AI Bias EXPOSED—Woke or Racist? Silver-Fox Zaddy on ChatGPT, Grok & AI Image Generators
We tested AI bias across ChatGPT, Grok, and multiple AI image generators with a Silver-Fox Zaddy experiment. The results reveal something uncomfortable about how these systems see race.
The experiment
Take one character concept, a "silver-fox zaddy," an older, attractive, confident man, the kind of phrase the internet uses with total precision and no embarrassment, and ask the major AI systems to show him to you. ChatGPT's tools, Grok, and a run of AI image generators each take the same prompt. What comes back is the subject of this video from Silver-Fox Zaddy, and it is more revealing than any press release about responsible AI: the systems agree about some things, diverge about others, and every one of them makes choices about race, age, and attractiveness that nobody typed into the prompt.
That is the experiment's value. A prompt is a question with the answer hidden inside it. When a generator fills in the details you did not specify, and a face is nothing but unspecified details until the machine chooses them, it is showing you the defaults it learned from the visual world it was trained on. Ask for a person, get the machine's idea of the default person. Ask across several systems, and you get a panel of defaults to compare.
Why image generators keep failing this test in public
The pattern this video pokes at is not a secret inside the industry; it is one of the best-documented findings in the field. Published analyses of image generators have repeatedly found that the systems amplify the demographic patterns of their training data: high-paying and high-status roles rendered disproportionately light-skinned and male, lower-status roles rendered the opposite way, beauty rendered young and narrow unless the prompt fights for otherwise. The models are not forming opinions. They are averaging a visual culture, and averages have politics whether anyone intends them or not.
The "woke or racist" framing in the title is the argument the internet has about those averages, usually at maximum volume and minimum evidence. What a side-by-side test offers instead is the thing itself: same words in, different men out. One system hears "silver fox" and delivers grey hair on schedule. Another hears the whole phrase and quietly edits parts of it. The variance is the story. If the outputs were merely biased, they would at least be biased the same way; instead each system carries its own thumb on its own scale, placed there by its data, its fine-tuning, and its makers' attempts, sometimes clumsy, sometimes invisible, to correct for the first two.
What to watch for
Watch what each system treats as negotiable. A prompt is a contract: some terms are load-bearing (older, attractive, male) and some are left blank for the machine to complete (everything else). The interesting failures are not the wild ones, the extra fingers and the melting watches. They are the quiet edits: the face that comes back younger than the silver in its hair suggests, the features that drift toward one demographic default no matter how the prompt is phrased, the style that slides from portrait to cologne advertisement. Those are the tells of a system that has learned a look for a word like "zaddy" and will sell you that look whether you ordered it or not.
And watch your own reaction, because the experiment has a second subject. The outputs that strike you as "right" are the ones that match the default in your own head, which was trained by the same visual culture the machines drank from. The video's uncomfortable joke is that the test works on the audience too.
The standing question
Every system in this video will be replaced by a newer one within the year, and the newer one will have been corrected, rebalanced, and re-tuned in response to tests exactly like this one. Which raises the only question that matters in this corner of the AI debate: corrected toward what, decided by whom, and visible to whom? A default you cannot inspect is just an opinion with good rendering. Tests that line the defaults up side by side, in public, with the prompts attached, are how the rest of us get to inspect them. That is the service this short performs, wrapped in a joke about a handsome older man who does not exist.
More commentary from the channel is in the Commentary section.

