ChatGPT - the Potential of Artificial Intelligence for digital Accessibility
"Cool, what AI can do" and "Oh, it really can't do anything" are two statements you hear lately, especially in connection with ChatGPT. The latter statement is heard mainly in Germany. If there is a bug somewhere, someone from Germany will definitely find it. It's no wonder that many companies and researchers are leaving the country.
I am excited not only by what technology can already do, but above all by the potential it holds. And it saddens me that, despite the enormous intellectual and entrepreneurial potential in Germany and the EU, we are not achieving anything comparable. Some of the biggest developments of the last years like the MP3 format and a lot of basic research came from Germany/EU and Others are making it usable. But let's not go there. Today I want to look at the potential of AI for digital accessibility.
It is legitimate to point out gaps and problems, it is equally legitimate to point out potential. Young persons don't need to care about digital accessibility because they often don't need it and there are many other topics where they can contribute their ideas. If they keep hearing that AI doesn't do anything in digital accessibility, they will look for another topic.
- Opportunities for AI in accessibility and criticism.
- Technical conservatism in the scene
- Final thoughts.
- Further Reading
Opportunities for AI in accessibility and criticism.
I'm not necessarily talking here about the things AI can already do today, especially not flawlessly, but about the potential I think it could have in the foreseeable future.
What ChatGPT and other generative AIs are capable of is composing existing knowledge. One can find everything that a Google search would also bring up, but usually more compactly than Wikipedia. So you can get answers to questions - an enormous relief for persons who don't have sophisticated research skills or access to databases. In other words, the vast majority of humanity. Wikipedia may be all sorts of things, but for many persons the language is too complicated.
It is true that ChatGPT also produces wrong answers. For example, I asked for studies on accessibility in finance. ChatGPT gave me five studies, 4 of which could not be found via Google, so they probably don't exist. But other persons and Google also unearth wrong answers on other topics. The question is whether you want to measure the quality of an application entirely by what questions you specifically ask it to lead it up the garden path.
At this point, I'd like to pick up on a few of the criticisms I've heard.
It is said that AI cannot adequately describe images because it does not know the intention of the person providing them. That's correct, but 1. another uninvolved person wouldn't know that either and 2. the interesting thing about ChatGPT is that you can ask questions, I can do the same with a person and they probably won't answer. The AI is patient and varies its answers, many people are not.
The next argument is the bias of the machine. Of course, you can also be discriminated through software if it only reproduces discrimination. But the following also applies here: 1. Does the software only reproduce human prejudices and 2. can a machine be truly neutral, while a human is always a prisoner of their prejudices. Those who believe in anti-bias training have never considered the effects of training and entrenched assumptions. We are tolerant as long as we can avoid people that are different from us.
The argument about the poor quality of texts and image descriptions is also correct. I would categorize what is publicly available so far as nice to useless. There are now some applications that can describe scenes or recognize individual objects in images. But I also see great potential here. The problem today is that images are described once and blind people have to cope with that. But some want a brief description, others a detailed one. With complex information graphics, a description rarely includes exactly what you want. But with an AI I can ask questions. But can I do that also to a person too? True, but it doesn't make an AI bored or impatient (unless it's Marvin from The Hitchhiker's Guide to the Galaxy). As a blind person, you don't always want to be dependent on the good will or whim of a sighted person, and perhaps there are things that you don't necessarily want to show even to a person you trust. An appropriately trained AI should be able to adequately describe complex information graphics from studies and work. Infographics consist of structures, patterns and labels that should be easy to capture with machine learning, especially if they are vector graphics, i.e. if the AI can access the underlying code.
The argument of limited intelligence is not directly related to disability: software primarily aggregates accessible information. It couldn't produce a Mozart or Shakespeare. The argument is also correct. However, 99.9 percent of humans are unable to do this either. The creativity of the web designers lies primarily in the question of whether the font should be 12 or 13 pixels and which colors should be put where. No offense, but if you take away their fonts and colors, all websites look relatively the same, especially on smartphones.
Other creative works like pop songs or books, consist of recurring and modified patterns. We're not that far away from an AI creating a Grisham or an Eminem song. There's maybe ten percent creativity in most works, the rest is routine work like fleshing out scenes, developing characters or settings. In other words, I could formulate a few ideas and tell the AI to write a novel in the style of Agatha Christie, then add a little fine-tuning and the bestseller is ready. Utopian? Perhaps today, but probably possible in the foreseeable future.
