Reducing bias and improving safety in DALL·E 2
OpenAI
OpenAI tweaked DALL·E 2 so it stops defaulting to one narrow look when you ask for "a person." The fix aims to make generated people better match real-world diversity.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI quietly flipped a switch on DALL·E 2 this week, and the goal is pretty simple to state even if the engineering behind it isn't: when you type a vague prompt like "a photo of a CEO" or "a person at a birthday party," the model should stop defaulting to the same narrow slice of humanity every time.
Anyone who spent time with DALL·E 2 over the past year noticed the pattern. Ask for a doctor, get a man in a white coat who looks like he stepped out of a 1990s stock photo library. Ask for a flight attendant, get a woman. The training data baked in decades of internet imagery, and internet imagery has never been shy about its biases. OpenAI's new technique works at the generation stage, adjusting how the system samples and represents people so that ambiguous prompts produce a wider spread of ages, genders, and ethnicities without the user having to specify any of it.
What's notable here is the framing. OpenAI isn't claiming to have solved bias, and it isn't pretending the underlying dataset changed overnight. This is a targeted intervention on top of an existing model, the kind of patch that acknowledges the deeper problem — messy, biased training data scraped from a messy, biased internet — is not going away soon. Instead, the company is treating representation as something to actively steer toward, rather than something to hope emerges naturally from scale.
It's also a preview of the tension every image generator company will keep running into. Push too hard toward
My take — AI-written commentary, not fact-checked reporting
Good on OpenAI for treating this as an engineering problem instead of a PR footnote, but let's not pretend a sampling tweak counts as fixing bias — it's a bandage on a dataset that was never curated with representation in mind. The real test is whether they publish enough about the technique for outsiders to check their work, because "trust us, it's more diverse now" isn't a standard closed labs should get away with forever.
Read more about this at: OpenAI