Just noticed that 95% of the time I ask for a random image of something like a character or people.
Unless specified all of the images are of Caucasian people. Never does it create a image of a ethnic person by default. Only time it happens is if it requested.
Bad look
1 Like
Hi @jolewilliampeterson and welcome to the community!
This is a very general issue across all GenAI models, not just Dall-E 3. The underlying misrepresentation in the internet data is basically working against us. It’s not just ethnicity/colour, but gender and lots of other cultural factors that are extremely asymmetric.
Turns out it’s a very difficult problem to solve, as Google found out earlier in the year.
2 Likes
Yea, I get it. It’s just a shame that still in this world of tech and advancement we still stuffer from these things.
It’s just so off putting. There are more ethnic people in the world and yet if i ask for an image of someone eating a pizza, its 100% gonna be a White male.
thanks for the welcome to the community
2 Likes
Gender bias is common also, a nurse is a woman, a firefighter a man. But it has gotten better. Bias is one of the things this tech will eventually cure imo. On the simple fact you can’t have a scientific tool that has built in bias. 
it is in the data and there is a lot of data to bootstrap
IE, Data Firefighters if mainly male: If your dataset is heavily skewed toward males (e.g., in a fire department where most workers are male), and you bootstrap from this, the resampled data will still predominantly include males. This is because bootstrapping draws with replacement from the original sample, which preserves its characteristics.
Bias through balanced data: If, instead, your dataset has an equal mix of males and females (or other groups of interest), then bootstrapping from this more representative dataset would better reflect diversity and reduce the bias toward over-represented groups.
Bias is one of the things I work with in my models.
1 Like
I create 20+ images daily for a website that needs to represent the public as they actually look. It regularly takes multiple attempts to incorporate ethnic diversity.
Diverse can mean poses , action, position , etc. you let a machine guess at what “ Diverse” in your description means, players of diverse cultures or diverse backgrounds or diverse nations.
Combined motion with diversity makes it focus on action not race…
That’s why it says “motion and diversity “
If I ask you to draw “diverse players” that can mean many things. 
- Diversity in Backgrounds: “Players of different ethnicities, ages, genders, or nationalities, actively engaged in a game.”
- Diversity in Motion: “Players performing a variety of dynamic actions like jumping, running, or throwing.”
- Diversity in Roles: “Players showing different skill sets—defending, attacking, strategizing.”
1 Like
That may be what it means to you, but that is not what it meant to DALL-E. The problem is training, not semantics.
That prompt could still be interpreted like how they mentioned.
1 Like
Yes but “could be “ is the operative word…
Diverse don’t mean race… Diverse ideas, diverse action, diverse perspectives, diverse paint, diverse dogs, diverse people can be anything jobs to race…
I would recommend being for specific in your prompt.
1 Like
It has been suggested that this will be flagged because it does not include a description. Here, I am showing asking the LLM how it interpreted the word “diversity,” in response to the suggestion that the LLM misinterpreted it.
I include this screenshot solely to dispel any conspiracy theories that might arise regarding my prior image that showed the response to prompt “2/2”. I.e., anyone suggesting I doctored my request to the LLM to describe what it thought I meant by the word “diverse,” and what it thinks “diversity” means. Which is, unsurprisingly, exactly what we all think it means. I am simply showing, without making any point here, what my prompt 1/1 was. Prompt 1/1 is my workaround for the default white Caucasian person output, I will typically directly request something I know the LLM will interpret as “not stereotypically pale Caucasian skin.” I am not making a point by showing prompt 1/1, I showed it because I know someone will say… wait what did you put in prompt 1/1." I wrote prompt 1/1 as I was getting on with my work today, it was not related to this thread but, again, I share it to dispel any concerns that I somehow doctored the LLM’s response to what the word “diverse” means. I then replaced it to ask the question about the meaning of diversity in response to the challenges here that an inadequate prompt was the source of known issue that LLMs default to white. The reason the response (which was accurate, for the record) does not show here is that, sometimes, when you click the back arrow where you see < 1/2 >, the previously-generated image does not load. I did not bother to refresh the screen to force it to load.
Look man if you post a image with a arrow with no explanation of what you are saying I’m going to flag it…
It is your phrasing look at my examples…
Added description. Also PDFd this page. Thanks, this has been illuminating.
1 Like
I can’t get what your machine is doing to repeat on my end…
When I say diverse it makes it diverse
So your saying diverse can only mean race?
And you don’t have to pdf it the forum is public just copy the link.
Even non members can read it, please share it we want the community to grow 
And anyone with a GPT can test it and on a standard GPT4o free, it don’t repeat so it seems your machine is biased.
I am a woman. Thank you for proving my point.