There are already programming assistants such as Copilot, which, according to reports, make programming easier and improve the code. That's all we want at the moment; we need tools that do the tedious or superfluous work for us. The criticism is somewhat reminiscent of the early days of Office, when some thought that typographers could design documents better and that Excel would constantly miscalculate.
Automated testing for digital accessibility is possible. Contrary to popular belief, accessibility is not excluded from test automation. The fact that common tools do not work very well so far is due to the fact that hardly any of the big players are active in this area. Matching image descriptions with an image, recognizing links as meaningful, analyzing form labels - these are all things that can be automated.
The general argument is the very limited capabilities of AI so far. Now this is pretty much nonsense. It's like criticizing a three-year-old because he can't yet speak or write fluently. Yes, the capabilities are currently limited. But we are still at the beginning of development. Machine learning is characterized by the fact that it can get better and better.
Humans can better translate or simplify texts or describe images. Some humans, not all. There are good texts or image descriptions translated by AI and there are really bad translations or image descriptions created by humans. It's just not true that a human automatically makes everything better. I have seen so many bad translations from English that I would prefer DeepL any day. This is even more evident with plain language: There are plenty of translators out there with moderate to poor quality. The WCAG will soon be 25 years old and we are still talking about image descriptions and semantic HTML like on day one, human intelligence seems to have failed, maybe we should give AI a chance.
I find comparing an AI to an expert less helpful. The AI should rather be compared to an average human and then it often performs better. An expert will today often perform better. In my opinion, however, it is only a matter of time before specially trained applications are available that can keep up with experts. This makes perfect sense: specialization has increased more and more, even within professions. Today, no person is able to keep track of everything, even within a relatively small area. An expert system could be very helpful for a scientist or doctor, for example. Or - my A11Y-bodies will stone me for competently answering questions about digital accessibility. In case anyone hasn't noticed - there's a shortage of skilled workers that won't be solved in the foreseeable future.
Again, I see benefits to digital accessibility. I especially recommend that newcomers to the topic read at least two articles from different sources on a topic, because persons in the scene tend to be sometimes difficult to assess and opinionated. The AI could aggregate the essence of several articles and thus take away the trouble of reading many different opinions.
Technical conservatism in the scene
Of course, criticism is always legitimate and also helps us. I just don't like the somewhat arrogant way in which criticism is often expressed in Germany. It often obscures the view of what is possible. Telephoning in the early days of the telephone - very poor quality, Internet in modem times too slow, with horses you didn't have to crank, etc. pp. If one would be just as critical of green technologies like insulation materials, solar cells or wind turbines, nothing would progress at all.
It is true that under AI in general and especially in AI and accessibility a lot of junk is sold. These are, for example, the so-called overlays or toolbars, i.e. tools that promise automatic accessibility, but in reality worsen accessibility. Automatically generated image descriptions were so far disappointing. And even with reputable organizations, you have to use common sense. The US-Americans like to turn the marketing up three notches: They are talking of "revolution", even if it is only about small improvements. On the other hand, it seems to me that there are also persons in the scene who believe that everything should be tested manually as it was ten years ago, as if that were still state of the art. Test automation is one of the big topics in software testing and we have missed this train. If we work on digital accessibility at the same pace as we have in the last 20 years, we won't get much further than we are today. The Germans are particularly critical here, but the international scene also seems to want to remain closed to development on the whole. In my opinion, the line between technology criticism and technology conservatism has been crossed by many persons. The goal should not be to close oneself off to technical development, but to point out deficits and to improve them further.
It also seems to me that demographic change is having an impact here. Many pioneers are now reaching an age where - it seems to me - they have a rather negative attitude toward technical development. If you've done it that way for 20 years, there's no reason to change it.
As with many developments, there are good and bad sides to AI. The WWW has made minorities as well as right wingers and conspiracy idiots louder. AI has a lot of potential for surveillance and manipulation. But it can also facilitate access for disadvantaged persons, for example by simplifying the coding of simple solutions or by making texts at different levels more understandable by translating texts or correcting orthographic errors. These are problems that many of us don't know, but are helpful or even existential for many other persons.
