OpenAI tweaks DALL·E 2 to fight bias, sees 12x diversity jump

OpenAI has started rolling out a system-level mitigation in DALL·E 2 designed to force more diverse outputs when users don’t specify race or gender in their prompts. It’s a blunt but apparently effective fix. The company’s internal testing shows users were 12 times more likely to say the resulting images represented people of diverse backgrounds after the change went live.

The technique kicks in on generic prompts like “firefighter” or “CEO,” where the model might otherwise default to generating a homogenous set of results based on skewed training data. OpenAI didn’t publish a detailed technical breakdown, but the move directly addresses one of the loudest criticisms of text-to-image generators since they burst onto the scene: they tend to amplify stereotypes. Early testers in the research preview were the ones flagging these sensitive and biased images, and that feedback loop is what informed this patch.

Beyond the diversity fix, the company detailed a handful of other guardrails hardcoded into the preview. You can’t upload a realistic face to use as a prompt, and any attempt to generate the likeness of a public figure—think celebrities or politicians—gets rejected outright. The content filters have also been tightened to catch policy-violating prompts without, ideally, gutting creative expression. It’s a tough balance, and the statement is honest about the system’s ongoing limitations, noting they’ll keep iterating as more data comes in.

Expanding access is the real stress test. OpenAI frames the wider rollout not just as a product launch, but as a core part of responsible deployment. You can’t fix what you can’t see, and real-world use is going to surface edge cases that internal red-teaming would never dream up. The 12x number is a compelling stat, but the real question is how this automated diversification holds up across thousands of users with genuinely weird and unexpected prompts. That’s the data they’re after now.