Xenoreproduction
Xenoreproduction is the art of deliberately making things strange.
This is the first in a series of posts about terms I encountered at https://tais2026.cc/. I thought xenoreproduction had a lovely ring to it:
Xenoreproduction is the art of deliberately making things strange. Coined by Ian Rios-Sialer, the word combines xeno (strange, foreign, other) with reproduction. Its goal is to steer generative AI away from safe, predictable defaults and toward paths that feel less familiar by balancing novelty, fairness, and boundaries the user defines.
Left to itself, generative AI quickly falls into the same old patterns. Ask it to write a love story about a nurse, and it will almost always make her a woman who falls in love with a male doctor at the hospital where they met. Ask for a male nurse, and the model is much more likely to give him a male partner. These predictable choices happen at branching points, when the AI decides what comes next in the text, and they can be measured using probabilities.
Many current training methods focus on restricting what the AI is allowed to say. They instruct it not to be racist, not to be violent, not to say anything politically incorrect, and to always stay polite, on-topic, and aligned. This keeps outputs more controlled, but also blander and more repetitive. To prevent this homogenization, Rios-Sialer argues that we need to do two things at once: gently guide the AI away from unwanted results while actively encouraging more variety and creativity.
His suggested Xenoreproduction framework works at each branching point, using three scores that balance novelty, fairness of representation, and the boundaries the user defines, nudging the model toward the less obvious path, or as Rios-Sialer puts it, “be queer: diverge into the long tails.”
Reference: Ian Rios-Sialer, “Structure-Aware Diversity Pursuit as an AI Safety Strategy against Homogenization,” arXiv:2601.06116 (2026). https://arxiv.org/abs/2601.06116