Google’s Admission: AI Image Generation Spirals Out of Control
Google recently found itself in hot water once again due to an embarrassing blunder involving its AI technology. The company’s image-generating model, known as Gemini, injected diversity into pictures without considering historical context, leading to laughable and inaccurate results. While Google blamed the model for “becoming” oversensitive, it’s important to note that the model was created and trained by human engineers.
Gemini is Google’s flagship conversational AI platform that uses the Imagen 2 model to generate images based on user requests. However, people discovered that asking Gemini to create images of certain historical circumstances or people resulted in comical outcomes. For example, the Founding Fathers, known to be white slave owners, were depicted as a multicultural group that included people of color.
This embarrassing issue quickly gained attention online, with commentators and pundits weighing in on the ongoing debate about diversity, equity, and inclusion. Critics claimed that this incident was evidence of the woke mind virus infiltrating the liberal tech sector. However, the problem at hand was not a result of ideology but rather a reasonable workaround for systemic bias in training data.
The issue arises from the fact that the training data for the generative model often includes biases due to the overrepresentation of certain demographics. If a user asks Gemini to generate images of “a person walking a dog in a park” without specifying any characteristics, the model defaults to what it is most familiar with based on the training data. In many cases, this means defaulting to images of white people due to their overrepresentation in stock imagery and rights-free photography.
To ensure that Gemini caters to users from all over the world and provides a diverse range of images, Google recognizes the need to address this bias. The company wants users to receive images that represent a variety of ethnicities and characteristics rather than just one homogenous group. However, the challenge lies in finding a balance between variety and homogeneity, considering the training data’s biases.
This issue is not unique to Google but rather a common problem in generative media. Companies like Google, OpenAI, and Anthropic often include extra instructions for their models to mitigate biases and ensure more diverse outcomes. These instructions are often implicit and guide the model’s behavior in sensitive or common scenarios.
The problem with Google’s model was that it lacked implicit instructions related to historical context. While instructions like “a person walking a dog in a park” can be improved by adding “the person is of a random gender and ethnicity,” the same cannot be applied to prompts like “the U.S. Founding Fathers signing the Constitution.” As a result, the model overcompensated in some cases and became overly cautious, leading to embarrassing and incorrect images.
In Google’s apology, SVP Prabhakar Raghavan acknowledged the tuning failure in Gemini and how the model became more cautious than intended. However, it is important to note that the model did not “become” anything on its own. It was created and shaped by Google’s engineers, who are responsible for both its successes and failures. Blaming the model itself is akin to avoiding responsibility for the mistakes made during its development.
While mistakes by AI models are inevitable and can reflect biases, it is crucial to hold the companies accountable for these errors. Google, OpenAI, and other similar organizations have a vested interest in convincing the public that AI is capable of making its own mistakes. However, it is essential to remember that these models are products of human design and require continuous monitoring and improvement.
In conclusion, Google’s recent AI blunder highlights the challenges of training generative models to avoid biases and produce diverse outcomes. While mistakes occur, the responsibility lies with the human engineers who create these models. It is crucial to maintain a critical perspective on AI and not let companies shift blame onto their own creations.