After an embarrassing rollout and a bunch of attention from all the wrong people, Google forbade its Gemini AI from generating images of people — but, for whatever reason, it’ll still draw clowns. Which are people. Right? Right!?
This debacle began last week when anti-woke crusaders online got big mad about Gemini, which was formerly known as Bard until this rebrand, generating “racially diverse” images for prompts such as “American,” “Viking,” and “Founding Fathers,” that they expected to be entirely white. Adding to the fracas, it even depicted Nazis as people of color.
In response, Google apologized for vague and unspecified “inaccuracies in some historical image generation depictions.”
Initially, the tech giant took the entire image-generating capability offline. Now the image generation feature, which lives inside the chatbot itself, is back up, but it’s forbidden from depicting people. When you ask it to do so, it responds with a boilerplate statement that reads very similarly to a company blog post admitting that the AI “got it wrong” and that it would be pausing the image generation of people as it works on improvements.
“We are working to improve Gemini’s ability to generate images of people,” Gemini’s response reads. “We expect this feature to return soon and will notify you in release updates when it does.”
But while tinkering around with Gemini, Futurism found that the image generator is still happy to generate images of clowns inside specific environments. If you ask for just a clown it still demurs, but if you ask it for images of clowns in settings such as submarines or spaceships, it happily spits out goofy illustrations — or, in some cases, some unsettling and realistic-looking clowns.
What’s more, there are some pretty simple workarounds to get Gemini to draw pseudo-humans as well. The phrase “little guy” seems to work bizarrely well to this end, with the AI spitting out at least one shockingly creepy little guy when asked to do so.
As our session with Gemini went on, it became clear that the chatbot was wising up to our tricks. By the end of it, it was refusing to generate some of the images we prompted, such as “clown inside a submarine” and “little guy in a spaceship.”
But others, like the actually terrifying clown inside an amoeba below, still came up.
And eventually, it totally refused to do any more images at all.
The whole thing once again goes to show just how easy it is to circumvent AI guardrails — and how hard it is for the creators of AI, even an indescribably well-resourced one like Google, to anticipate the infinite number of potentially problematic things users could ask for.
We’ve reached out to Google to ask what’s up with the clowns and little guys that have creeped around its no-people-generating rules, but nobody from the company has gotten back to us yet. By the time they do, this particular jailbreak may be patched up — but rest assured, someone else will find another.
More on Gemini: Google Chatbot Refused to Say Whether Elon Musk Is Better Than Adolf Hitler
Share This Article