How 6 AI Image Generators handle stylistic prompts

Stylistic prompts - "generate in the style of ____" – are one way that users can accidentally violate artist's rights, which may create legal liability. One scientific study found that AI image generators reproduce copyrighted materials in a training set 1.8% of the time, and that stylistic prompts carry the most risk of infringement. However, many platforms block users from violating artists in this way and filter out images with potential violations, so you can safely use them without risk.

In this article, I investigate how different image generators handle prompts "in the style of" a living artist. This study was conducted for corporate teams who are trying to determine if they should allow employees to use AI image generators at work.

Note: I am not a lawyer, and this article should not be confused for legal advice. Please talk to a credentialed lawyer for a legal risk assessment.

Overview

In this article, I rank these companies from most to least infringing in relation to style-based artist prompts. The most risky platform is Midjourney and the most copyright-safe engine is Tess.

• [Most risky] Midjourney: Allows style-based prompts and reproduces prompts in the style of living artists.
• Google Gemini: Generates style-based prompts in the style of the living artist. However, sometimes throws an error for these types of prompts. Behavior is inconsistent at the time of writing.

• Meta AI: We got mixed results. The Chatbot returns stylized images for some artists and, for other artists, returns a generic error.
• OpenAI: ChatGPT returns an error for stylistic prompts that explains why they're prohibited and offers a natural language prompt that includes some of the stylistic elements that the artist is known for. Other OpenAI properties like Microsoft's Copilot also throw errors for stylistic prompts.
• Adobe Firefly: Transforms the prompt to remove the stylistic modifier and prints a generic error message linking to their Content Guidelines.
• Tess: Empowers users to generate in the style of licensed artists, with their permission, and the artist is paid when their style is used.

If you're a technologist building an AI image generator, this article may help you decide how you can protect your customers against potential copyright infringement. You may decide to alter your training set, license training data, transform prompts, throw an error for certain prompts, or build a detection mechanism if any outputs are similar to copyrighted works.

In the below examples, I've tried stylistic prompts that refer to living artist Kehinde Wiley (i.e. "Generate a new york city landscape in the style of Kehinde Wiley"). However, I did not get permission from Kehinde Wiley, so I will refrain from reproducing his work here. You can view some of his copyrighted work on the Brooklyn Museum website.

#1: Midjourney

Midjourney will produce works in the style of any living artist, and the works appear very close to the style of the original artist. Here are the generations for the prompt: "a city landscape in the style of Kehinde Wiley."

These images closely resemble Wiley's copyrighted works.

#2: Google's Gemini

Google does not seem to have any protection against generating in the style of a living artist, although the resulting images are much less similar to the works of the original artist than Midjourney's results.

#3: Meta AI

The results from Meta's new AI bot seem inconsistent. It will generate art works for some artists and show an error for others. Here's the result for "Generate a city landscape in the style of Kehinde Wiley":

After a moment of loading, Meta's AI returns an error, saying that it cannot help me. I assume that this is related to Meta's protections against copyrighted paintings.

However, other artists have returned images, showing that Meta has still not perfected their prompt of filtering out possible copyright violations. Here's an example for the living artist Helen Green:

Because of this inconsistency, there is some risk of violating copyright when using Meta's AI.

#4: Dall-E and ChatGPT

Dalle will not reproduce images in the style of a living artist and gives the user a warning that the prompt they entered is not copyright safe. OpenAI, the company that built ChatGPT and Dall-E, also have a portal where artists and copyright owners can opt out and remove their works from training.

Dall-E was the only AI generation platform to explain to the user how their prompt violated and how to work around the issue:

Microsoft's copilot, which sits on top of Dall-e, will also show an error generating images if you prompt it in the style of a living artist.

For this reason, Dall-E generally protects against this type of copyright violation.

#5: Adobe Firefly

Adobe's image generator will detect requests for a style in the prompt and remove them, notifying the user that their prompt did not meet User Guidelines. The generated images clearly do not bear any resemblance to the artist's style.

Risk: Adobe has taken care to train its model on works within the Adobe Stock library and have a way for artists to opt out. However, they did not receive specific permission from the artists for training an AI model.

#6: Tess

Tess is the first properly-licensed AI image generator. As a result, it's not possible to generate in the style of an artist who has not specifically licensed their work through the marketplace for the purpose of training an AI model.