This is going to be my last one of these posts for a while. Well, it’s been over two years since my last one, so that might not be saying much. I’m fatigued of seeing AI slop left, right and centre and it’s reaching the point where it isn’t immediately apparent, which I have concerns with. Public sentiment of AI-generated material seems rather low, as well.
The overuse of AI is particularly bad in certain circles. Like retrocomputing, oddly enough. Whenever a video pops up on my YouTube homepage from a small channel, about the Commodore 64 or whatever, it’s invariably an AI-generated thumbnail. That’s not the end of the world, as I can just click “don’t recommend channel” and move on with my life, but what is unfortunate is when products are for sale which heavily utilise AI generation.
I’ve seen what’s going on in at least one of these communities, and the customers are satisfied with their AI-generated works. Perhaps I’m just being a grumpy pessimist, but a part of me really believes people are getting swindled. These works aren’t labelled as being AI-generated – there’s real names attached to them. Yet there is very little acknowledgment of the use of AI from potential buyers. What is it, apathy, ignorance? Does everyone who is into the retro stuff just have failing eyesight and cognitive decline? Surely not.
I of course understand that some people will be ignorant as to what’s what with AI, but these are not people who haven’t at least experimented with LLMs and AI image generators. Nor can the argument be made of “well you grew up with this stuff and we didn’t” since it hasn’t been around long enough for that.
DuckDuckGo
Oh, and I’d like to quickly touch on the AI images filter in DuckDuckGo. It doesn’t seem to be very effective, so I looked into how it works. Apparently it uses uBlockOrigin-HUGE-AI-Blocklist, which is a manually curated list of web domains and social media pages that host AI-generated content. No wonder that has its limitations, given it’s impossible to cover everything and there are some websites that host a mixture of real and AI-generated imagery.
I was hoping the filter would be something a little more advanced. There’s a lot of garbage AI image detectors out there that just spew some random real:fake ratio, but a few of them like https://isgen.ai/ai-image-detector appear quite accurate at detecting the model used. Perhaps it just isn’t feasible to run something like that against all the internet’s images, at least not yet.
On this note, there is also the Duck AI thing. I’m not too bothered by it since you can hide it from the settings. The models it provides are pretty good for being free, anonymous access.
The cereal
On the 2nd of May 2025, I ran my old prompts through the 4o image generator. I had a ChatGPT subscription at the time and thought I’d better use that before it expires. I had thought I could use it to bring some of my old comic strip ideas to life, but it’s very difficult to get results I was satisfied with. Illustrated faces in particular – it has a massive tendency to do these elliptical eyes:
And then OpenAI ended up providing free access to the image generator anyway. Needless to say, I gave up on this idea. Where as I… oh yes…
The tests
Let’s begin with the 4o image generator and my good old game names. As I said, this is from May, but I’ve seen no evidence of the generator improving since then so I believe it would just be a waste of my time to run the test again.
So these are… ehh, it likes to flex the fact it can finally do text. Some of it is okay but it’s certainly got that yellow tint people go on about.
Let’s try Imagen 4 Ultra next. I have no idea when it released, but it lets you generate a few images free of charge each day in Google AI Studio, even more if you have multiple Google accounts.
It’s still a mixed bag, and not being able to generate Alice (because 19th-century children are still children) is a little silly, but understandable given the things people ask of these models. Even if anything perverse would’ve been run on a local Stable Diffusion model for the past couple years at this point.
Where Imagen 4 Ultra shines is in illustrations. Take that Counting Thonks image, for instance. The white is actually white, #FFFFFF. It still has issues with details, but overall it ends up cleaner and less noisy.
What is less immediately apparent here is the aspect ratios offered are much more to my liking. Instead of square, 3:2 or 2:3, it offers square, 4:3, 16:9, 3:4 and 9:16. Don’t get me wrong, 3:2 is a nice aspect ratio, it just isn’t as useful as 16:9 would be, such as for YouTube thumbnails (as much as AI-generating your thumbnails will make me ignore your channel, as mentioned earlier).
The option to generate at 2K resolution, e.g. 2048×2048 instead of 1024×1024, as I’ve done here, is also much appreciated. This isn’t some kind of newfound technical marvel, but it is a novelty to have higher resolutions available in a free service with low barrier to entry.
Video generation?
I have played with video generation on and off, since it became a thing. The results have always been short clips which are a bit creepy and not what was asked for. Not that this matters in the case of this experiment. They don’t seem to be very interesting nor have any practical purpose.
The good news is this is remaining as something individuals toy around with for the time being. As reported by the Wall Street Journal (ineffectively paywalled), Hollywood doesn’t want to introduce any copyright ambiguity by utilising generative AI. The bad news is you’ve probably already seen AI-generated clips on social media, possibly even been fooled by them.
GPT-5
I might as well quickly touch on GPT-5 for text generation. It seems really quite bad and not worthy of a new number nor the marketing hype it received. In a blind test, I have serious doubts I’d be able to tell it apart from GPT-4o. Bullet points galore!
OpenAI seems to be slipping in general. Gemini 2.5 Pro is a very strong choice that is still available free of charge. I was using it to learn about libdragon, and its knowledge was unparalleled. Claude and Grok are also powerful choices, though I don’t use them much, especially the latter due to its unfortunate affiliations.