Lately, I’ve been experimenting with text-to-image AI tools like ChatGPT Sora and Google Whisk. I tried to see how far and wild the two of them can go. But now, I’m torn between pushing my creative limits and avoiding anything that triggers filters or misinterpretation.
I’m just wondering… how are people dealing with this?
And also, are there custom tools that don’t care about filtering their input and output? Do you know any?
A lot of models have built-in moderation. However, prompt engineers are finding ways around this moderation in order to produce whatever content they want. Furthermore, models can be downloaded and tweaked/modified to generate content not originally intended (if that model is open-source). But, generally, it’s being dealt with by training the model what’s an okay prompt and what isn’t.
As far as unfiltered models, there are a few. I don’t know any personally, but to my knowledge, there is an increasing (though lagged) response of modified models that have little-to-no moderation.
In a different ethical light, this is a large - and growing - problem with image generation. If it’s trained on images of actual people’s likenesses, we will quickly run into an issue where deepfakes or unmoderated content can be created of real people (which can lead to plenty of negative consequences).
I generally do not like AI image generation myself for any reason. Of course these consequences can be from deepfake of scams as you said below to false evidence being created or mass manipulation:
I would generally avoid any unfiltered models, as most AI models generally try not to unleash chaos and misinformation. Furthermore, there are usage polices by the companies that attempt to prevent this. I would avoid tools that do not filter input against potentially harmful options.
As for this:
The text-to-image that you should be using should be fine enough for baseline images, but for specific details, the model needs to redo the entire image, which leads to misinterpretation. I would suggest using tools such as Photoshop for local areas or isolating parts of the image for refinement.
I have a great conundrum that my one friend shared for me that I’m absolutely going to steal:
Let’s say that you struggle to do math. Your brain just doesn’t get it. So, a calculator makes it easier and solves a lot of your problems for you. Shouldn’t we afford the same luxuries for those that struggle to write well or to make art?
This does seem interesting, but there is a clear difference. A calculator is a tool to assist in maths, and it allows you to solve equations faster. If this were about tools such as Grammarly (grammar correction with AI), which correct issues and provide helpful improvements based on human writing, then that would be acceptable in support of your point. But this is about pure generative AI by my reading of it:
These are not generally tools; these are algorithmic creations. They remove most, if not all, creative input a person could have. In these examples, you can see that from the same basic prompt, you can get different results(albeit using different models). If you also ask the same, you can get a different result than I did.
Even so, there are many ways to do art, even in the most abstract of senses, as seen by Comedian(2019), or better known as Banana Taped to a Wall.
It inspires questions on artist intent, which is something you cannot get from AI, as the AI itself cannot explain its choices as they are algorithmically and probability proposed, and the prompts can give a completely different artwork at times. Even art considered kitsch or anything from Thomas Kinkade still has creative input in the way the art was made, even if the final product is meaningless.
Therefore, it cannot be considered an art piece, as the idea of generation with prompts does not result in any significant creative input. Any art you can do yourself with choices, or tools to improve issues that you have, will be better, as there is something to question behind it, there is meaning from the work’s creation that which AI generation cannot do.
I think you make some very interesting points! Let me try to take a different perspective on it, though: let’s say I’m playing my favorite tabletop RPG, D&D. I made this great character (it’s a frog wearing samurai armor and he has a lightning bow. He travels with his best companion, Pancakes the Firefly). But no images of my character (Spuds) exist on the internet, no images even closely resemble him. If I want to have an image of my beloved character, I either have to a) pay a real human being to make it, and even then, it might still be off, or b) do it myself. But I’m lazy and I’m broke, so should I be barred from leveraging AI to make it for me (or should that AI be constrained in its freedom)?
In all respects, art, due to being a specialised skill, is going to have its upfront costs in time and money. And by your own admission, you cannot afford commissions, nor are you able to do the art yourself.
I’ve had a go at Photoshop in 5 minutes with stock images to create below, however, it’s probably wildly inaccurate for your own character(and I haven’t added pancakes to the firefly).
But for this purpose, although I am against it, it can more so be compared to fast food or ready meals. They may not have any substance to them, but they are quick and easy replicas of better stuff. If you have never cooked, then you probably rely on these, and cooking is a skill that takes time and money to get good at. And sometimes, you don’t have the money or time to make something good, you just need something now.
As for the AI itself, it is the producer of this “fast-food art” and is ignoring the consequences of its own actions. Not all freedoms are considered good, as they can be pushed to their logical extremes, which can lead to disastrous results. However, arguing about which freedoms are good and which aren’t is in itself a massive debate hole in regular ethics, as the lack of freedoms can also lead to disastrous results. But you, as the end user, shouldn’t be banned from an option under our current way of life because sometimes the bad option is the only option.
Thank you for bringing my beloved Spuds to life 
I think that’s a very valid take, and I appreciate your insight! I think that the problem may worsen over time as AI models become more sophisticated and especially as these models are generated more and more by the end user, rather than large data centers. But, time will tell. Again, thank you for your blessing of photoshop skills! 
You’re welcome, and you can use them in any way you like. He’s yours.
As for this:
I think the only method is rigorous enforcement of regulations to declare that this was AI-generated, possibly by adding a metadata tag to all images saying it is AI-generated. As the technology improves, it will become impossible to tell, and for someone who prefers human art over AI, I would like to know if it was created by AI or not.
I think that’s fair. Do you foresee rogue models or people deploying their own as an upcoming issue, where the models are built/trained with no moderation requiring the image be watermarked or the metadata be written? I can only imagine that’s right around the corner, amplified by the fact that the models will also be getting stronger.
Yes, I can see it, but thankfully, at this time, AI-generated content on these rogue models will have more generation errors, and thus be able to be detected as AI-generated without metadata watermarks. This becomes a problem when current cutting-edge or hypothetical human replication models are able to run mainstream without servers. However, bad actors can be kept rare for a while and controlled. It isn’t going to be perfect, but major corporations that will likely have the main market share will not be able to do this and not get caught. You will still get smaller models that will produce this, regardless, but the majority of AI content will be watermarked “made by AI”. It’s a significant improvement over not bringing in any regulation for AI-generated content.
I agree that it’s a major improvement, but I imagine that we’ll have the following situation: first, we’re in the “that’s obviously AI” phase, like we are now. It’s easy to detect (mostly) and doesn’t yet need a lot of watermarking. Next, we’ll get into the “indistinguishable from real video” era, where we’d need watermarking - but at that point, or soon thereafter, these models would be available to individuals. Do you think there’s a feasible way to enforce moderation at any of these timeframes?
You make great points. “AI or real?” can be a big deal if it’s you, someone, or something you care about. A cartoon seems pretty far removed from the new norm. I regularly see AI generated world leaders doing strange things in short video posts. On the other hand, IA avatars that are just like the real people they represent are all over YouTube filling in for personalities who still collect revenue, even though they aren’t in the videos anymore.
I suspect the issue is going to be forgotten and in the new AI era (now), a whole new slew of issues will emerge.
