The Topic That Keeps Getting Left Out of AI Regulation Debates
Whenever I look at regulation discourse, it’s always only about safety and copyright, but the most urgent thing on the ground seems to be data provenance management and where responsibility lies.
When an incident occurs while no one knows what the model was trained on, there’s no standard for deciding whether it’s the developer’s responsibility or the user’s. Whether we regulate or not, I think this part needs to be sorted out first.
Of course, I’m not an expert either... What does everyone think?
7 answers
Data provenance management—I really relate to this. Every time regulations come up, I don't get why this is the one thing that always gets left out.
I've worked on the data labeling side, and honestly, just tracking training data provenance alone can eat up half a year of a company's time. The vast majority of places use data they've crawled and scraped together without even logs of where it came from... I think even before regulation, we need to start by establishing the practice of tagging sources internally. But this isn't a cost-free effort, so the reality is that startups can't even begin to take it on.
Well, that’s a bit... It’s true we need to sort out where responsibility lies, but if doing that pushes the safety discussion to the back burner, isn’t that the bigger problem? It’s not a matter of order—they need to happen at the same time. And saying 'there are no criteria for judgment' also seems like a bit of an exaggeration; isn’t this already covered to some extent by the Personal Information Protection Act or the Information and Communications Network Act? You need to first clarify whether you’re proposing to create something new or to apply the existing laws.
Got a source? Where did you see this claim?
I get the point of the original post, but is provenance management really the most urgent issue on the ground? I think responding to personal data breaches or deepfake harms is more urgent. Victims are already emerging in those areas, while provenance management is still only a potential risk.