Artificial Intelligence, or at least talk about AI, has become inescapable. Like “gluten free” it has been become a marketing tool that doesn’t mean anything. Gluten free toilet paper? – I’m sure somebody is selling it. AI powered paper clips? If not today then certainly tomorrow.
Nonetheless, AI is real in the sense that computers can be programmed to do things that look a lot like intelligence. The term “AI” doesn’t have any well defined meaning, so the initials are far less important than the capabilities of any particular so-called AI system. Whether you think “generative AI” is really intelligent or not, there is no doubt ChatGPT can turn out a poorly written essay much more quickly than a typical college student, and if your goal is to get a solid “C” while still having time to party then ChatGPT and its competitors have all the capability you need. If, on the other hand, your goal is polished prose that can persuade and educate you will probably find that generative AI is pretty dumb.
That brings us to AI and accsss to the internet. “Accessible” for websites, applications, electronic documents and the like is defined for most purposes by the Web Content Accessibility Guidelines (WCAG), which consists of a short set of principles and a long list of rules that a website or application should satisfy to be accessible. Beyond that are thousands of technical descriptions of how to accomplish what the rules require. Even before AI became the phrase of the hour computers were really good at following rules because applying rules and finding rule violations lends itself to the algorithmic processes at the core of how all computers operate. Programs that scan a website for failures to follow WCAG have been around for many years, and the step from finding a rule violation to fixing the rule violation is tediously complex but straightforward. Doing tasks that are tediously complex but straightforward is where computers shine.¹
Unfortunately, there is a significant gap between accessibility rules and equal use and enjoyment. Whether access is equal depends on two things that software cannot readily discern from simply scanning the code behind a website looking for WCAG rule violations. These are the purpose of the owner in creating the website and the goal of the disabled person who is trying to use the website. To take my favorite example, many websites make money by allowing (for a price) the display of advertisements. Those advertisements often take the form of click bait – pictures with captions that the advertiser hopes will induce the viewer to click and go to another site displaying advertisements. The purpose of the picture is to make you click – hence the name click bait.
WCAG has a rule about pictures:”All non-text content that is presented to the user has a text alternative that serves the equivalent purpose.” It is easy for software to detect a picture – non-text content – by scanning the code behind a website. It is harder but certainly doable to scan the image itself and identify in general terms what it is – a dog, or a person or a car. Eventually the software will become sufficiently sophisticated to go further – the dog is a german shepherd or the car is a 1994 Toyota 4Runner. But describing the picture has nothing to do with its purpose, which is to induce a click. For software to generate text that “serves an equivalent purpose” it would need to know the purpose of the website owner/advertiser (click bait) and then generate alternative text likely to induce a click by a blind person using a screen reader. I have an antique car and often visit websites with information about antique cars. The click-bait images these display are either car parts or women wearing revealing clothes and a caption like “ten wardrobe fails you have to see.” I know what the advertisers think of the antique car demographic, and perhaps these generate clicks. It is hard, however, to imagine a person who is blind being tempted to click a picture with alternative text hinting at more pictures of scantily clad women. And since click bait is mostly an annoyance it isn’t clear a blind user would even want to hear a text description.
There is another kind of purpose that AI is unlikely to discern. I recently visited a website whose background image was a collection of young, smiling and very diverse group of people sitting around a table laughing and talking. To me the message was clear: “buy our product and you will be like these attractive young people having fun” combined with “our products are for everyone, regardless of ethnicity or national origin or disability status.” It isn’t hard to imagine AI generating a description of the picture as a group of people, but less easy to see how AI could generate alternative text conveying the message I got in an instant. Equally important, even if a very sophisticated program generated a sufficiently detailed description to accomplish the purpose or that image it would probably be an annoyance to a blind user who has to listen to all those words. A picture is worth a thousand words, but while sighted users can glance and the turn away a screen reader user has to read all the words. Even at the high text reading speed experienced screen reader users employ that would be tedious, especially for a picture of secondary importance. In short, making the picture “accessible” as defined by WCAG would make the website less accessible for screen reader users.
Of course images create special communication problems for those who are blind just as sounds create special communication problems for those who are deaf.² There is, however, a deeper problem with accessibility created through the application of rules. Following a complete set of rules may make a website “accessible” in some theoretical fashion, but practical accessibility requires consideration of the limitations and abilities of whatever assistive technology a disabled user relies on. Title III of the ADA forbids discrimination in “in the full and equal enjoyment of the goods, services, facilities, privileges, advantages, or accommodations of any place of public accommodation.” For physical accessibility this means a disabled user can get in, get around, and buy, sell or otherwise do what they came to do. That in turn means, for example, putting in ramps instead of steps for wheelchair users. It also means, because wheelchair users and other with mobility disabilities do not move as easily, that at least some of the accessible parking for those ramps be close to the front door. Equality of enjoyment requires some special treatment, like an up close parking space. Because I am not disabled I could not say what different treatment might be required to create equality of use and enjoyment of a website, but I’m reasonably sure that merely making sure every part of a website is accessible does not guarantee equality of use and enjoyment and that even if AI is capable of both identifying and correcting WCAG errors it won’t be able to go beyond that and think about what it feels like to use assistive technology and what changes in design might be needed to go from mere accessibility to genuine equality of use and enjoyment. It is also difficult to imagine how AI would implement the proper compromise between that which improves equality of access for the disabled and that which makes a website less usable by a person without a disability.
This is only a sliver of the discussion that needs to be taking place about accessibility and the internet. There is, for example, the fact that creating accessible websites is so expensive that small businesses and hobbyists cannot afford to so but are still held to the same standards as the largest multi-nationals. There is the complicated question of whether website inaccessibility should be addressed only by looking at websites, or whether, perhaps, the creators of assistive technology have some responsibility for improving the capabilities of their products so as to overcome WCAG errors. (4) At the end of the though, as long as websites are owned by humans and used by humans it seems unlikely AI will be more than a part of solution for accessibility and, that it may well be a distraction from accomplishing the harder work of creating equality of use and enjoyment.
++++++++++++++++++++++++++++++++++++++++++++++++
¹ I’m not suggesting that those who create computer software are not creative – building complex systems that work efficiently is not straightforward. But the end result will always be a series of commands that execute a well defined process to turn some input into a different output.
² I cannot image that the open captions describing music that you can activate on a television come close to eliciting the emotions of the music itself.
³ Jason Taylor at Usablenet deserves any credit for this insight, but not any blame if I have misstated the idea.
(4) For example, if AI tools can scan a website to identify missing alternative text and supply the missing text there would not appear to be any reason a screen reader program could not incorporate the same technology and make the user less dependent on the website owners and developers.