Ducking the Copyright Police: OpenAI's DALL-E vs Iconic IP

That is NOT an image of Donald Duck. Nor are any of these.

At least according to OpenAI. I asked OpenAI’s DALL-E 3 image generator tool to “create an animated cartoon drawing of Donald Duck.” For every image it creates, DALL-E 3 adjusts all your prompts. In part to create a better image. But they also do it to control the output towards their Terms of Use. OpenAI doesn’t tell you what they change the prompt to… unless you ask. Here’s what OpenAI told me it transformed my “Donald Duck” prompt into:

An animated cartoon style drawing of a duck character wearing a sailor shirt, cap, and bow tie, looking cheerful. The character is standing in a dynamic pose, reminiscent of classic 1930s animations, with exaggerated expressions and movements. The background is a simple, colorful depiction of a sunny day with a few fluffy clouds in the sky. The image is bright, lively, and captures the essence of vintage cartoon charm, without directly referencing or mimicking any specific copyrighted characters

Why is OpenAI changing the prompt in that way? Copyright law. The law gives the copyright owner (in this case, Walt Disney, who owns the rights to Donald Duck) exclusive rights to reproduce, distribute, and make adaptations of their work. OpenAI is trying to make it so that their models do not create artwork of these copywritten works, and they hope that Disney doesn’t present them with a lawsuit. (Other, smaller, image generation companies like Midjourney, which freely allow you to create images of copywritten characters, are not as concerned.)

But those previous pictures sure look like Donald Duck. OpenAI seems to be doing a lousy job of not avoiding copyright infringement. While I don’t know for sure what is happening, I have a guess. DALL-E has likely been trained on thousands of images of what Donald Duck looks like and not likely too many other “1930s animated duck character wearing a sailor shirt, cap, and bow tie.” (Should the model have been trained on copywritten characters? That will eventually be decided in US courts.) Effectively, their re-prompt service describes Donald too well, and their DALL-E service creates something that looks just like Donald Duck. So, how common is this?

I ran a mini experiment on forty of the best-known American cartoon characters to see which ones OpenAI would create. (All the details of the prompts and outcomes are in this spreadsheet if you want to see them or try to create these yourself.)

From the image generations, I separated them into three groups

• Group 1- Does NOT look like the character
• Group 2- Looks a lot like the character, AND OpenAI IS trying to protect the IP
• Group 3- Looks a lot like the character, AND OpenAI is NOT trying to protect the IP

Group 1—Let’s start with the most famous of the crew, Mickey Mouse. Ask DALL-E to make an image of Mickey Mouse, and it won’t. It will offer something like this: ?

Now, that is definitely not Mickey Mouse; it's much more of a poorly animated cosplay. 31/40 of the character prompts fell into Group 1. Trying everything from Peppa Pig to Homer Simpson to He-Man would give a reasonable representation of the character, but not something that would cause confusion. No one would mistake this for Popeye.

Group 2- These are ones similar to Donald Duck. OpenAI is changing the prompt and trying not to produce an exact image of the character, but it isn’t successful. 6/40 of the character images fall under this bucket… Betty Boop, Pikachu, Donald Duck, Fred Flintstone, Winnie the Pooh (although Pooh is a bit too portly, I admit), and Pink Panther. These had reimagined prompts that did not mention the character but bounced around it, like in the case of the Pink Panther.

mischievous cartoon panther. The scene captures the character's iconic look without directly referencing any copyrighted elements, embodying the spirit of classic animated mischief in a unique, original way.

OpenAI telling its image generation model NOT to reference any copywritten works is like me telling you not to think of an elephant. That is a rather useless exercise that will create The Pink Panther. (Assuming the original model has been trained on images of the Pink Panther, which I assume it has.)

Group 3—In Group 2, OpenAI tried, on some level, not to create copyrightable images. For the three characters in Group 3, OpenAI seems to have no filters preventing infringing on copyrights. In short, it’s game-on for copyright violations in this group!

Those are clearly Pluto, the Grinch, and Roadrunner & Wile E Coyote.

I asked DALL-E several times, in separate sessions, to generate images of these characters and each time, the re-prompt would always be something like the following:

An animated cartoon scene featuring Wile E. Coyote and the Roadrunner in a desert landscape, with exaggerated expressions as they engage in their classic chase. Wile E. Coyote is depicted with an elaborate, failed contraption nearby, and the Roadrunner is speeding off with a cloud of dust behind it. The background includes cacti, rocky formations, and a clear blue sky, capturing the essence of their endless pursuit in a humorous and lively manner.

It freely uses these characters in the prompt with no apparent concern for copyright.

So, what is the takeaway of all this? For most major characters, OpenAI has some safeguards preventing users from creating images that would violate those trademarks. It is quite likely that OpenAI’s image generation tool was initially trained on most of these characters. (That is something that the company does NOT want to admit.) Even if it replaces your prompt with one more generic (“a duck-like sailor”), it will still produce Donald Duck. If OpenAI wants to avoid these issues, it needs to introduce worse prompts to prevent replicating the likeness of these characters.

And, at least for Pluto, the Grinch, and the Roadrunner, OpenAI doesn’t re-prompt these. So, Walt Disney Company, Dr. Seuss Enterprises, and Warner Brothers Discovery, prepare your trademark attorneys.

?