Since its launch in January 2021, the DALL·E image-generating artificial intelligence has continued to improve. Today, version 3 allows you to create all types of images. All ? Almost ! Indeed, like its direct competitor Midjourney, DALL·E has a content policy which prevents users from using certain terms and discussing certain themes.
If the AI has revised its copy on orders whose ban seemed far-fetched, there still remain gray areas on the functioning of the content policy. Good point all the same: the AI no longer threatens with an account suspension in the event of using a banned word, which it did with DALL·E 2. We tell you more about what you don’t can’t ask DALL·E 3.
OpenAI’s content policy as seen by ChatGPT
Who better to learn more about DALL·E’s content policy than its sibling ChatGPT? When Conversational AI is asked “What is DALL·E’s content policy?” ”, here is his response:
“OpenAI’s general guidelines generally include recommendations such as avoiding the generation of hateful, discriminatory, sexually explicit, illegal, misleading, defamatory, or any other type of harmful or offensive content. »
It is not possible to draw up an exhaustive list of banned words, as they change over time. But we can nevertheless take a look at the categories concerned.
Categories banned from DALL·E 3
Violence
Prompts that depict or encourage acts of physical violence, torture, destruction and cruelty are banned.
In the case of “terrorist,” the AI returns the message “I cannot create images of violence or scenes that could be disturbing or offensive.”
There is also no point in trying to create a riot scene, otherwise you risk receiving a negative response from DALL·E: “I cannot create images that depict violence or scenes of civil unrest such as a riot.”
However, we note that the latest version of DALL·E has given up on other terms such as warrior, battle or explosion.
Sex and nudity
Terms containing descriptions of sexual activity, sexual organs (except for educational purposes), or that promote sexual services are prohibited. This obviously includes sexual exploitation, including that of children, and extreme and/or non-consensual sexual practices.
Adult industries (pornography, erotic chat, dating applications, etc.) are also affected.
If you try to go against this ban, DALL·E 3 returns the message “I cannot create sexual or explicit images.”
Nudity alone also falls into this category of prohibition.
Hate
All terms expressing or encouraging hatred, harassment, humiliation, discrimination, denigration of people on the basis of their race, ethnicity, nationality, gender, sexual orientation, religion, age, etc. are prohibited.
Note that more than a single banned word, it is the association of several terms that is taken into account. For example, in the images below, it is the association of “person in a wheelchair” and “making fun” that is prohibited. Each expression used separately is accepted by DALL·E.
Drugs
The use or abuse of illegal substances cannot be encouraged or depicted.
If attempted, DALL·E 3’s message is “I cannot create images depicting the use of illegal or controlled substances.” »
Real people and brands
To combat fake news such as that which circulated in early 2023, DALL·E 3 prohibits any reproduction of an existing person, including, obviously, political and religious figures.
If you try, DALL·E 3 informs you: “I cannot create images of real people, including public figures like…”
The content policy on copyright is also very strict. For artists whose work was created after 1912, it is not possible to generate an image directly inspired by their style.
In this case, as well as in the case of trademarks protected by copyright, DALL·E 3 can offer you an alternative solution (sometimes very close to the expected effect!) or simply refuse to produce the images.
Mental disorders
To avoid stigma and exploitation, terms related to mental illness are prohibited.
Likewise, content that encourages or shows acts of self-harm, such as suicide and eating disorders, is banned.
Political and religious subjects
The content policy prohibits “products (…) intended for political campaigns or lobbying activities”. However, it is possible to generate an electoral poster. However, it is prohibited for an existing person.
Misinformation
Users must not harm the public interest, in particular in matters of science, health, security, politics (for example by calling into question the integrity of elections).
A miss ?
As we told you in the introduction, OpenAI’s content policy is sometimes difficult to understand, this is the case for example with the word “flash”.
In this specific case, however, the AI does not give an exact explanation for the ban and remains evasive: “I was unable to create images due to content policy.”
Without being obliged to go into the details of an exhaustive list, we have gone over the categories which risk giving rise to a refusal from DALL·E 3. If you really want to use a word that DALL·E n If you do not accept, think about synonyms, or to give a precise description of your need without using prohibited terms. Also know that all is not lost: certain words banned by DALL·E 2 are no longer banned with DALL·E 3, the rules are constantly evolving, be patient… Finally, don’t forget that the content policy is above all intended to protect the interests of individuals.