People Are Using AI to Generate Disturbing Kids’ Bedtime Stories

Advanced AI models like GPT-3 and DALL-E have been all the rage lately, with the latter inspiring a recent surge of interest in creating AI-generated images from text prompts.

But it turns out, combining those images with advanced text prediction can result in some extremely elaborate and disturbing stories—including an illustrated children’s book about a knife-wielding Pikachu leading his comrades into battle under ominous dark skies.

That’s just one result achieved by Edwin Chen, an engineer at Surge AI who used a combination of OpenAI’s DALL-E and GPT-3 natural language tools to generate a Pokémon-themed story for his kindergartener. GPT-3 is known for its hauntingly accurate ability to fill in the blanks when given only a short text prompt, and OpenAI’s Davinci builds off the model to make it especially good at understanding intent and context. When given a brief description of a story about Ash and Pikachu going to war against Team Rocket, the model generated a detailed story about the ensuing battle. Each generated line was then used as a prompt for DALL-E, resulting in some fairly unsettling illustrations.

DALL-E is still in testing phases and only available to a select list of artists and researchers. But OpenAI’s text-based generation tools are open to anyone who registers for a free account on their website.

After Chen’s illustrated Pokémon tale began making the rounds on social media, others were inspired to generate stories of their own. Here is one by Motherboard’s own Edward Ongweso, about Donald Trump’s love of the anime series Attack On Titan:

View on X

For many AI researchers and ethicists, the thought of using machine learning to generate content for kids will be particularly alarming. GPT-3 has been shown to reinforce sexist and racist stereotypes, and the researchers behind DALL-E have demonstrated that the image-generating system will also frequently produce biased results—such as the prompt “CEO” returning only images of white men in business suits.

In response, companies and organizations like OpenAI have developed what they claim are mitigations that prevent abuse of generative AI systems. This can involve various methods, like re-weighting the training datasets to increase the frequency of image types the system normally wouldn’t display.

No matter how fascinating or cute the results, AI researchers will have to contend with the fact that these systems will always produce skewed and harmful content that simply mirrors pre-existing societal inequality.