cross-posted from: https://lemdro.id/post/11955
TL;DR
- Google has updated its privacy policy.
- The new policy adds that Google can use publically available data to train its AI products.
- The way the policy is worded, it sounds as if the company is reserving the right to harvest and use data posted anywhere on the web.
You probably didn’t notice, but Google quietly updated its privacy policy over the weekend. While the wording of the policy is only slightly different from before, the change is enough to be concerning.
As discovered by Gizmodo, Google has updated its privacy policy. While there’s nothing particularly notable in most of the policy, one section now sticks out — the research and development section. That section explains how Google can use your information and now reads as:
Google uses information to improve our services and to develop new products, features and technologies that benefit our users and the public. For example, we use publicly available information to help train Google’s AI models and build products and features like Google Translate, Bard, and Cloud AI capabilities.
Before the update, this section mentioned “for language models” instead of “AI models.” It also only mentioned Google Translate, where it now adds Bard and Cloud AI.
As the outlet points out, this is a peculiar clause for a company to add. The reason why it’s peculiar is that the way it’s worded makes it sound as if the tech giant reserves the right to harvest and use data from any part of the public internet. Usually, a policy such as this only discusses how the company will use data posted on its own services.
While most people likely realize that whatever they put online will be publicly available, this development opens up a new twist — use. It’s not just about others being able to see what you write online, but also about how that data will be used.
Bard, ChatGPT, Bing Chat, and other AI models that provide real-time information work by scraping information from the internet. The sourced information can often come from others’ intellectual property. Right now, there are lawsuits accusing these AI tools of theft, and there are likely to be more to come down the line.
Sounds like a good way to speed up sample collapse. Eventually AI training data will include enough AI written comments and articles that the training data or “sample” will be contaminated by it resulting in training data that makes the AI worse, kind of like inbreeding in humans expressing recessive genetic disorders like diseases and birth defects. Once that happens they’ll need to find a way to accurately detect and remove AI data from the sample to continue improving the AI. But even with the early generations of AI we’re using now it’s incredibly hard to automate that with any kind of real accuracy and attempts have caught a lot of false positives and let a lot of legitimate AI generated texts through.
It’s interesting that the limiting factor on AI development may be the fact that it was released to the public before further training could be done.
A big hurdle of AI is the fact that they really can’t ‘learn’, at least not like humans can, where we filter out bad data or go back and correct previous assumptions (not that we do this perfectly). Seems like anyone who’s able to truly figure out how to teach AI without needing super-clean data sets will have basically unlocked something pretty close to the singularity. Which makes me assume that we’re honestly no where close to figuring that out and that sample collapse is much more likely (with possibly the internet as a whole being effectively ruined, same as voice calls have been effectively ruined by rampant spam).