For example, DALL-E v3 appears to generate images and then evaluate the generated images before rendering to the user. This approach is essentially adversarial, whereby the evaluative engine can work at cross-purposes to the generative engine.

Its this layered, adversarial approach, that makes the most sense; and there is a very strong argument for a robust, open sourced evaluative AI anyone can deploy to protect themselves and their systems. It is a model not dissimilar from retail anti-virus and malware solutions.

In sum, I would like to see generative AI well funded, limited distribution and regulated; and evaluative AI free and open. Hopefully, policy makers see this the same way.