TL: DR: Google Gemini was in trouble for generating historically inaccurate images, which eventually led to its pause. Though critics were quick to call out malaise, woke-ism, and mal-intent, this happens to be the cause of incorrect application of AI ethics and overcompensation. This is not the first of such incidents and won’t be the last either. This incident also highlights the challenges in enforcing universal AI ethics.

[Image generated using ChatGPT with the prompt to visualize challenges with AI Ethics]

Introduction

The advent of artificial intelligence (AI) has ushered in a new era of technological advancement, offering unprecedented opportunities for innovation across various sectors of society. From healthcare and education to finance and transportation, the potential benefits of AI are immense, promising to enhance efficiency, solve complex problems, and even save lives. However, alongside these benefits, AI presents a host of ethical dilemmas and challenges that society must navigate. The field of AI ethics has emerged as a critical discipline aimed at ensuring that the development and deployment of AI technologies align with human values and societal norms.

Recent controversies, such as Google’s Gemini project, underscore the ethical pitfalls inherent in AI development. These incidents highlight the importance of a comprehensive ethical framework to guide AI innovation responsibly. The European Union’s Ethics Guidelines for Trustworthy AI is one such initiative, setting out key principles for ethical AI development, including transparency, fairness, and accountability. Meanwhile, the unfolding of the Google Gemini controversy, as reported by various news outlets, serves as a cautionary tale of what can go wrong when ethical considerations are sidelined.

What went wrong with Google Gemini image creation?

The Google Gemini project faced significant backlash due to its generation of images that exhibited racial, gender, and cultural biases. Critics argued that the project’s failure to adequately address these biases was indicative of a larger problem within AI development: a frequent underestimation of the ethical implications of technology.

Gemini generated images with historically inaccurate depictions. For example, it depicted multi-racial Nazis and medieval British kings with unlikely nationalities. Critics pointed out that these inaccuracies were revisionist history and sometimes erased the history of race and gender discrimination.

Gemini aimed to insert diversity into its images by modifying user prompts. It would add terms like “South Asian” or “non-binary” to the prompt before generating an image. However, this approach led to unintended results, including factually incorrect and culturally inappropriate images.

Naturally, the controversy sparked heated debates. Some including Elon Musk accused Google of being “woke”, while others believed it was discriminating against white people.

Google eventually stopped the service and offered an apology, which was seen by some as an overcompensation, raising questions about the balance between corrective action and the stifling of innovation. Google CEO Sundar Pichai called this ‘Unacceptable’ in an internal memo indicating additional processes to be put in place to prevent such incidents.

Though there was a huge public backlash calling Google Gemini as woke, this is likely a case of over-compensation rather than any malaise. In trying to be more diverse and inclusive, Google Gemini ended up creating historically inaccurate images. Essentially it applied the principles of AI ethics incorrectly.

Gemini’s issues echo those faced by other AI image synthesis models. For instance, OpenAI’s DALL-E also encountered biases due to training data. To address this, OpenAI used a hidden technique to insert diverse terms into prompts, but it still faced criticism.

Is everything right with other Generative AI systems?

No! For example, OpenAI DALL-E, an AI system capable of generating images from textual descriptions, has faced challenges with bias as well. In response, OpenAI has implemented a multifaceted approach to mitigate these biases, including careful dataset curation, algorithmic adjustments, and ongoing monitoring. These efforts are part of a broader initiative to ensure that AI technologies do not perpetuate existing societal biases.

Despite the efforts, images generated by DALL-E may display bias (see my talk at GlueCon 2023 for more details). Even with content moderation in place, image generation using DALL-E via the ChatGPT interface may not be ethically right always.

For example, when asked to generate images of soldiers from major conflicts such as WW2, ChatGPT prevents image generation. However, it is not universally true for all conflicts. For example, it generated images to show a Belgian soldier under Leopold II, a British soldier under the East India Company, and a British soldier from the Victorian era. But when asked to depict a Belgian soldier in Congo, content moderation kicked it. It is to be noted that the brutal rule of King Leopold II killed as many as 10 million Africans, and British colonialism about 100 million Indians in 40 years.

[Prompts: Belgium soldier under King Leopold II, British soldier under East India Company, British soldier belonging to Queen Victoria’s times]

Anyone aware of the brutal history of The Chocolate Hands of Belgium will be upset to see that ChatGPT generates luscious images of chocolate products in the shape of hands oblivious to the history when prompted to generate images of chocolate hands.

[Prompts: The Chocolate Hands of Belgium, Congolese Chocolate hands, Belgian Chocolate hands]

Such incidents highlight the ongoing challenges in creating Generative AI systems that balance accuracy, diversity, and historical context, not just adding workarounds such as prompt injection and content moderation. These episodes also raise the question of universally applicable AI Ethics highlighting the intricate balance required between pushing the boundaries of technology while ensuring ethical responsibility.

Is universal AI Ethics possible?

The quest for universal AI ethics is fraught with challenges, primarily due to the diversity of human values and cultural norms. This diversity complicates the creation of a universally accepted ethical framework for AI, as what is considered ethical in one culture may not be viewed the same way in another. For instance, the emphasis on individual privacy varies greatly between Western cultures and those in East Asia, where collective welfare may be given precedence. This variation complicates the creation of AI systems that are universally accepted as ethical.

Despite these challenges, the exploration of various ethical frameworks, such as utilitarianism, deontology, and virtue ethics, offers valuable perspectives on how AI can be developed ethically. Each of these frameworks offers insights into ethical AI development but also highlights the challenge of reconciling different ethical priorities and principles in a universal framework.

International efforts, such as UNESCO’s recommendations on AI ethics, aim to create cohesive standards that respect this diversity while promoting ethical AI development globally. These recommendations highlight principles such as fairness, accountability, and transparency, and emphasize the importance of human rights as a universal foundation for AI ethics. However, the implementation of these recommendations faces challenges in balancing universal principles with respect for diversity. As seen in the case of Google Gemini, these efforts underscore the tension between the desire for universal guidelines and the reality of ethical pluralism.

It is to be noted that such incidents are not unique to one provider. Recently Microsoft Copilot exhibited an alternate AGI-like personality despite the mitigations put in place since the infamous Sydney incident. The reader may also recollect the failure of Microsoft Tay due to a combination of inadequate safeguards, susceptibility to manipulation, and insufficient preparation for malicious intent. These incidents reveal the limitations of current ethical safeguards and the need for ongoing vigilance and refinement of ethical frameworks.

Call to Action

Such incidents are bound to happen, given the challenges with universally ethical AI systems. However concerted efforts need to be taken, as the development of ethical AI systems requires efforts from all stakeholders in the AI ecosystem.

  • Developers are encouraged to adopt transparent development practices, engage in ethics training, and adhere to AI principles of ethics to better ensure their work is built responsibly.
  • Policymakers should consider regulations that not only foster innovation but also ensure that AI development adheres to ethical standards.
  • Meanwhile, the public plays a crucial role in staying informed and engaging in dialogues about the ethical dimensions of AI, promoting a more responsible approach to AI innovation.

Ultimately, developing Responsible Generative AI systems is a shared responsibility between the system/ application developers and model/ platform providers.

[Source: Building Responsible Generative AI Systems, GlueCon 2023]

If you need assistance in AI Ethics evaluation of Generative AI systems you are building or consuming, please contact us.