My colleague Yuri Westplat showed me an AI that could generate a working website by following a few commands that you type into a textarea. He also showed me a tool that could generate a pretty complicated interactive web thing, again by typing a few descriptive sentences into a textarea. And the results weren’t even much worse than the sites we use today. So should we worry?
Machine learning needs data. Without data these machines can’t do anything. The tools that Yuri showed me used the source code of millions of websites, and millions of repositories in Github. And with this data these tools create all kinds of new things, based on what you ask it to do. This means that it is very good at creating average websites. It also means that the results are not of very high standards.
What is the quality of this training data?
WebAIM has been doing some very interesting research that looks at the quality of the million most popular websites from an accessibility point of view. Over 95% of the homepages of these websites had detectable accessibility issues — remember that most accessibility issues can’t be detected by a robot. These are statistics for the million most visited websites. I’ll assume that these websites have a budget to hire good developers. So if even these developers can’t create code of acceptable standards, what would the quality of the rest of the web be. In other words, what is the quality of the data that these bots use to train themselves?
The data is bad. The results will be bad. If, from now on, we switch from humans to AI to create our websites, our websites will forever be as bad as they are today. (On the other hand, if we don’t switch, our websites will be bad as well.)
Should we worry?
I think quite some people should worry, yes. People who build mediocre websites will definitely loose their jobs. An AI will be much more efficient in making mediocre websites. And if you look at the stats that says that 95% of the top websites have detectable accessibility issues, you could assume that 95% of the developers should fear for their jobs.
Organisations that want a higher standard will still need humans, I think. So not every developer will lose their job. The quality that is expected from these people will be much higher than the quality of things we make today, and much higher than the AIs that are trained on the things we make today.
Is web design AI a good thing?
Some people may argue that this will further democratise web design. More people will be able to create an average website, even people without the budget to hire designers and developers. From this point of view, this is a good thing. This will enable hobbyists and small non profits to create their own independent sites. This is very nice indeed.
Others will argue that by letting an AI do the tedious jobs, developers will have more time for important things.
I have two issues with this.
First of all: in practice we know that automation always results in loss of jobs. The result is never an improvement of working conditions. It’s a middle management argument, purely based on efficiency, and not on humanity. So no. Most developers will not have time for important things, they will be fired.
The other issue I have is: who decides what the tedious jobs are? If I look at how popular web development is among my students, and if I look at how much fun they have with typing code, I can’t help but wonder why should we ever want to automate something that is so much fun to do?
Who makes these AIs?
Finally, a very important question to ask is: who makes these AIs and are they to be trusted. I will not answer this question in this post. Instead you should take a look at this video about AI art that answers it clearly.