I dream of a tool that helps persons with technical problems to use digital applications. It could analyze the structure of an application.
In general, I still believe in the potential of speech-based assistants. As mentioned above, ChatGPT can not only read out information from Wikipedia - not very helpful for persons with little text comprehension experience - but also summarize information in an understandable way and answer queries. This would be an interesting tool for functional illiterates or for tutoring.
Another topic is speech to text. Speech to text is a simple way to control computers. However, it is not possible for many persons to use this method due to pronunciation problems. Machine learning could gradually allow software to understand voices that are less articulate.
According to evolutionary theory, AI is the new great mortification of head persons in particular. There we sit for hours on a text or a code and a software - also developed by humans - can generate something comparable in a few seconds and maybe even better. Everyone thinks that AI could make many jobs redundant, just not their own work. But I'd argue that any job done on a computer can be done, at least to some extent, by AI. AI can't, however, watch our children, care for our parents, build our homes, or take out our trash. These are persons for whom we have little respect - and little compassion when their jobs have been abolished, as with miners, or taken over by machines in factories. We shrug off the end of many small farms before turning to the next nonsense from Elon Musk. But our work must be preserved at all costs; we are creative.
By nature, most of us are information professionals - that comes with the job. But I miss the perspective of those who are not in the snooty posts of some contemporaries. There is a lack of understanding of the challenges of those who can't handle the information overload.
One should take the term Artificial Intelligence as what it says: machine intelligence unequal to human intelligence. The English term Intelligence is not identical with the German term Intelligenz. It means, among other things, data processing. AI will not completely replace human intelligence or even manpower in the foreseeable future, and that is not the point. Neither do the terms learning, understanding and knowing fit with computers, if the same is to be expressed with it, which takes place with beings with a biological brain. However, in my opinion, for lack of better terms, one can use these concepts and should not hang one's criticism on them. Just in passing, it should be mentioned that there is no generally accepted definition for the term intelligence.
The argumentation against AI in accessibility starts from some - from my point of view wrong assumptions:
- Experts can do better, by and large - I addressed that argument above. Experts tend to be little divas and can talk quite a bit of nonsense. Ask three experts and you get six opinions (and all of them can be correct). That said, I've heard enough nonsense from accessibility experts to fill another book with it, unfortunately I don't do comics.
- Further, it's nonsense to bring up the AI vs. human dichotomy. I don't know any human who enjoys creating video transcripts. There is not enough manpower to tag all documents, describe all videos, or translate all content into plain language. The alternative is not AI or human, but in many cases AI or nothing. We don't just have a shortage of skilled workers, we have a shortage of manpower, of labor time, and of financial resources.
- One argument is that we can't do it without consulting/explanation. The argument is correct, but counseling does not have to be done by a human. I even believe that the inhibitions to chat with a bot are lower than with a human. Some - not all - persons are quite difficult to deal with in person or explain things in such a complicated way that you don't understand them. I'm sure you're familiar with the situation where you'd like to ask, but for some reason you don't want to with the person you're chatting with. You can grill a chatbot and it won't answer you with arrogance. The premise is false that only a human can answer correctly and adequately, if the software has been trained appropriately, it can do just as well or better. I, at least, would have no problem chatting with a machine, it doesn't care if you have the camera active or if you are sitting in front of it in a baggy sweater.
Don't get me wrong: AI will certainly not solve all the problems of humanity or accessibility, just as the WWW, genetic engineering, nanotechnology or robotics have not or will not. All of these technologies were subject to a certain amount of hype. But these technologies have actually brought benefits and the main thing is to bring these benefits to the persons and not to billionaire corporations so they can earn a few more euros. If we in DE don't mine the potential of this technology, then others will do it in our place and then they will determine what happens to it, as is the case with most web platforms today. Technology and automation will not solve all accessibility problems, but it can foreseeably solve a large portion of them.
The answer, as is so often the case, is open source; no one can have an interest in a few large corporations controlling this technology and using it as they see fit. A lot is currently happening here, and we can hope that the open source community will make a big impact.
- Artificial Intelligence and digital Accessibility
- Cognitive Bias and Accessibility
- Can Software reduce discrimination