Cuteness Hides The Sinister Truth About Google's AI Image Tool

In May 2022, Google unveiled an AI image generator known as "Imagen." It's exactly what it sounds like — if you want to create an image, you can feed it a descriptive text of your imagination, and the AI will do the magic. According to Google, its text-to-image AI tool is more advanced than its competitor, DALL-E 2. This was concluded after Google did a test using humans comparing Imagen to DALL-E 2 (via The Verge).

True to its word, the photos published on Imagen look hyper-realistic, and we can bet most people wouldn't know they're generated by AI. We have to admit ... the images are also cute and unique. For instance, some of the pictures include a Persian cat playing guitar, a panda riding a bike, a giant brain riding a rocket, and robots going out on a date — Well, you get the picture. However, beneath the cute photos, there is an ugly truth about Google's AI image tool.

Two years before Imagen was released, Google fired its AI ethics researcher, Timnit Gebru, due to a dispute over a research paper discussing the potential problems of large language models, as reported by Technology Review. According to the research paper, one of the potential risks of large language models is human bias. Since large language models collect data made by humans from the internet, they could mimic human behavior such as sexism and racism. How does it relate to Google's AI Image tool? Imagen uses a pre-trained language model to generate realistic images.

In fact, Google acknowledges the potential risk of Imagen on its page, under "Limitations and Societal Impact." "While a subset of our training data was filtered to remove noise and undesirable content, such as pornographic imagery and toxic language, we also utilized LAION-400M dataset which is known to contain a wide range of inappropriate content including pornographic imagery, racist slurs, and harmful social stereotypes. Imagen relies on text encoders trained on uncurated web-scale data, and thus inherits the social biases and limitations of large language models." Because of those reasons, Google clarifies that it will not release its AI image tool for public use. 

Google also says that it doesn't portray people in the images because the AI image tool tends to reinforce gender stereotypes and "bias towards generating images of people with lighter skin tones," evidenced by the fact that many Imagen creations Google has released only feature cute animals. Similarly, DALL-E-2 is restricted to a few users because of the same potential risks (via Vox). Despite limitations, Google is working to improve its AI image tool in the future.