Spotlight: Access for All — How DocAccess Advances Digital Equity in the Public Sector

[MUSIC PLAYING]

VOICEOVER: Spotlight, Access for All: How Docaccess Advances Digital Equity in the Public Sector. Speakers: Mac Clemmens (Main Presenter), CEO, Streamline. Shawn Jordison, Founder, The Accessibility Guy. Carter Temm, Accessibility Strategy and Training Lead, QualityLogic

MAC CLEMMENS: Hello, and welcome. I wanna give a special thank you to the Vista Center for putting on Sight Tech Global, and for bringing us all here together today. My name is Mac Clemens, and I’ve spent my life as a website developer, and working with organizations across the country on how to make their websites and communication more accessible. But today, we’re here to talk about one of the hardest problems facing institutions, local governments, universities, and that’s document accessibility.

Um, despite all the work and progress that’s happened on the web, there’s still big gaps when it comes to PDFs, and with upcoming DOJ enforcement, and the evolving WCAG standards, it is tough to get all those PDFs compliant. And what’s happening out there is very disturbing. In order to be compliant, people are deleting their documents en masse. They’re getting rid of historical documents. They’re, um, having basically a digital book burning in the name of accessibility and compliance, but it’s coming at the cost of transparency.

So we, as a society, face a crisis. Do we somehow use technology to bridge this gap and make accessibility of documents available, and available now, to stop this? Um, or do we give up and leave this as a hard and unsolvable problem? Well, we know the answer to that. We know the right answer to that, but it’s not easy to do.

So, I’ve enjoy- I have invited two experts with me today to talk about this problem. First, I’d like to introduce Carter Temm. Uh, Carter is a lifelong screen reader user. He’s helped develop NVDA. He has helped consult with numerous organizations, from Wells Fargo in banking, to many others, to help them improve their accessibility experience. And he’s gonna talk about some of the tools available and some of the things we’re working on.

I also have invited Shawn Jordison, who is The Accessibility Guy on YouTube, and has remediated maybe a million documents. I don’t even know. Shawn, how many documents have you remediated?

SHAWN JORDISON: You know, it, it’s really hard to, to guess, but throughout the… This has been my only grown-up job since I was 18, and I’m 35 now, so, uh, a lot. (laughs)

MAC CLEMMENS: So basically, between the two of us, we’re completely unemployable. Uh, Shawn can’t work for anybody. Uh, so Carter’s the only hope, uh, for this group. (laughs) But, you know, I think we’ve come together as, like, passionate mavericks to be like, “There’s gotta be a better way to solve this. I cannot watch another local government delete all their PDFs ’cause they can’t deal with it.” Transparency is fundamental to democracy, and if we don’t have trust in institutions, which is under threat today, then we will fail in our mission to empower communication and connect communities.

So, one of the things that we’re gonna be talking about is, like, this problem, and how do we solve it. So, um, Carter, I wanna open it up to you. Like, what is the state of PDFs today, for you, as a screen reader user?

CARTER TEMM: Yeah, for sure. Uh, well, I think a lot of people listening and watching this have experienced a very similar thing, right? You see a PDF document on a website, and there’s this internal feeling of dread. You don’t know whether it’s going to work. You assume it’s not going to work. And there’s a spectrum, from a PDF that’s been, uh, a bunch of pictures kind of sewn together, scanned in, with nothing else added to it. We’ve seen those many, many times. You have the documents that have been scanned in and a little bit of text added, but it’s hundreds of pages of just unstructured information. And then you have the documents, uh, like the ones that Shawn creates, that are excellent, but they unfortunately require a ridiculous amount of time to actually get right, and to tag correctly, and to send over to users to make sure that they’re able to navigate them with all of the assistive technology that they use.

Um, and when you’re dealing with this DOJ legislation, I’ve heard so many people say, “Well, we have hundreds of thousands of documents. We wanna make them perfect. But Shawn’s only got so many hours in the day, and there’s no possible way to do it through conventional means.” Uh, now, most people, when they think of a PDF, they’re going to think of the Adobe Reader, which has made some pretty big strides in the last couple years especially, uh, to try and create accessible versions. Of course, it still relies on being tagged. Um, you have Microsoft Edge and Chrome in your browsers that open PDF documents as well, and maybe you can read the text, uh, assuming it wasn’t scanned, until you run into structural information, like a table or lists, uh, at which point you get all of the information read out to you at once. So, uh, in summary, it’s a pretty, pretty wild world out there.

MAC CLEMMENS: Yeah. And you know what? I’ve heard stories, like when Colorado was passing HB21 1110, which is a law that would require accessibility, not only of websites, but documents, uh, passed a few years ago, you know, there were horror stories. Like, a woman had a, a wildfire evacuation notice read to her one letter at a time. It was a PDF, it had been, you know, put in uppercase or something, and it was reading it like a giant abbreviation. Like, Shawn, you know, is this easy for local governments to, to make these adjustments? Like, what are, what are you encountering when you’re training people on this?

