Chapter 1: The Concept of AI Self-Portraits

Recently, a couple of notable incidents in the world of artificial intelligence have sparked interest. In June, Blake Lemoine, a now-former Google engineer, claimed that a program called LaMDA had achieved sentience, leading to his dismissal in July. In another development, OpenAI's DALL-E 2, an AI visual design tool that generates images on request, entered public beta.

While I won't delve into the debate over whether AI can truly possess sentience or creativity, I aim to merge these concepts and explore how an AI would visually depict itself. To investigate this, I assigned an AI program a task reminiscent of an eighth-grade art project: to create a self-portrait. Since DALL-E 2 is currently invite-only, I opted for an AI web application known as Craiyon, previously branded as DALL-E mini. Craiyon's team includes members who contributed to OpenAI's DALL-E, and the platform is free to use, supported by ads.

The outcome of my request for a self-portrait from the AI was intriguing.

Chapter 2: Analyzing AI's Artistic Output

After running the same test a few days later, I noted that the AI's self-portraits remained largely unchanged.

The faces in both sets of portraits displayed distortions. Craiyon acknowledges this as a limitation, stating in its FAQ that the strangeness of faces arises from the current image encoding technology, which is a work in progress. Notably, the AI consistently produced self-portraits that depicted a European male with facial hair. Does this imply the AI identifies as a bearded white man? Likely not. The uniformity in its output indicates that the training data, rather than a defined self-image, is the primary influence. The AI's results are shaped by biases inherent in the dataset it was trained on, where most images labeled "self-portrait" depict white males.

The developers note in their model card that the AI tends to generate images of individuals perceived as white, while people of color are underrepresented. They aim to enhance this aspect in the future by ensuring a more balanced and equitable representation in the training dataset.

Despite the noted facial distortions, the AI can produce diverse representations when prompted with specific instructions, such as "self-portrait oil painting of a multiracial woman."

While these outputs do not genuinely reflect a self-identity, they illustrate the AI’s ability to create images akin to what it has been trained to recognize as a "self-portrait." An AI with a sense of self might choose to anthropomorphize that identity, akin to how humans create self-portraits.

Chapter 3: The Implications of AI Identity

The philosophical inquiries surrounding AI identity may seem abstract, but they hold real implications in product design today. From a development standpoint, the prevailing approach is to leverage human intelligence over artificial. By allowing users to choose the identity traits of the AI they engage with—like in Apple’s Siri—product managers and developers shift the responsibility of identity onto the users.

As technology evolves, the aspects of design that depend on identity—such as voice, personality, and visual representation—may transition from user-defined choices to algorithmically optimized selections based on input data, reflecting the AI's "experiences." This shift would change the nature of choices from intentional to latent, as AI systems respond to the data generated from human interactions.

Ultimately, how AI represents itself to users is shaped by the values and perceptions we hold regarding identity traits, informed by the data we provide.