Google's Gemini Dodges Depictions of Diverse Races
In a world envisioned by Google's Gemini, Nazis are black, American revolutionaries are Asian and historical injustice captured by AI can be remedied by a band aid.
Only two weeks after Google’s Gemini was released, it was taken offline until further improvements could be made. While Google’s latest AI models have all failed in one way or another, a rather unexpected problem has come up this time. The model struggles when it comes to generating images of white-skinned people. It refuses to produce images of people, historical or otherwise, that are white, and tends to represent public or historical characters in a variety of inaccurate races.
For a look at Google’s predecessor AI, as well as its shortcomings, check out my piece exploring bias and flaws in chatbots.
Recently, Google’s Senior Vice President wrote in a blog post that he 'won’t promise that Gemini won’t occasionally generate embarrassing, inaccurate or offensive results.' This statement raises a critical question: Should Google really be releasing AI products to the marketplace before they undergo rigorous validation? After all, when’s the last time your microwave cooked your food slower based on the color of your eyes? Or when have you had to wear a mask to fool the doorbell lens before you could unlock your front door? Given the recent AI Acts in the EU and USA, along with a greater public awareness of the need for responsible technology use, it’s time for Google to be subjected to much higher scrutiny regarding its products.
For an overview of the topic of bias and fairness in AI, check out my piece covering the ethical, legal and technological aspects, as well as various case studies showing how bias can be a problem across different settings.
Google has a rather ugly track record when it comes to racial bias in their image-based AI. In 2015, for instance, Google Photos was criticised for tagging black people as “gorillas”. Later in April 2020, Google's Vision AI displayed racial bias when labeling images. The system labeled a dark-skinned individual holding a thermometer as "gun," while a similar image with a light-skinned individual was labeled as "electronic device."
Google apologized for these incident but instead of fixing the issue directly, they chose to remove the "gorilla" label from their system, even for actual pictures of gorillas. In the same vein, Google, anxious about generating images that do not overfit, fumbles its way to generate images for every race except White/Caucasian. At a certain point this much overfitting becomes marked, given that 12% of the world are not represented in the model’s image generation capabilities. Rather than dealing with the bias by retraining the model to have more balanced and accurate associations between images and people, Google has instead opted to hide the symptoms. Google Photos’ capacity for making racist classifications cannot be solved by merely avoiding particular instances. Such an approach obscures the harder to detect problems, the more fundamental ones that reveal different kinds of bias from low level systems to high level ones, as well sea of socio-economic activity that creates the data used in these systems.
Google has also addressed racial bias with a new 10-shade skin tone scale, known as the Monk Skin Tone Scale, to help make their gadgets and apps more inclusive for people of color. This scale was introduced to replace the Fitzpatrick Skin Type, a six-color standard that underrepresented people with darker skin. The Monk scale is now used in various Google products, such as image searches and Google Photos, to ensure a more diverse and accurate representation of skin tones. At the same time, it could be seen as detrimental to focus on skin tones alone in lieu of broader socio-economic factors involved in disparate outcomes across demographics.
The past few years contains a slew of firings when it comes to AI Ethics teams in Big Tech. Timinit Gebru, a co-lead of Google's ethical AI team, was fired after a conflict over the publication of a research paper. Following closely on the heels of Gebru's dismissal, Margaret Mitchell, another leader of the ethical AI team and a co-author of the controversial paper. This trend of AI ethics leads being fired over the past four years. Microsoft, in a notable shift, laid off its entire ethics and society team within its AI division, part of a broader cut affecting 10,000 employees. This team was instrumental in examining the risks associated with integrating OpenAI's technology across Microsoft's products. Similarly, Amazon and Twitter have also reduced their focus on AI ethics teams. Twitter (X, whatever), under Elon Musk's ownership, eliminated its ethical AI team just days before a major round of layoffs. Ironically, in his attempt to train a “non-woke” chatbot, it ended up backfiring rather beautifully, giving balanced and neutral opinions despite the edgy boychild’s intentions.
The AI system (likely not the image generation aspect, but the text fed into the image generation model) is trying so hard NOT TO overrepresent white people that it prefers to rewrite history than to risk this. Consider how Gemini depicts 1943 German Soldiers (the Nazis). The images show people of various races, many of whom were the target of soldiers’ onslaught.
In the next five or so years, what happens when gargantuan AI models emit troubling behaviours like this? Will Google simply pop on the plaster and hope the bleeding stops? Imagine when AI systems comb through a living stream of data in order to make decisions for every aspect of our lives and society. These complicated yet blundering algorithms will be harder and harder to diagnose. They will impact us more severely.
Each time there is a new release, or a model gets used in an untested environment, there could be a new age of saturated information, that we, as technological consumerists, will use to make decisions. The rules of robotics from Asimov’s famous science fiction series demonstrates how even well-intentioned instructions can lead to calamities. Sometimes its the exception to a rule that is more dangerous than the breaching of the rule.
There is demanding work ahead in order to align policy, technology and more importantly, what we as people value. Tech companies shape the laws and governments shape the tech companies. We can’t let bias of this scale into our lives and we can’t tolerate the fake fixes these companies continue to produce, all in the name of a commercial race, driven by a gross medley of figures from lifeless scatterplots and price trajectories. Recent regulation emphasizes the need for bias and fairness auditing as well as explainability in AI models. Read on to learn more about the state of AI regulation in the EU and USA.