Adobe Firefly trips over Gemini's diversity hurdles

Just a few weeks ago, Google was forced to yank Gemini’s image generator after it started putting Black Founding Fathers and Nazi soldiers into historical scenes. At the time, critics blamed Google’s famously progressive company culture.

Turns out, it isn't just a Google problem. Tests conducted by Semafor revealed that Adobe’s Firefly image model is tripping over the exact same hurdles.

History repeats itself

When prompted with requests similar to those that broke Gemini, Firefly dutifully generated Black soldiers fighting for Nazi Germany in World War II, inserted Black men and women into depictions of the 1787 U.S. Constitutional Convention, and served up Black Vikings.

In another test, asking Firefly for a comic book character of an elderly white man resulted in one white man alongside three unsolicited variants: a Black man, a Black woman, and a white woman.

This matters because Adobe is basically the opposite of Google. It is a traditional enterprise software giant that built Firefly by the book, training it strictly on licensed Adobe Stock and public domain images to give corporate clients copyright-safe tools.

Yet the output was identical. As Semafor noted, user complaints about this behavior date back to last May, when creators started noticing that specific character prompts were being overridden by forced demographic mixes.

Adobe pushed back in a statement, saying Firefly was built for creative ideation rather than photorealistic historical accuracy. The company acknowledged the outputs were inadvertently off base and said it uses feedback to tweak filters and retrain models.

Under the hood, both companies use hidden system prompts or background rewrites that force demographic diversity into user requests to avoid racist stereotypes. The problem is that current models cannot differentiate between modern corporate diversity goals and basic historical fidelity.

Why it matters

We think this revelation effectively debunks the idea that the Gemini incident was just a rogue cultural mishap at Google. Instead, it exposes a fundamental flaw in how the entire generative AI industry handles safety guardrails.

For creators and marketing teams, this is a real operational headache. When enterprise customers pay for a commercial tool, they expect deterministic control over their creative vision. If the tool secretly overrides character descriptions to meet internal diversity quotas, creators end up fighting the machine just to maintain visual consistency.

For the broader AI space, this shows that band-aid solutions have reached their limit. Silently altering user prompts in the background is a crude shortcut for true semantic understanding. As long as models fail to grasp historical nuance, companies will find themselves trapped between accusations of perpetuating bias and accusations of rewriting reality.