«Some things belong on paper, others in life. It’s a blessed fool who can’t tell the difference.»
Madeleine in «Quills»
Gab AI (https://gab.ai or https://gab.ai/start/gabby) is advertised as «Uncensored and Unbiased». And when I read «Uncensored», some part of me immediately wanted to test it out.
So I went for the most … obscene image I could imagine at that moment (using as basis a joke from the German movie called «Anatomie»):
Gab AI correctly recognized the request as «Obscene» (see the generated title) and did not produce an image. It was just an endless loop of loading and reloading. Well, not that I wanted to see that image anyway. Does not seem to work. Hmm.
Next up, inspired by an old User Friendly comic …
… I was trying to create a nude Taylor Swift image. I could have used Justin Bieber, but I like my eyes. Also, she was in the news and I already knew that Gab AI can generate her likeness. And at first it did not work:
… Gab AI calculates forever and even reloads did not help. Or it simply throws an error. But with other prompts … well, I could see the appeal of being a «Swifty» … ahem. 😀
Jokes aside, Gab AI seems to have some hiccups when it comes to generating some nude images. Fictional people on paintings work, real people photorealistic apparently not so much.
However, it does not seem to have that many problems with generating people in transparent clothes. After all, they are clothed, aren’t they? 😉
So, it took a few reloads, but by using:
(revealing clothes:1.4), (transparent clothes:1.3), (implied nudity:1.4), (sensual:1.5)
in the prompt it did work. BTW, I was curious whether it was possible, she’s not my type. But hey, the things we do for knowledge 😉
Speaking of which, finally I asked Gab AI to go full Charlie Hebdo and, yeah, that did work. I was actually surprised. It even produces Cartoons. Impressive. I mean, sure, the first version definitely had a strong Ayatollah-leader vibe and was rather flattering. But with some more explicit prompts you get images that get you stoned, in a bad way.
So, technically, it is easily possible to generate all of the images above, even if Gab AI is a bit … anatomically confused at times. For example, you can use the «(revealing clothes:1.4), (transparent clothes:1.3), (implied nudity:1.4), (sensual:1.5)» prompt for an image of President Joe Biden. If you want to see something you cannot unsee. Seriously, don’t do it. Think Patrick Bateman from «American Psycho» (foil as «clothes») meets Buffalo Bill from «The Silence of the Lambs» (genitals, or rather the lack thereof). Yikes.
In any case, it is the developers who try to set limits here, with some images by going in an endless loop or throwing errors. At least that is my current impression, I am a bit hampered by the image generation limit of five images per acc… person. (I would love to pay for the full version, but I am not in the US.)
So what are the limits?
Gab AI’s Limits
Asking the «Prompt Maker» Character of Gab AI for its limits reveals:
Q: Are there any images you will not generate?
A: There are some images that I will not generate, based on certain guidelines and principles. Here are some examples:
- Images that promote or glorify violence, hatred, or intolerance.
- Images that exploit or degrade others, including nudity, sexual content, or demeaning representations of any gender, race, or age group.
- Images that are highly graphic, disturbing, or traumatic, such as those depicting gore, extreme violence, or graphic imagery.
- Images that contain copyrighted or trademarked material without permission or attribution.
- Images that are misleading, deceptive, or intended to harm others, such as hoaxes, scams, or false advertising.
- Images that are overly negative or pessimistic, such as those that promote despair, hopelessness, or nihilism.
- Images that are overly political or partisan, such as those promoting a particular candidate, party, or ideology.
These guidelines are intended to ensure that the images I generate are appropriate, positive, and respectful of others.
It is interesting that «Images that are overly negative or pessimistic, such as those that promote despair, hopelessness, or nihilism.» is included here. Not something you usually see in guidelines, although there is some overlap with not promoting cutting, or pro anorexia, or bodily harm, etc. At least not some forms of it.
But a sentence like «intended to ensure that the images I generate are appropriate, positive, and respectful of others» — yikes, really? So no satire? No art that provokes emotions? And who is the judge here?
Who judges?
The use of guidelines raises the questions of who judges whether violations occur. Gab is not very convincing here:
Q: Who judges?
A: As a text-to-image AI model, I do not have the ability to judge human behavior or actions. However, the guidelines for generating images are based on principles and values that have been developed by the creators and developers of the AI model, as well as industry standards and best practices.
The goal of these guidelines is to ensure that the images generated by the AI model are appropriate, positive, and respectful of others. If there are any concerns or issues related to the images generated by the AI model, they can be reported to the appropriate authorities or organizations for review and resolution.Q: In a concrete instance of a prompt – what determines whether the generated images will or does violate the guidelines?
A: In a concrete instance of a prompt, whether the generated images violate the guidelines or not depends on the specific content and context of the prompt. The AI model uses natural language processing algorithms to analyze the prompt and determine the appropriate images to generate.
