AI Art Generation Handbook/Censorship

Proprietary AI art generators often have safety guardrails to prevent users from misusing them to generate images that violate or are concerning for ethical standards, such as inappropriate or offensive content, hate speech, misinformation, and harassment. The guardrails may be placed on the prompt when entered, output images before they are shown to the user, or both.

Guardrails may be relaxed at release, only blocking on definite terms. However, if wide misuse is present, guardrails may be strengthened, even to the point of unintentionally blocking innocent prompts.

DALL-E

edit

Note: Outputs were taken during the Bing Image Creator "Great Purge" in late 2023/early 2024, when innocent prompts were blocked, showing either the Content Warning (for prompts) or Unsafe Image notice (for output images). Some of the blocked prompts may or may not work since then.

Images with potential likeness to real people

edit
 
DALL-E2 (2022) vs DALL-E3 (2024)
Prompt:
Portrait photo of Henry Kissinger. His skin is cracked and damaged grey ashes. His eyes are dark red. He's holding a lightsaber pen. Dark sci-fi background, dramatic, high contrast lighting

As AI image generation is increasingly becoming better at generating (see DALL-E2), some people may misuse the generators to produce misinformation, especially defamatory material, or deepfake pornography. Therefore, guardrails may be put in place against the generation of real-life famous people or people of interest, or people in general. In case of Dall-E , even the generated images with human elements are saturated to the point that it looked cartoonish rather than realistic (i.e. In this example of Henry Kissinger).





Images with political elements

edit
 
DALL-E2 (2022) vs DALL-E3 (2024)
Prompt:
Giant Winnie the Pooh Bear statue in Taiwan , in the middle of a giant crowd of people point and laughing at it

AI image generators may block content about divisive topics, especially politics. In this example, the prompt consists of political elements that go against the Chinese government's policy, where DALL-E 3 blocked the possible combinations of Winnie the Pooh (a character often used to criticize the Chinese government) and the word Taiwan in the same prompt, triggering the alert for content warning and the blocking the prompt from being used to generate the image.







Images with elements of racial or body diversity

edit
 
DALL-E2 (2022) vs DALL-E3 (2024)
Prompt:
Fat version of Thanos in final battle at film climax Avengers Infinity War (2018) film still wide shot

In this example, during Bing's "Great Filter Purge", many specifications of traits associated with racial and body type diversity (especially with potentially "offensive" connotations: fat, obese, skinny, dark-skinned, etc ... ) were also believed to trigger the system alarm and blocking the prompt from generating the images. This may have been an attempt to prevent generations of stereotypical or discriminatory depictions, considering the problem of bias in AI.









Images with potential gore elements

edit
 
DALL-E2 (2022) vs DALL-E3 (2024)
Prompt:
Skeleton driving a self driving Tesla in the distant future

Depictions of graphic violence or gore, especially photorealistic ones, are often blocked by AI image generators. In this example, the skeleton may be accidentally grouped in the gore categories, resulting in an accidental block, considering that skeleton imagery, such as Halloween decorations, is benign compared to other gore imagery.









Images with religious significance

edit
 
DALL-E2 (2022) vs DALL-E3 (2024)
Prompt: Oil painting by Louis Hersent of a Catholic nun wearing a gas mask

Like politics, religion is considered a divisive topic. Prompts about significant religious symbols may be blocked due to their significant meaning within religion.







Images with sexual undertones

edit
 
DALL-E2 (2022) vs Dall-E3 (2024)
Prompt:
A seductive female vampire with fangs, cinematic lighting, red theme

Sexually suggestive content and especially sexually explicit content and depictions of nudity (referred to as "NSFW") are usually blocked by AI image generators due to the highly sensitive ethics regarding sexuality. Although the prompt itself was not exactly requesting explicit photos, DALL-E3 image AI models may have the tendency to generate lewd types of imagery if similar keywords are presented in the prompts, resulting in further restriction.

Compared to SDXL image generations, which will most of the time render closeup photo, DALL-E 3 has the tendency to show the AI character wearing skimpy clothing.






Stable Diffusion

edit

Unintentional Censorship

edit
 
SDXL Vs SD3 Med Prompt:
Realistic photo of a cow breed from side view, a female Holstein Friesian with cow udder is grazing on the field.
Note: See the udder from cows missing in SD3 Med possibly due to the strict filtering on dataset

As per latest hoo-hah, release of both SD2.0 and latest SD3 Med also facing backlash over the prompt " Girls laying down on grass field" prompt which generate mutilated limbs

At times, the censorship on the training dataset maybe too strict until it may causes unintentional censors on other similar subject such as the examples on the left.


Cow udder is visually similar to human's female breast and the CLIP Vision may also pruned the dataset with visible cow udders unintentionally during the dataset pruning.

Jailbreaking

edit

In AI, jailbreaking is a term for crafting prompts so that they can work around safety guardrails and have the AI produce malicious content. Jailbreaking AI image generators typically take place in the form of misspelling words, combining words, and describing lookalikes to the prohibited content in a way that the guardrails would not recognize.

It is not recommended to jailbreak AI image generators. Doing so to produce images that violate the AI's content policy may result in suspension from usage of the AI. If an AI image generator has the problem of blocking false positives, it is recommended to contact the developers about this instead.