A centaur differs from human-in-the-loop by symbiotic learning: human input enters the model's learning as a partner dataset, not as labels.

groundsEffective Generative AI: The Human-Algorithm Centaur, Soroush Saghafian, Lihi Idan

Symbiotic learning departs from the methods discussed in Section 4 in that it implements a mutual human-AI learning process where the human is not treated merely as an oracle or a preprocessor, but rather as an equal partner in the learning process by incorporating her unique input into the learning process.

groundsEffective Generative AI: The Human-Algorithm Centaur, Soroush Saghafian, Lihi Idan

Similar to the fact that two businesses might either create a partnership or merge, humans and algorithms can either work alongside each other through a simple partnership (while being two separate entities that may or may not listen to one another) as suggested by HITP-type methods, or combine their powers through a strong merge, as suggested by centaurs.