I used to think that the power requirements might be the downfall of AI. Now I think the lawyers may kill it.
I'm peripherally involved in getting AI approved at our company. It's almost pernicious. Aside from the obvious things, like what are approved or disallowed AI sites for the business, we have to deal with various vendors bundling varying types of AI into their products. The trend seems to be to turn the damn things on by default and then hope we never figure out the legal implications that entails.
We're working on ongoing policy development and locking down some of these things, but it feels like a losing battle, and these vendors aren't making it any easier. Add to that the various flavors of AI have different concerns and vulnerability vectors. Something like ChatGPT might potentially have access to strategic or proprietary information that an employee submits, while something like Claude might be able to infer a business direction from coding requests, or even worse, train the AI to provide a solution that we may have spent months perfecting in-house with the potential to hand it to our competitors with the right prompt.
None of this even considers the unanswered questions about the potential copyright concerns building these models in the first place.
For the supposedly smart people building these things, they sure haven't thought a lot of it through. Or maybe they have and just hope to be so integrated before legal catches up that they can't be removed.
Curious what others are seeing in this area.
Here's what Grok says about potential code leakage through Claude:
"you paste (or describe in detail) a novel algorithm, architecture, optimization, or proprietary implementation into a regular Claude chat:
The model can learn patterns from it during training. Future users (including competitors asking similar questions) could receive outputs that are influenced by or closely resemble your solution. LLMs don't usually spit out exact copies of training data, but they can generalize and reproduce distinctive approaches, especially if the idea is rare or highly specific. This is a known industry-wide issue (memorization, regurgitation, and IP leakage risks).
It's not that your code is "publicly exposed" in a searchable database—it's internalized into the model's weights. But the net effect for competitive advantage can be the same."
I wonder how many companies rubbing their hands together about replacing employees with AI have considered the potential intellectual and strategic leakage if they don't affirmatively opt out of training (and if the vendors don't change the service agreement later and start using their info anyway)
so wait, does this mean I can steer public AI's into sugguesting my own ideas to other people who interact with them later?
Cuz i've got some heaters, let me tell ya
I think there's a very real possibility that you could use one AI to corrupt another AI if you can get your prompts included in the training process.