Addressing Racism in AI Image Generators: The Latest Updates in AI This Week

Keeping up with an industry as fast-moving as AI is a tall order. So until an AI can do it for you, here’s a handy roundup of recent stories in the world of machine learning, along with notable research and experiments we didn’t cover on their own.
This week in AI, Google paused its AI chatbot Gemini’s ability to generate images of people after a segment of users complained about historical inaccuracies. Told to depict “a Roman legion,” for instance, Gemini would show an anachronistic, cartoonish group of racially diverse foot soldiers while rendering “Zulu warriors” as Black.
It appears that Google — like some other AI vendors, including OpenAI — had implemented clumsy hardcoding under the hood to attempt to “correct” for biases in its model. In response to prompts like “show me images of only women” or “show me images of only men,” Gemini would refuse, asserting such images could “contribute to the exclusion and marginalization of other genders.” Gemini was also loath to generate images of people identified solely by their race — e.g. “white people” or “black people” — out of ostensible concern for “reducing individuals to their physical characteristics.”
Right wingers have latched on to the bugs as evidence of a “woke” agenda being perpetuated by the tech elite. But it doesn’t take Occam’s razor to see the less nefarious truth: Google, burned by its tools’ biases before (see: classifying Black men as gorillas, mistaking thermal guns in Black people’s hands as weapons, etc.), is so desperate to avoid history repeating itself that it’s manifesting a less biased world in its image-generating models — however erroneous.
Anti-racist educator Robin DiAngelo expresses in her popular book “White Fragility” that the ignorance of race or “color blindness” adds to systemic racial imbalances instead of lessening or remedying them. When individuals make claims of “not seeing color” or believe that merely recognizing the plight of individuals of other races is enough to consider themselves “woke,” they perpetuate harm by avoiding any substantial discussions about the subject, DiAngelo asserts.
Rather than avoiding the issue, Google’s delicate handling of race-based prompts in Gemini endeavored to hide the worst of the model’s biases. Some may argue (and many have) that these biases should not be ignored or downplayed, but instead should be addressed in the broader context of the societal training data from which they emerge.
It’s true that the data sets used for training image generators often contain more white individuals than Black ones. Moreover, the images of Black individuals in these data sets perpetuate negative stereotypes. This is why image generators tend to sexualize certain women of color, portray white men in authoritative positions, and generally favor affluent Western viewpoints.
Some might argue that AI vendors are in a no-win situation. Regardless of whether they address or ignore the biases in models, they face criticism. While this might be true, I argue that these models are missing explanations, constructed in a way that minimizes the manifestation of their biases.
Addressing the flaws within AI models honestly and openly, using simple and transparent language, would be much more beneficial than making hasty attempts to correct what is essentially an inherent bias that can’t be completely eradicated. The reality is that all of us possess biases and consequently, we don’t treat everyone the same. This extends to the AI models we create. Recognizing this truth would be beneficial.
‘Embarrassing and wrong’: Google acknowledges its inability to control its image-generating AI
Here are a few noteworthy AI-related stories from the previous few days:
The perception is that AI models are highly knowledgeable, but what do they truly comprehend? In reality, they don’t understand anything. Nonetheless, if you ask the question slightly differently, it becomes evident that they have internalized certain “meanings” that are similar to human comprehension. Even though no AI has a true understanding of what a cat or a dog is, could it have an understanding of similarity encoded in its representations of those words that differs from, for example, cat and bottle? Researchers from Amazon seem to think so.
Their research analyzed the “trajectories” of similar but different sentences such as “the dog barked at the burglar” and “the burglar caused the dog to bark,” compared to those of grammatically similar but logically different sentences, like “a cat sleeps all day” and “a girl jogs all afternoon.” They uncovered that the sentences humans perceive as identical were indeed internally regarded as more similar, despite their grammatical differences, and the opposite for the grammatically similar sentences. This paragraph might feel a bit baffling, but what it all comes down to is that the meanings encapsulated in LLMs appear to be more robust and nuanced than anticipated, not totally simplistic.
Neural encoding is providing beneficial outcomes in prosthetic vision, as discovered by Swiss researchers at EPFL. Artificial retinas and other methods of replacing segments of the human visual system typically have constrained resolution due to the limitations of microelectrode arrays. Regardless of the intricacy of the incoming image, it must be transmitted at a considerably low fidelity. However, there are various methods of downsampling, and this team discovered that machine learning performs excellently at this task.
Image Credits: EPFL
“We realized that when we utilized a learning-based method, we received better results in terms of optimized sensory encoding. But what was even more surprising was that by using an unrestricted neural network, it learned to reproduce aspects of retinal processing on its own,” stated Diego Ghezzi in a press release. In essence, it carries out perceptual compression. This was tested on mouse retinas, thus it’s not merely theoretical.
Researchers from Stanford University have unveiled fascinating insights into children’s drawing skills using computational vision. They received and evaluated 37,000 drawings of various animals and objects submitted by children and assessed how recognizable each artwork was based on the children’s feedback. Notably, it wasn’t the presence of distinct features such as a rabbit’s ears that enhanced the recognizability of drawings among other children.
Judith Fan, the lead investigator, explained, “The factors influencing the recognizability of older children’s drawings are not simply due to a single common feature that they all incorporate into their artwork. It’s a much more complicated process the machine learning systems are detecting.”
Chemists at the EPFL, discovered that LLMs are remarkably proficient at assisting with their tasks after minimal training. Instead of directly performing chemistry, these systems are fine-tuned on a wealth of knowledge that individually surpasses the abilities of chemists. For example, among thousands of research papers, there might be several hundred statements discussing whether a high-entropy alloy is single or multiphase. A system built on GPT-3 can be trained to answer this type of binary question and quickly starts making accurate extrapolations.
This is not necessarily a groundbreaking development, rather more proof that LLMs are effective in this regard. Berend Smit, one of the researchers, commented, “This process is as simple as conducting a literature search, which is suitable for many chemical problems. A foundational model might become a routine querying tool to kickstart a project.”
Lastly, Berkeley researchers shared a cautionary note, upon my re-reading, it’s evident that EPFL also contributed to this research. Kudos to Lausanne! The team identified that images sourced via Google often reinforced gender stereotypes associated with certain jobs and expressions, a phenomenon more pronounced than that observed in corresponding textual references. Moreover, a glaring disparity was observed in the number of men featured in both scenarios.
The researchers not only made these observations, but also conducted an experiment. They found that participants who researched a role using images, as opposed to reading text, were more likely to connect those roles with a specific gender, a notion which influenced their perceptions even after several days. “This isn’t constrained to the prevalence of gender bias online,” explained researcher Douglas Guilbeault, “The narrative is also about the profound and lingering impact image representations of people have that text simply doesn’t.”
In light of ongoing issues like the Google image generator diversity controversy, it’s important not to overlook the well-documented and oft-confirmed truth that the data source for many AI models demonstrates significant bias. This bias tangibly impacts real people.
Discover the pinnacle of WordPress auto blogging technology with AutomationTools.AI. Harnessing the power of cutting-edge AI algorithms, AutomationTools.AI emerges as the foremost solution for effortlessly curating content from RSS feeds directly to your WordPress platform. Say goodbye to manual content curation and hello to seamless automation, as this innovative tool streamlines the process, saving you time and effort. Stay ahead of the curve in content management and elevate your WordPress website with AutomationTools.AI—the ultimate choice for efficient, dynamic, and hassle-free auto blogging. Learn More