SHAWN JORDISON: Yeah, it’s, it’s tricky. There’s typically two methods that we train to organizations. It’s…… uh, the, the primary method is start with accessibility in mind. Open up a fresh Microsoft Word document and enter your data with styles and, you know, then convert it to PDF and validate. However, if you’re working with, uh, a PDF that has been given to you or that you are editing yourself, maybe from some other legacy platform, or even, like, some online tools that create PDFs, when you go to edit that document, let’s say in your case of having the, the single letter read out- oftentimes you would need to actually update a certain property, or you have to literally change the text in the document.

So what feels like some simple process to, like, swap out, uh, a few letters, what tends to happen is as soon as you hit that edit button in PDF and you try to change some letters, it will actually strip the tags out of that entire container for that object in the PDF. So something as simple as changing a couple of letters can turn into, unfortunately, an hour of work, depending on the context.

MAC CLEMMENS: Yeah, and we’re seeing this at the local government level, where board clerks are going to trainings and they’re like, “I’m barely making my agenda packet deadlines. How am I supposed to do all of this and get those right, and figure out my multi-page tables?” Like, they’re losing their mind over this, and they’re worried about the impact. And then we see on the other end in universities, faculty are preparing lesson plans and they’re trying to make sure that things work, and MathML even alone is pretty sketchy and hard to get into documents.

SHAWN JORDISON: I, I would just add into there, like, one thing that we see a lot with, uh… I, I work a lot with educational institutions along with government, but more education. What we see a ton of is the print to PDF feature. And, you know, raise your hand if you’ve ever been, uh, personally attacked by the print to PDF feature in the chat. (laughs) But, um, honestly what it does is it strips out all of your styles and tags. And so seemingly, even if you did all of the hard work in your source program, it is so easy to accidentally, like, create, uh, a terrible PDF. And that at scale is just, like, such a problem that people are looking to solve.

MAC CLEMMENS: Yeah. And so I think we come to the crossroads of this. The problem is very extreme. We have millions, billions of PDFs. Um, in our legal system, in our judicial system, depositions, contracts, bankruptcy filings, every single board meeting of every local government in America, meeting agendas and minutes, going back decades. This is historic information that’s critical to our democracy and to our learning institutions. And we’re getting to a point where, like, people are just saying “Burn it down. Let’s get compliant.” I actually just had a meeting with a university, I will not name which one, and I was talking to them about this problem, and they said, “You know what Mac, we’re gonna get rid of all of our PDFs for any non-critical documents, and we’re going to move everything to a Canvas portal where we just upload it to an app. It’s all locked down, it’s all proprietary, nothing will be indexed. We’re going to encrypt it. But we won’t have PDFs anymore.”

And I was like, “That is not a solution. You’re breaking transparency, you’re breaking search indexability, you’re breaking the ability for knowledge to spread, for people to link to it, for people to cite it.” Like, this is not okay. We’ve gotta actually solve this problem. So that led me down this journey, and I was in awe of what AI models can do. Like, these generative AI models have gotten so much better. Um, but I also was aware of the limitations. Like, last year, I brought on Carter Temm, and I was like, “Carter, I know you’ve been working on NVDA. We need to write a speech synthesizer that is specifically for PDF accessibility that reads at high speed, but in a format that, like, is useful and makes sense.” And the previous approach in speech synthesis for documents was just to read it out loud as you scan through it, left to right, top to bottom. But when you have these multi-column formats, that doesn’t work. You read across both columns at once, and it makes no sense.

So we were breaking new ground trying to get these things to scan in a way that makes sense. And I realized very quickly we need to transition to HTML. PDFs are inherently non-accessible, and we need to transition them to HTML. But there’s a lot that goes into that. You need meaningful alt text. You need to understand the intention of what something is. You can’t just mindlessly translate what’s there. And I want to take us through a couple of examples of how we’ve been working on solving this problem with DocAccess.

So this is the actual interface. This is a city manager’s report from San Rafael. And what DocAccess does is it allows you to drop a PDF into it and have it generate an accessible HTML version of that document. And the way that we approached it, we need to have a high-quality visual model. We cannot just use OCR. OCR is oftentimes not accurate enough, because we need to be able to understand and read hand-written notes. We need to be able to read low-res images. We need to be able to read things that are skewed or tilted. And I had someone in government literally last week say, “Hey, I have this PDF, and I know who sent it to me. They wrote it on their smartphone using their thumb, like, they hit print to PDF, and it’s a little bit crooked, and it’s low resolution. Can your system do this?” And the answer is yes. Like, we can actually handle that stuff now because of the quality of these vision models.

But there’s more to it than that. We have to understand what’s in the visual. And I’ll give you an example. If there is a form field, we need to be able to preserve that form field’s name and have it be in the proper form format. If there’s a hand-drawn signature, we should honor that and have alt text that says this is a signature. If there’s a table, we need to preserve column headers, row headers, scope. We need to keep it intact. But if there’s a section that’s just two columns of text, don’t turn that into a table. Recognize that that is just two columns of text and let it flow into each other. These are nuanced things that are really difficult to build into a system.

