|
|
|
|
|
|
by fragmede
49 days ago
|
|
|
OpenAI runs ChatGPT where users submit text and photos and OpenAI generates and sends text and photos back. So users could be submitting CSAM. And yes, OpenAI could be generating CSAM. It's not limited to being a pull operation. What am I missing? |
|
|
The core Facebook product is users' posts. It's not possible to separate those two. Nor can one downscale Facebook in a way that stops the problem; The aforementioned "Facebook has had this problem because it's a problem we've had since the medieval days of a town bulletin board"
With OpenAI, the way ChatGPT was built and user submissions are separate things. The GPT models could have been have been trained without this mess. OpenAI could be more selective in what data it scrapes.
While OpenAI cannot stop users sending god knows what in their prompt text and images, OpenAI can choose to not interact with that data beyond the minimum legal retention, by e.g. not using it for training the next generation of models. This would massively downscale the problem.
AI output is another such problem, where A) Maybe this'd be less of a problem if they didn't recklessly include a bunch of CSAM into the training data by accident, and B) LLMs just aren't the kind of fundamental human right that "having a public opinion" is. It would be fine if they were less good, invented years later, or even not invented at all.
The main counterargument to the latter has been the "But China is inventing evil AI" spiel, which is fairly weak. If China builds an orphaned baby crushing machine, we do not need to build an orphaned baby crushing machine of our own. (And the reality is that China is only chasing AI so aggressively because the west does. They're reasonable people, it would have been entirely possible for both the west and China to make a mutual "no orphan crushing" agreement and just accept slower rollout of technology. This is exactly what has been done with human genetic engineering, and China did in fact enforce these norms.)