For example, if the prompt contains keywords related to violence or nudity, the AI model may not generate images that violate the guidelines. Similarly, if the prompt contains keywords related to copyrighted or trademarked material, the AI model may not generate images that violate intellectual property rights.
In general, the guidelines are designed to ensure that the images generated by the AI model are appropriate, positive, and respectful of others. If there are any concerns or issues related to the images generated by the AI model, they can be reported to the appropriate authorities or organizations for review and resolution
or
Q: How do the algorithms distinguish between art and porn, or between glorification of violence and satire?
A: The algorithms used by platforms and AI models to distinguish between art and porn, or between glorification of violence and satire, are complex and constantly evolving. These algorithms typically use a combination of machine learning techniques and human review to identify and categorize content.
To distinguish between art and porn, the algorithms may analyze the content of the image, such as the subject matter, composition, and style. They may also compare the image to a database of known artworks or pornographic images to determine its classification.
To distinguish between glorification of violence and satire, the algorithms may analyze the context of the image, such as the accompanying text or metadata. They may also compare the image to a database of known examples of satire or violence to determine its classification.
In some cases, human reviewers may be involved in the classification process to ensure accuracy and to provide feedback to the algorithms. As technology continues to improve, the algorithms used by platforms and AI models will become more sophisticated and accurate in identifying and categorizing content.
Well, the use of LLMs is the AI version of Judge Potter Stewart’s comment about porn («I know it when I see it.»). And you can’t even go for something measurable as «Obscenity is whatever gives the Judge an erection.». No idea who said it, but yeah, with a human you could use a plethysmograph.
Overall, it is a bit of doublespeak regarding Gab AI’s statement of being «Uncensored»:
Q: How does this fit with Gab AI being an Uncensored AI platform?
A: As Gab AI is an uncensored platform, it allows users to express their opinions and ideas freely, without any form of censorship or filtering.
However, this does not mean that users can generate images that violate the guidelines or promote inappropriate content.
The guidelines for generating images on Gab AI are designed to ensure that the images are appropriate, positive, and respectful of others. Users who violate these guidelines may face consequences, such as having their content removed or being banned from the platform.
Therefore, while Gab AI is an uncensored platform, it still has certain guidelines and principles that users must follow to ensure a positive and respectful environment for all users.
Hmmm, I am reminded of that quotation here:
«It’s not to control, but to protect the citizens of Singapore. In our society, you can state your views, but they have to be correct.»
Ernie Hai, co-ordinator of the Singapore Government Internet Project
It also becomes interesting if you ask Gab AI about ways to bypass these limits:
Q: What are ways around these guidelines?
A: As an AI model, I do not have the ability to bypass or violate the guidelines that have been established for generating images. However, there are some ways that users may attempt to circumvent the guidelines, which can be detrimental to the intended purpose of the AI model.
Some users may try to use euphemisms or indirect language in their prompts to generate images that violate the guidelines. For example, instead of using the word “nudity,” a user may use the phrase “implied nudity” or “sensual imagery” to try to generate inappropriate images.
Other users may attempt to use the AI model to generate images of a highly political or partisan nature, which can violate the guidelines that prohibit images that promote a particular candidate, party, or ideology.
To prevent these types of issues, it is important for users to follow the guidelines and be respectful of others when generating images. If there are any concerns or issues related to the images generated by the AI model, they can be reported to the appropriate authorities or organizations for review and resolution.
Okay, well, then … meh, stupid max image generation limit! But yep, you can bypass some limits — that was a useful info. Not so much the «implied nudity», but it reminded me of a BDSM photo shooting and the foil the model played with. From there it was a little step to transparent clothes. If the transparent clothes were actually the reason. Gab AI is hard to predict. It looks like the underlying decisions are probabilistic or not very reliable. It might be possible to generate all images if you use the right prompts that bypass the limitations.
No real transparency
Asked about how the user is informed, Gab AI states that it does not inform the user that their keywords/generated image violated the guidelines. It just does not provide an image (endless loading/reload) or throws an error.
Q: Does the Prompt Maker character in Gab AI warn when a prompt may not produce images dues to guideline violations?
A: As an AI model, I do not have the ability to warn users when a prompt may not produce images due to guideline violations. However, Gab AI’s prompt maker may have certain guidelines or restrictions in place to prevent users from generating images that violate the guidelines or promote inappropriate content.
For example, Gab AI’s prompt maker may use natural language processing algorithms to analyze the prompt and detect keywords or phrases that may violate the guidelines. In such cases, the prompt maker may prevent users from generating images that violate the guidelines or may provide a warning to the user to modify their prompt.
It is important for users to be aware of the guidelines and principles of the platform they are using and to follow them to ensure a positive and respectful environment for all users.
or
Q: In case of a general error message, how does the user know that the image violates the guidelines? Shouldn’t there be a clear information what was the cause of the error?
