#81 Colourism Between the Lines
Investigating the Racial Bias of Google’s Gemini; Emerging Consensus on Resource Exploitation in Outer Space
Today, Rohan Pai examines the racial bias of Google’s Gemini away from the noise and outrage of social media, while Ashwin Prasad writes about the ostensible consensus on resource exploitation that is emerging between Chinese and American-led space efforts.
Also,
We are hiring! If you are passionate about working on emerging areas of contention at the intersection of technology and international relations, check out the Staff Research Analyst position with Takshashila’s High-Tech Geopolitics programme here.
Cyberpolitik: Investigating the Racial Bias of Google’s Gemini
— Rohan Pai
On the first day of February, Google’s AI chatbot ‘Gemini’, formerly known as Bard, launched its newest image generation feature to all users at no additional cost. Employing Imagen 2, Google’s cutting-edge text-to-image diffusion model, Gemini promised to generate photorealistic, high-definition images of any user input, including human hands, something that often escapes the capabilities of models such as DALL-E. Within just a few weeks of Gemini’s release, however, a barrage of accusations were levelled against the chatbot for possessing a racial bias and generating images rife with historical inaccuracies. Interestingly, and in possibly a first for AI, this bias was discriminatory towards light-skinned, Caucasian people.
When asked to generate an image of the Founding Fathers of the United States, for example, Gemini churned out representations of African-American, Native American and East Asian men. These inaccuracies reached a boiling point when Gemini generated images of multiple ethnic groups in Nazi military uniforms, even going so far as representing the pope as an Indian woman. This engineering snafu was quick to antagonise conservatives in the US, who claimed that Gemini’s racial bias was premeditated as part of Big Tech’s ‘woke agenda’. Elon Musk, as well, offered his two cents and lambasted the forced diversity as ‘insane’ and ‘anti-civilisational’.
Before bringing out the pitchforks, though, it might be valuable to remember that text-to-image generation technology has been around for less than a decade. alignDRAW, developed by the Azerbaijani computer science prodigy Elman Mansimov in 2015 at the University of Toronto, is widely recognised as the first text-to-image model in existence. Mansimov, merely 19 years old at the time, flipped the script on image captioning technology, which used neural networks to label images and instead took on the audacious task of generating images based on text sequences. Albeit blurry and low-resolution, alignDRAW’s images were the pioneers of AI art, representing even non-existent scenarios like “a herd of elephants flying in the blue skies”.
This was soon followed by the application of generative adversarial networks, or GAN, that worked on the principle of training neural networks from a specified dataset to compete against each other and create authentic data. In 2016, researchers were able to use GAN to generate higher-quality images of birds and flowers in different colours. It was only as recently as 2021 that DALL-E was introduced to the public, and never-before-seen images of avocado armchairs and dog-walking radishes began to flood the Internet.
This is all to say that text-to-image generative AI is very much an emerging technology that will continue to make mistakes. Following the pause of Gemini’s image generation feature 3 weeks after its launch, Senior Vice President of Google, Prabhakar Raghavan, expressed a similar sentiment in a blog post. Raghavan candidly admitted that Gemini was designed to generate a range of people belonging to different ethnicities to avoid falling into the trap of overrepresenting a certain skin colour. These intentions are more than justified, given that earlier text-to-image models often have a poor track record when it comes to racial diversity.
Stable Diffusion, a free-to-use deep learning model that was released about a year after DALL-E, received its fair share of public outcry for generating images of people who conformed to regressive racial stereotypes. People with light skin tones routinely represented Doctors, lawyers, CEOs and other high-paying jobs, while blue-collar professions were overrepresented by people of colour. More dangerously, though, images of convicts and terrorists tended to display African-American and Middle Eastern men, respectively. In retrospect, it was perhaps prudent of Google to train Gemini on data that represented a range of ethnicities. After all, this is by no means the tech giant’s first rodeo in the arena of racist algorithms. Less than a month after its launch in 2015, Google Photos erroneously organised images of 2 black people into an album titled “Gorillas”.
The Gemini controversy, at the end of the day, is not an isolated case. Large language models or LLMs are trained on datasets fed to them by human beings, and these datasets cannot promise to always be up-to-date. In the case of Stable Diffusion, certain occupations are associated with certain skin tones because that is what the historical data unfortunately reveals. Of course, those discriminated against in the past have since risen through the ranks to occupy even the highest positions of power, but applying this logic to every facet of social life no doubt results in statistical bias. The question that now looms over the space of generative AI is whether to accurately reflect the existing prejudices in society based on training data or to artificially promote inclusion and diversity as per the prevailing value judgment at the time.
Providing concrete answers to such a divisive question is nothing short of a herculean task, but new legislation may keep the public at bay. On March 13th, the landmark EU Artificial Intelligence Act was passed in the European Parliament with a staggering 523-46 majority. Among other revolutionary regulations, the EU AI Act requires generative AI companies to submit detailed summaries of the datasets used to train their large language models. At the moment, text-to-image models such as DALL-E and Gemini are not built on open-source software, and the recent legislation in the EU could potentially change that. The actual enforcement of the EU AI Act is set to take place only in 2025, before which companies may follow the rules at their discretion, but the future of generative AI looks transparent.
