Uncanny AI

OpenAI's ChatGPT and Codex were referring a lot to goblins, trolls, and raccoons earlier this spring. And that happened inadvertently when the AI lab prompted those models to have a "nerdy" personality, says Janelle Shane, writer of the blog "AI Weirdness."

That sentence structure is what once made William Shakespeare's writing great. But, chatbots tend to style outputted words as "it's not X, it's Y" to sound nuanced, says Will Oremus, writer at The Atlantic.

Because the advanced chatbots are trained on global data, it may choose to spew a word relating to a conversation in a different language. This could mean the large language model decided to cue up the next relevant token erroneously, says Janelle Shane, who writes the AI Weirdness blog.

When interacting with AI chatbots, the large language model might reference something the user mentioned without relevance to the conversation. Janelle Shane, creator of the AI Weirdness blog, explains why LLMs' “memory” is different from human memory.