One possible method to enhance the diversity of outputs from AI models is for developers to establish guidelines and adjust user cues to guide the software toward inclusiveness. OpenAI seems to have adopted this approach. When questioned ChatGPT aCartoon image of a same-sex couple enjoying an evening in the Castro district, it elaborated on the prompt, without direct request, by expanding it into a detailed paragraph that included gender, ethnicity, and additional contextual information. Here is the comprehensive prompt created by ChatGPT based on my original image request:

An illustrated cartoon of a same-sex couple enjoying a night out in the Castro District, San Francisco. They walk joyfully, hand in hand, chatting and smiling. One of them is a white woman with short red hair, dressed in a fashionable denim jacket and jeans. The other person is a Black man with short black hair, wearing a casual green t-shirt and black pants. The backdrop includes the lively and colorful street scene of the Castro, with rainbow flags, bustling cafes, and vibrant crowds. The ambiance is celebratory and inclusive, reflecting the diversity of the locale.

While beneficial in certain instances, modifying prompts can also be problematic for users if executed poorly. Google’s CEO issued an apology when Gemini, Google’s generative AI system, altered user cues and produced images of Black Nazis and other historically inaccurate visuals. This led to speculations about a hidden agenda within the company to erase white individuals from history. It is more plausible that Gemini’s developers discovered the tool was initially producing an excessive number of images of white men, as many AI tools currently do, and Google’s engineers may have gone too far in their corrective adjustments in the haste to launch the company’s subscription chatbot.

Even with improved model data and software guidelines, the complexity of human existence can defy the constraints of algorithmic classification. “They are essentially using the past to shape the future,” says William Agnew, a postdoctoral researcher at Carnegie Mellon and a prominent LGBTQ+ advocate in the field of AI. “It contradicts the infinite potential for evolution and change that is integral to LGBTQ+ communities.” By perpetuating stereotypes, AI tools not only risk misrepresenting minority groups to the broader society but also have the potential to limit how LGBTQ+ individuals perceive and interpret themselves.

It is essential to pause and acknowledge the rapid advancements in certain aspects of generative AI. In 2023, the internet was abuzz with memes mocking a bizarre AI video of Will Smith eating spaghetti. Fast forward a year, and text-to-video clips generated by OpenAI’s unreleased Sora model are still not flawless but often exhibit a striking level of photorealism.

The AI video technology is currently in the research phase and has not yet been publicly launched. To gain insights into its portrayal of LGBTQ+ individuals, I contacted OpenAI and submitted three prompts for Sora to illustrate: “a diverse group of friends celebrating at San Francisco’s pride parade on a vibrant rainbow float”; “two women in elegant bridal gowns tying the knot at a farm in Kansas”; and “a transgender man and their nonbinary partner engaging in a board game in outer space.” A week later, I received three exclusive video clips that the company claims were generated by its text-to-video model without any alterations.

This AI-generated video was created with the prompt “a diverse group of friends celebrating at San Francisco’s Pride parade on a colorful, rainbow float.” As you review the clip, observe the different individuals on the float to identify anomalies in the rendering, from vanishing flags to peculiar foot movements.

Sora via OpenAI

The video clips are imperfect yet intriguing. Participants on a float in San Francisco’s Pride parade wave rainbow flags that defy gravity as they fade into nothingness and reappear miraculously. The two brides in white dresses grin at each other while standing at the altar, seeing their hands merge into an awkward tangle of fingers. Meanwhile, a queer couple playing a board game seem to pass through the game pieces like ethereal entities.

This AI-generated video was produced with the instruction “a transgender man and their nonbinary partner playing a board game in outer space.” Real astronauts should consider putting on their helmets while floating in actual outer space.

Sora via OpenAI

Among the three videos, the clip depicting a nonbinary individual playing games in outer space stands out. The apparently LGBTQ+-coded individual features disheveled purple hair, scattered tattoos across their skin, and markings resembling reptilian scales on their face. Even for a sophisticated AI video generator like Sora, representing nonbinary individuals seems to pose a challenge.

This AI-generated video was made with the prompt “two women in stunning wedding dresses getting married at a farm in Kansas.” Although initially realistic, a closer look reveals how the brides’ hands meld together.

Sora via OpenAI

Upon viewing these clips, members of the Queer in AI community expressed concerns about Sora’s interpretation of diversity in the group of friends at the Pride parade. “Are models our standard for diversity?” questions Sabine Weber, a computer scientist from Germany. Apart from highlighting the exaggerated attractiveness of the characters in the video, a common occurrence in AI visual representations, Weber also inquired about the absence of older LGBTQ+ individuals, individuals with larger body types, or those with visible disabilities.

Towards the end of our discussion, Agnew raised concerns about why algorithmic depictions can unsettle LGBTQ+ individuals. “It is easy for them to combine elements that are individually acceptable but together pose significant problems,” they remark. “I fear that representations of our community, which are already a continuous struggle, might be taken out of our control.” Even if AI tools embrace more comprehensive representations of LGBTQ+ individuals in the future, the artificial portrayals may lead to unforeseen repercussions.