They could still use AI for that step of matching backing with warrant?
Would this work if you prompted the same prompt but gave a mix of images that were and were not truly humiliating rituals? What if the images were almost completely either not humiliating or not rituals? What if the images were culturally contextually not humiliating but perhaps a Western view would consider them humiliating, or vice versa?
Ooooh this is a great question, and would make a great mini-research project. You could specify that you want the evidence interpreted in a specific cultural context and it would try to do that, but I don't know how successful it would be. Additionally the fact that a non-western context requires specification and a western context does not is itself a sticking point. If you find a real-world parallel example we could use let me know.
On the other question -- yes, I'm working on some tests around how to make sure that the analysis is not *too* influenced by what it expects to see. All analysis benefits from some expectation (inference is key) but seeing how much expectation might deform an analysis is something I am working on.
They could still use AI for that step of matching backing with warrant?
Would this work if you prompted the same prompt but gave a mix of images that were and were not truly humiliating rituals? What if the images were almost completely either not humiliating or not rituals? What if the images were culturally contextually not humiliating but perhaps a Western view would consider them humiliating, or vice versa?
Ooooh this is a great question, and would make a great mini-research project. You could specify that you want the evidence interpreted in a specific cultural context and it would try to do that, but I don't know how successful it would be. Additionally the fact that a non-western context requires specification and a western context does not is itself a sticking point. If you find a real-world parallel example we could use let me know.
On the other question -- yes, I'm working on some tests around how to make sure that the analysis is not *too* influenced by what it expects to see. All analysis benefits from some expectation (inference is key) but seeing how much expectation might deform an analysis is something I am working on.
I think at the very least, try putting all neutral (not humiliating) images, or a mix? I'll think about particular cultural ones