Accessible Template Design: The file template should be created through a process that ensures maximum usability and compliance of the output files.
    • This necessitates that a person who is expert in both Word and Section 508 remediation be part of your template design process to avoid a ‘garbage-in, garbage-out” template workflow that will require human intervention at the end of the process.
  • Minimize Decorative Elements: Decorative elements should be avoided in the template/documents because automated systems often cannot differentiate between decorative and functional images.
    • If present, decorative elements will require human intervention during the process to ensure they are “artifacted”.
  • Simple Tables Only: Tables with subheadings, merged cells, footnoted content or other complex elements cannot be remediated by an automated process.
  • One Level Simple Lists: Lists, especially nested lists, and tables of contents are generally not well-handled by automated systems
  • No Complex images, Diagrams, Maps, or Charts: Complex images such as maps, diagrams, or charts cannot be remediated by a machine, even if they are scanned by OCR first.
    • OCR scans are also not 100% accurate and require human intervention to fix the text the OCR scan detects with low-confidence (“suspects”). If the OCR machine cannot interpret or fix the ‘suspects’, the AI machine cannot do it either. This work requires a human being who can look at the content, determine the intent, and make the correction.

AI Remediation Cannot Replace Human Remediators

AI automation cannot fully evaluate usability, context, or intent. A scanner may confirm that an image has alt text, but it cannot reliably judge whether the alt text communicates the function or meaning users need. It cannot know that there is content in the document that has not been captured by the tagging process – only human eyes can review the content and tagging to ensure that all elements are captured and tagged correctly. It cannot understand complex content such as strikethroughs for deletions and underlines for additions and will not differentiate the content to the end user, resulting in the content being read without context, creating confusion. It may create nonsensical table structures, create a reading order that does not reflect the correct flow of the document, or mishandle symbols, abbreviations, and other content that may make sense to a visual user but will be read incorrectly by a screen reader or other alternative technology. And sometimes, seemingly simple items can create accessibility problems!

As an example of “simple” content that makes sense to a visual reader but will be mangled by a screenreader, consider the phrase “lbs per day’. If you are reading this content visually, you most likely read this phrase as “pounds per day” even though that is not literally what the phrase says. You don’t even really think about the translation process going on in your mind that causes you to read “lbs” as pounds.

A screenreader does not have the capability to make these types of translations a sighted user finds effortless. A screenreader will read the phrase literally – i.e., “L…B….S….per day”. AI has no way of knowing that it should wrap that phrase in a tag and use the Actual Text function to force the screenreader to read what the phrase means – “pounds per day” and not what it says – “L…B….S….per day”.

The key takeaway is that the person using the accessibility feature relies on the tagging structure to present an accurate, complete, and equivalent version of the contents but they have no way to verify this. The burden is on the organization providing the document to ensure that the remediated document accurately represents the contents in an accessible, usable format – and it is simply not possible. A document can get a good automated score. It may still have serious accessibility problems. The fact that an automated check has been passed does not mean that equal access has been provided.

Section 508 Document Remediation by Melissa Raulston

Headshot of Melissa Morisco Raulston.

Melissa Raulston is a highly experienced document remediator, a certified Microsoft Word expert, and DHS Trusted Tester. From document template creation in Microsoft Word through the final, accessible document, Melissa has the expertise to help you achieve Section 508 compliance for your organization’s complex, high page count documents. Melissa serves a small group of select clients with the provision of on-call services that work within your schedule and deadlines. Drop me a line to discuss your document production needs!