And so what we’ve done is we’ve used what’s called reasoning tokens, where we’re actually forcing the model to not just respond with the answers, but to think through, “What am I dealing with here? What is the intention? What should I do?” And then we layer on top of that. We also have quality control processes. So after it generates the initial HTML, we have another model go and validate it and check it and make sure it looks right, and look for hallucinations, and look for anomalies. And we’re documenting all of that now.

But one of the challenges, and I really wanna highlight this, is that you can’t just mindlessly convert a PDF to HTML. You have to understand the document as a whole. And let me give you a specific example. On every page of this document, it says “City of San Rafael” at the top. But that’s not the heading. That’s actually just a header decoration. The actual heading is what’s unique on that page. And so we have to understand, “Okay, this text is repeating on every page, so it’s probably not a heading. It’s probably decoration. Put it in a header div or ignore it.”

But then you have pages where the information in the header is actually meaningful. Like, if it’s a chapter title that spans multiple pages, that should be preserved as part of the structure. So we have to have logic that understands, “Is this repeating text meaningful or decorative?” And that requires processing the entire document first before we can process any individual page. That’s one of the things that makes this so complex.

Another example: tables that span multiple pages. If you have a table that starts on page 3 and continues on pages 4, 5, and 6, only page 3 has the column headers. But for accessibility, every page needs those column headers. So we have to detect, “Oh, this is a continuation of a previous table,” go back, grab those headers, and apply them to subsequent pages. That’s the kind of intelligence that you need to make documents truly accessible.

And then there’s the question of accuracy. Like, Shawn has spent his career getting these things perfect. And I don’t want to replace that. But I also recognize that perfection at the scale of billions of documents is impossible. So what we’re trying to do is get to, like, 95% accuracy, 98% accuracy, where it’s good enough that someone can use it, that it’s meaningful, that it’s helpful. And if there are minor issues, like, maybe a heading level is off by one, or maybe the alt text isn’t perfect, those are things that can be fixed over time. But the important thing is that the document is usable, that people can access it, that information is not being deleted in the name of compliance.

And I’ll be very transparent with you. Like, this has been terrifying to work on, because I know that if we get something wrong, if we generate inaccurate information, that could have real consequences. Like, imagine if a building permit has a wrong number in it, or a legal document has a wrong date. That’s serious. So we’ve built in safeguards. We have quality control agents. We have human review processes. We’re documenting everything in implementation logs so that people can audit what the system did and why.

But at the end of the day, I believe that technology can rise to meet this challenge. And I believe that we have a moral obligation to try. Because the alternative—deleting millions of documents, locking things behind proprietary portals, breaking transparency—that’s not acceptable. We need to preserve access to information while also making it accessible to everyone, including people who use screen readers.

And so that’s what DocAccess is trying to do. We’re trying to use AI to bridge this gap, to make it possible for local governments, for universities, for any organization with large document collections to make those documents accessible at scale, affordably, quickly. And we’re doing it by converting PDFs to HTML, which is the right format for accessibility.

Now, I know there are concerns about AI. People worry about accuracy, about bias, about hallucinations. And those are valid concerns. But I also think that the perfect can’t be the enemy of the good. Right now, people have no access to these documents at all. Or they’re deleting them. So if we can provide 95% accurate access, that’s infinitely better than 0% access. And we can continue to improve over time.

CARTER TEMM: Yeah, and if I can just add to that, from my perspective as a screen reader user, I would much rather have a document that’s 95% accurate and that I can use today, than to wait another 5 years for it to be perfectly remediated, or worse, to have it deleted entirely. Like, imperfect access is still better than no access. And what Mac and the team have built with DocAccess is genuinely impressive. I’ve been testing it extensively, and while it’s not perfect, it’s better than anything else I’ve seen at this scale.

SHAWN JORDISON: And I’ll add, you know, as someone who’s spent 17 years doing manual remediation, this isn’t about replacing what I do. It’s about making it possible to do what I do at a scale that was previously impossible. Like, I can remediate maybe 50 documents a week if I’m really pushing it. But there are organizations that need to remediate 50,000 documents. There’s no way to do that manually. So tools like DocAccess aren’t replacing manual remediation—they’re making it possible to get documents to a usable state quickly, and then manual remediation can focus on the most critical documents or on fixing edge cases.

MAC CLEMMENS: Exactly. And I think that’s the right way to think about it. This is about scaling accessibility, not replacing expertise. And I’m so grateful to both of you for your partnership on this. Carter, your testing and feedback has been invaluable. Shawn, your expertise and knowledge of what makes a good remediated document has been essential. And I’m grateful to all the local governments and universities that have been part of our beta program and have given us feedback.

We’re not done. There’s still so much work to do. But I genuinely believe that we’re on the right path. And I believe that with technology, with expertise, with commitment, we can solve this problem. We can make documents accessible at scale. We can preserve transparency and access to information. And we can do it in a way that honors the principles of accessibility and the needs of users.

So if you’re interested in learning more about DocAccess, please visit docaccess.com. We’d love to talk with you about your document accessibility challenges and how we might be able to help. And I want to thank the Vista Center again for putting on Sight Tech Global and for giving us this platform to share this work. Thank you all.

[MUSIC PLAYING]