What is the common point between OpenAI's release of text-to-video model Sora, and the nearly simultaneous, hasty rollback of human image generation of Google's Gemini? Both events collide with our perception of truth through artificial intelligence technologies. Both also raise the question of its sources.
With Sora ("sky" in Japanese), generative AI crosses a new frontier, and OpenAI once again sets an unprecedented technological standard. The videos are disturbingly realistic, raising questions about their inspiration. OpenAI remains quiet about the underlying training material, and the company has so far invoked the right to "fair use" of publicly available content to justify their massive collection.
Regardless of the legal outcomes of the disputes that are bound to arise, one question remains: can one rely on the "open" nature of such a large amount of content, to ultimately distill it into models that are so closed?
With its absurd range of racially and gender-diverse German soldiers, Gemini's image generation system revealed, in a quite obvious manner, an over-parameterization whose excesses quickly turned the model into a subject of ridicule. Yet, for the past 25 years, Google has consistently succeeded in the mission it set for itself: to organize information and rank it—especially the information that pays—but usually, the information produced by others.
Clearly released too early, Google's model also fell into the trap of an "ethics of compliance" added on top of the model—and in this case, most probably influenced by woke culture. As per diffusion model principles, its algorithm statistically reconstructs an image, a scene, from a noisy starting point. Much like a sculptor brings forth a piece of art from the original block of stone. And as a work of art, it will never satisfy all cultural canons or claim to represent reality without the bias of the artist.
What does Google's misfortune suggest to us?
First of all, that it is probably by trying to fix supposed biases from its training data that Google so clumsily distorted reality. Secondly, that the race to dominate the technological stage leads even tech giants to release products on a large scale with insufficient testing. Lastly, that technology advances much faster than its legal framework—see Sora—and that significant questions arise regarding the very outcome we are looking for.
By the way, what are we looking for? Is it reality? Generative AI is not built to honestly mirror reality, however large the training dataset is, since truth is not necessarily what is statistically the most probable.
But generative AI perhaps mirrors and amplifies our own uncertainties and our deepest doubts, starting with this fundamental question, which spans across all eras and cultures: what is truth?