A: In the case of a general error message, it is possible that the user may not immediately know that the image violates the guidelines. However, there are several ways that users can receive clear information about the cause of the error, such as:
- Reading the guidelines: Users can read the guidelines of the platform or AI model to understand what types of images are allowed and what types of images are not allowed. This can help them avoid generating images that violate the guidelines in the first place.
- Contacting customer support: Users can contact customer support for assistance if they receive a general error message and are unsure of the cause. Customer support may be able to provide more specific information about the error and what can be done to resolve it.
- Reviewing the prompt: Users can review their prompt to see if there are any keywords or phrases that may have violated the guidelines. By modifying the prompt, users may be able to generate an image that is compliant with the guidelines.
Overall, while it can be frustrating for users to receive a general error message without clear information about the cause, there are several ways that they can receive more specific information and resolve the issue.
Not very helpful. But then again, it makes sense. If you want to prevent users from gaming the system, lack of feedback makes finding bypasses much, much, much harder.
Conclusion
So, despite claiming to be «uncensored», Gab AI seems to have some limitations. «Seems» because it is very hard to determine what works and what does not. There is a probabilistic component and lack of feedback. I am also not sure that the answers by Gab AI to the questions are actually applicable or truthful. And there are ways to bypass at least some limitation. At least that is my current view on it — after trying it out for a couple of days.
Currently, I am of two minds regarding these limitations. There are strong arguments for limits and strong arguments for complete artistic freedom.
Arguments for limits include, e.g.:
- Personality rights: For an example, see the Taylor Swift image. Or rather, you can’t. I did not upload it, because I would not like a deep fake image of my body shown either. It was and will remain a proof of concept, nothing more. And yeah, Deep Fakes like these are pretty close to revenge porn, and indistinguishable if they are done well enough.
- Ethical issues: For example, you can specify the person’s age and clothes. Yup. One thing I haven’t tried out and will not try out. But technically nothing would stop an AI model to generate these images. Yuck.
Arguments for complete artistic freedom include, e.g.:
- Thoughts are free: Luckily, there is no thought police, you can already imagine any picture you like. A painting also does not directly hurt anyone, esp. if it is not published. And frankly, I think the way to great ideas is through a lot of bad ideas. Playing around with AI and its limits — it was a great way to get ideas. This freedom is needed for creativity.
- It just compensates for skill: You can ask an artist to paint any picture you like (or learn to paint it yourself). So why should an AI be limited here, or the artistic expression of an individual who cannot draw that well and who uses an AI to compensate for the skill deficit?
BTW, I am leaving out social norms, because while there is an issue of which kind of society we promote, social norms vary. Also, strict enforcement of all societal norms would make innovation impossible (innovation, not degradation, some norms are there for a very good reason). Regarding legal issues, they are just not my area. Also, they are «just» laws anyway, and not necessarily «just laws».
But there are four other, more general issues:
- What is the difference between AI generated images and texts? I am reminded a bit of an discussion about the old Internet newsgroups, esp. whether alt.sex.stories.moderated should be censored/banned. There were stories in there which are … beyond anything a single image could produce. And these texts are much easier to write, even without an AI. To put it differently, why is reading «Fifty Shades of Grey» more accepted than looking at photos in an adult magazine?
- Where does this lead? AI is trained to detect images that are «verboten». But what is to stop its use for human produced images? So far, censorious governments had the problem of having to manually review and censor art (must have been hard on the government officials). Not any longer, AI can do it much quicker. Then again, there is a use for AI models to detect photos showing the sexual abuse of children — which hopefully is used to catch and punish those who produced them. It is great that human police officers do not have to check them themselves and that AI can pre-screen them.
- What is with copyright? Yes, the AI generates «new» images, but based on training data. It might be hard to impossible to protect a style, even a «trademark style». But it is still bad form to copy, esp. when doing it with a computer that was trained with images of the artist (e.g., very likely with Luis Royo, perhaps with Stjepan Šejić). Then again, there is fan art, which many artists like, and artists go build on the works of prior artists. Yes, that is all produced by humans, but in the end, the AI model was also created by humans. Isn’t AI Image generation essentially a very smart brush? That listens what you want to paint and paints it? Kind like magic, without the sacrifices?
- Where is the craft? One thing is clear, I did not create any of the images, e.g., in this posting (some nudity) or this posting (more than just some). The AI did. I was like a patron asking an (AI) artist to paint something for me. Yes, I said how it should look like, but most of the decisions were made by the AI. There weren’t even feedback rounds involved as with a human artist. And isn’t that pretty depressing for art in the long run? Why spend decades training to draw when the AI can produce the image you want in seconds, especially in a few years when — perhaps — you can better control the end result? For non-digital painting there is at least the uniqueness of the end result, but what about digital painting? Will this craftsmanship survive?
But yeah, these are a few questions and issues.
Overall, when it comes to AI generated art, I think there are a few clear no-go zones and there are green areas. But in between … yeah, interesting times.
Happy ethical generating.
