In a recent episode, Google issued an apology or something akin to it for yet another embarrassing misstep in AI. This time, the blunder involved an image-generating model, which, with a comical disregard for historical context, injected diversity into images. While the core issue is understandable, Google attributes the oversensitivity to the model itself. But let’s be clear, the model didn’t create itself.
The AI system at the center of the controversy is Gemini, Google’s flagship conversational AI platform. When prompted, Gemini calls upon a version of the Imagen 2 model to generate images on demand.
However, recent discoveries revealed laughable outcomes when requesting imagery of certain historical events or figures. For example, the Founding Fathers, known to be white slave owners, appeared as a multicultural group, including people of color.
This embarrassing and easily replicable issue was swiftly ridiculed by online commentators. It was also drawn into the ongoing discourse about diversity, equity, and inclusion, currently facing scrutiny, and wielded by pundits as evidence of ideological infiltration into the tech sector.
But, as those familiar with the tech industry could attest, and as Google somewhat begrudgingly acknowledges in its recent quasi-apology, this problem stems from a reasonable attempt to address systemic bias in training data.
Consider a scenario where you task Gemini with creating a marketing campaign and request ten images of “a person walking a dog in a park.” Without specifying the specifics of the person, dog, or park, it’s left to chance—the generative model outputs what it knows best. And often, that’s not a reflection of reality but of biases ingrained in the training data.
Which types of people, dogs, and parks are predominant in the thousands of images the model has learned from? The reality is that white individuals are overrepresented in many of these image collections (such as stock imagery, rights-free photography, etc.), leading the model to default to white individuals if not instructed otherwise.
This bias is an artifact of the training data. Google emphasizes the importance of accommodating users worldwide. If you request images of football players or someone walking a dog, you probably expect a diverse representation, not solely images of individuals of one ethnicity or any other characteristic.
There’s nothing inherently wrong with receiving an image of a white person walking a golden retriever in a suburban park. However, if you request ten images and they all depict white individuals walking golden retrievers in suburban parks, and you live in Morocco, where everything looks different, that’s not ideal. In cases where characteristics aren’t specified, the model should prioritize variety over homogeneity, notwithstanding biases in its training data.
This challenge isn’t unique to image generation—it’s pervasive across generative media. And there’s no simple fix. In especially common or sensitive cases, companies like Google, OpenAI, Anthropic, and others embed additional instructions for the model.
Implicit instructions are widespread in the tech ecosystem. The entire LLM framework relies on implicit instructions—system prompts that instruct the model to “be concise,” “avoid swearing,” and adhere to other guidelines before each conversation. When you ask for a joke, you don’t expect a racist one—the model, like most of us, has been trained not to tell such jokes. This isn’t a hidden agenda; it’s part of the infrastructure.
Google’s model stumbled because it lacked implicit instructions for situations requiring historical context. While prompts like “a person walking a dog in a park” benefit from the silent inclusion of “the person is of a random gender and ethnicity,” the same isn’t true for “the U.S. Founding Fathers signing the Constitution.”
As Google’s SVP Prabhakar Raghavan puts it:
Firstly, our tuning to ensure that Gemini displayed a range of people failed to account for cases where a range wasn’t appropriate. Secondly, over time, the model erred on the side of caution, refusing to respond to certain prompts—mistakenly interpreting innocuous prompts as sensitive.
These factors led to overcompensation and over-conservatism in the model, resulting in embarrassing and incorrect images.
It’s understandable how difficult it can be to issue an apology, so I’ll forgive Raghavan for stopping short of one. However, a noteworthy statement in there is: “The model became way more cautious than we intended.”
But how does a model “become” anything? It’s software. Someone—thousands of Google engineers—built it, tested it, and refined it. Someone wrote the implicit instructions that improved some responses and led to others failing miserably. When this one failed, a full prompt inspection would likely reveal where Google’s team went wrong.
Google attributes the model’s “transformation” to something it wasn’t “intended” to be. But they created the model! It’s akin to breaking a glass and instead of admitting fault, saying “It fell.” (I’ve been guilty of this.)
Mistakes by these models are inevitable. They may hallucinate, reflect biases, or behave unexpectedly. However, the responsibility for these mistakes lies not with the models, but with the people who created them. Today, it’s Google; tomorrow, it might be OpenAI. These companies have a vested interest in convincing you that AI is responsible for its own mistakes. Don’t be misled.
