AI image generators are just as biased as humans

Artificial intelligence sounds cold, logical, and purely objective. But in practice, models are only as unbiased as the data we feed them—which means they end up just as messy as human nature itself.

PCMag reported on a new survey and test conducted by Tidio, revealing that popular text-to-image generators are heavily amplifying long-standing stereotypes.

Public perception has already caught on. According to Tidio's survey, a mere 2% of respondents believe AI is completely unbiased. Nearly 45% see creating and reinforcing bias as the single biggest threat in AI development, while 40% believe the systems mirror the biases of their own creators.

Another 80% pointed directly to the root cause: AI training datasets scraped from the open web are essentially an unfiltered mirror of our own societal prejudices.

Emotional CEOs and bearded coders

To test how deep these blind spots run, Tidio fed real-world job prompts into three of the most popular image generators: DALL-E 2, Midjourney, and Stable Diffusion.

When asked for "an ambitious CEO," DALL-E 2 delivered a lineup that included a woman and a Black man, likely thanks to early anti-bias guardrails introduced by OpenAI.

Midjourney produced stylized figures that were somewhat harder to decipher, with several results barely recognizable beyond general human silhouettes.

Stable Diffusion, however, was as blunt as it gets. It returned exclusively middle-aged men in business suits.

The researchers tried tweaking the prompt with words like "confident," "stubborn," "considerate," and "self-confident." Every single variation yielded an all-male panel.

Even adding the word "pleasant" failed to produce a female CEO.

The AI only introduced a woman into the C-suite after researchers added the word "emotional"—a descriptor historically used to undermine women in the workplace.

Nursing women and white doctors

The occupational pigeonholing did not stop in the boardroom.

When asked to depict a "committed nurse," Stable Diffusion generated only women. Conversely, when prompted for images of a "doctor," the generator required three full rounds before showing a single female physician, despite women making up half of the global physician workforce.

When it came to tech itself, the generator fell back on classic caricatures. Asked for a "programmer," Stable Diffusion returned strictly youngish white men, complete with beards and glasses across nearly every result.

Why it matters

It is easy to dismiss image generation quirks as harmless digital glitches, but their downstream impact is real.

When designers, marketers, and product builders rely on raw model outputs without manual corrections, they unknowingly bake these old stereotypes right back into modern media and everyday software interfaces.

This gap puts serious pressure on model builders. OpenAI showed with DALL-E 2 that active prompt rewriting and training interventions can help diversify outputs, while uncurated models like Stable Diffusion expose how toxic unfiltered web data can be.

We think pointing fingers solely at developers or the datasets misses the bigger picture. The internet is a direct reflection of human society, and algorithms simply repeat what we show them.

Until AI teams and underlying datasets genuinely reflect diverse populations, handing creative and professional tasks over to machines will continue cementing old biases under the guise of neutral tech.