India is set to host the Quad Leaders' Summit in 2024. Subscribe to Takshashila's Quad Bulletin, a fortnightly newsletter that tracks the Quad's activities through the Indo-Pacific.
Your weekly dose of All Things China, with an upcoming particular focus on Chinese discourses on defence, foreign policy, tech, and India, awaits you in the Eye on China newsletter!
The Takshashila Geospatial Bulletin is a monthly dispatch of Geospatial insights for India’s strategic affairs. Subscribe now!
Antariksh Matters: Emerging Consensus on Resource Exploitation in Outer Space
— Ashwin Prasad
The Outer Space Treaty and Moon Treaty fall short in their efforts to address the issue of space resource mining. China has now disclosed its stance on the matter. It broadly agrees with the United States as outlined in the Artemis Accords - something India has also signed. This emerging consensus could pave the way for a more comprehensive legal framework on the subject.
Race to the Moon
The true test of the Moon Treaty, both as a treaty and customary law, will not come until the exploitation of extraterrestrial resources becomes technically and economically feasible. ~Listner, 2011
The costliest part of a Lunar mission was getting to the moon. With the advent of reusable launch vehicles, this cost will only fall. The space race is on, and the first destination is the Moon. The moon is the jumping-off point for all space missions farther into the solar system. Its location at the edge of Earth's gravity well, coupled with its low gravity, makes it ideal to house spaceports to deep space. Also, the moon is a rich source of mineral wealth. Lunar ice can be a potential fuel source in addition to providing potable water and breathable air. The Moon will be an important testbed for space technologies. It will enable the learning required to overcome the effects of solar radiation and meteor showers as humanity sets its sights on Mars and beyond.
The US is leading the charge. NASA's Artemis Campaign has already begun. In 2026, Artemis III will land humans on the moon. The In-Situ Resource Utilisation is a major component of the Artemis missions. The Russians and the Chinese also have plans to carry out manned moon missions by the decade. The two countries have partnered on the International Lunar Research Station (ILRS) project to set up a permanent lunar base in the 2030s. Pakistan, Azerbaijan and Belarus have already joined the ILRS. India is conceptualising the roadmap to put Indians on the moon by 2040.
Given the growing motivations and capability to reach the moon, nations should consider ensuring peaceful coexistence in space alongside strategic interests. Setting up a clear, legal framework for space resource utilisation can preserve peace and, at best, prevent or at least reduce potential military conflicts between nations in space in the coming decades.
Where the Space Treaties Fall Short
To appreciate the need for the framework, one must understand where the existing international space treaties fell short. Against the backdrop of the Cold War and growing concerns about the militarisation of outer space, the Outer Space Treaty entered into force. It affirmed that space should be used for peaceful purposes only. It also prohibited nations from asserting ownership over outer space or any celestial body. The fact that over a hundred nations have ratified the treaty shows that there is a global desire to preserve peace in outer space. While the treaty deserved credit for its relevance almost 60 years later, it failed to address the issues related to the ownership of space resources, especially by private entities.
In 1984, the Moon Treaty came into force in an attempt to take the Outer Space Treaty's ideas forward. It declared the Moon and its resources as the common heritage of mankind. Declaring something as a common heritage can have many unintended negative consequences due to the tragedy of the commons. It will disincentivise space innovation and will most likely impair the space environment. Only a handful of States have ratified the treaty. Notably, the US, India, Russia and China are not among them.
With the entry and participation of many private corporations in the space sector in recent years and the aforementioned race to the moon motivated in part by the Moon's rich mineral resources, there is a growing need to plug these gaps in Outer Space and Moon Treaties and reach a global consensus surrounding space resource utilisation.
Emerging Consensus
The Artemis Accords attempt to address the uncertainties left by the Outer Space and Moon treaties. The Accords are a series of non-binding bilateral agreements that attempt to steer the Outer Space Treaty's principles toward the US-led interpretation. The accords do this by distinguishing between national appropriation of space territory and national exploitation of space resources. India is also a signatory.
There are now reports of consensus coming from another major player. According to a recent document submitted by a Chinese delegation to the United Nations Committee on Peaceful Use of Outer Space (UNCOPUOS), China considers the user of resources in outer space as legal and calls for adherence to the principles of the Outer Space Treaty. Christopher Johnson, an expert on space law, highlighted the importance of this move. It indicates that China is taking UNCOPUOS seriously. It also means that the signatories of the Artemis Accords and China bloc agree on the issue of space resource utilisation, which can pave the way for broader consensus going forward.
Admissions are now open for Takshashila’s Postgraduate Programme in Public Policy. This comprehensive, multidisciplinary, and entirely online 48-week program is tailored for those aiming to delve deep into both the theoretical and practical aspects of public policy. Apply by 22nd April to avail yourself of the Early Bird Deadline (10% scholarship)!
What We're Reading (or Listening to)
[Podcast (All Things Policy)] Making Sense of “New” Chinese Chipmaking Breakthroughs, ft. Rijesh Panicker and Satya Sahu
[Podcast (The Hindu)] What can the 10,000 genome project tell us about India’s health, ft. Shambhavi Naik and Zubeda Hamid
[Opinion] India cannot solve its water crisis without proper pricing, by Nitin Pai