AI, a brilliant nerdy sibling?

When updating my website, I also wanted to refresh the images with the various blogs. The existing illustrations were of varying quality, format and style. More importantly, the photos were found online, and copyright issues may have been involved. I thought I would address most of these issues by generating graphics with DALL-E (part of the ChatGPT 4). And I could, mostly.

It looks great. Or doesn’t it?

Some topics are easy for my new AI friend, DALL-E. When I asked for a picture illustrating that ‘Time = Money’ or ‘Comparing Apples and Oranges’, I had no complaints about the result. The ‘Baby on Board Sign’ was also easy; I could even decide which side of the glass I wanted the sign to be. ‘If you have a Hammer, every problem looks like a nail’ was apparently a familiar concept, although the perspective would not convince any art teacher. I was very impressed by capturing ‘Curiosity’ as an emotion on a little boy’s face, even when the little guy was not looking at anything interesting. Changing gender or age was easy as well. Again, impressive!

But then it gets weird

DALL-E seems to struggle with spelling. Whenever I asked her to put a word or a sentence on a toolbox, a T-shirt, or a sign, she got it wrong nine out of ten times. When I asked for the same picture again but with correct spelling, she mostly did better but hardly ever perfect. For that reason, you will see hardly any images or text in my blogs. I didn’t find the time to train her in spelling.

Sometimes, the results were plainly weird. The kingfishers look OK, although their anatomy is slightly puzzling in the details. And the fish jumping out of the water shows way too much similarity with its predators. Even more alarming are the hands. ‘Sourcing for a Better World’ is almost perfect, but eleven fingers? ‘Compliance and Integrity’ is all about the hands, and having six fingers to a hand is hardly acceptable. For illustrating ‘Elevator etiquette’, the generated image is elevator-like enough, but you shouldn’t give it a second glance because it is also pretty weird.

Great, but not perfect

I started this endeavour by uploading a blog and asking for an illustration. I will spare you the results, but it is safe to say that OpenAI still has some work to do in that area. At the same time, if my life depended on it, I would not been able to generate pictures with a quality even remotely close to what I ended up with. I am impressed by almost all the results. Even the ones that didn’t make the final cut and were replaced by improved versions were a good starting point. A couple of iterations with more focused instructions typically would help a lot.

At the same time, it is like having your brilliant but nerdy sibling do your homework for you. Please read it carefully before you hand it in as your own. You might get into trouble if you don’t check, re-check, and double-check the outcomes. At the bottom of the screen, OpenAI warns that ‘ChatGPT can make mistakes. Consider checking important information.’ Even while creating impressive results, that warning is still very relevant.