> So, kids, let's not be asking any AI chatbot whether you should divorce your husband, how to cheat on your taxes, or if you should try to get your boss fired. That information will be kept, it may be revealed in a security breach, and, if so, it will come back to bite you in the buns.
Just as a PSA - there's nothing unique to AIs here - whenever you ask a question of anyone, in any way, they then have the memory of you having asked it. A lot of sitcoms and comedic plays have the plot premise build upon such questions that a person voiced then eventually reaching (either accurately or inaccurately) the person they were hiding the question from.
And as someone who's into spy stories, I know that a big part of tradecraft is of formulating your questions in a way that divulges the least about your actual intentions and current information.
If anything, LLM-driven AIs are the first technology that in principle allow you to ask a complex question that would be immediately forgotten. The thing is that you need to be running the AI yourself; if you ask an AI controlled by another entity, then you're trusting that entity with your question, regardless of whether there's an AI on the way.
This was a surprisingly big thing back in the early 2000s with The War Against Terror. I think that it was mostly for reasons of 'chilling effect', but the media made everyone aware that the Department of Homeland Security were paying attention to what books people took out of the library.
What was curious about this was that, at the time, there were few dangerous books in libraries. Catcher in the Rye and 1984 was about it. You wouldn't find a large print copy of Che Guevara's Guerrilla Warfare, for instance.
I disagree about how libraries minimise the risk of anyone knowing who is reading what. On the web where so much is tracked by low intelligence marketing people, there is more data than anything that anyone can deal with. In effect, nobody is able to follow you that easily, only machines, with data that humans can't make sense of.
Meanwhile, libraries have had really good IT systems for decades, with everything tracked in a meaningful form with easy lookups. These systems are state owned, therefore it is no problem for a three letter agency to get the information they want from a library.
Libraries don't tend to have consolidated, centralized IT. As a result, TLAs have to actually make subpoenas to the databanks maintained by individual, regional library groups, and The ALA offers guidelines on how to respond to those (https://www.ala.org/advocacy/privacy/lawenforcement/guidelin...).
This, of course, doesn't mean your information is irretrievable by TLAs. But the premise of "tap every library to bypass the legal protections against data harvesting" is much trickier when applied to libraries than when applied to, say, Google. They also aren't meaningfully "state-owned" any more than the local Elk's Club is state-owned; the vast majority of libraries are, at most, a county organ, and it is the particular and peculiar style of governance in the United States that when the Feds come knocking on a county's door, they can also tell them to come back with a warrant. That's if the library is actually government-affiliated at all; many are in fact private organizations that were created by wealthy donors at some point in the past (New York Public Library and the Carnegie Library System are two such examples).
Many libraries also purposefully discard circulation data so as to minimize the surface area of what can be subpoena'd. New York Public Library for example, as a matter of policy, purges the circulation data tied to a person's account soon after each loaned item is returned (https://www.nypl.org/help/about-nypl/legal-notices/privacy-p...).
Have you seen the list of books fascists want to ban? I think GP's point was exactly to emphasize that when we're talking about "dangerous books", we're talking about books that indicate you might not be a toe-lining member of The Party. We're talking about any book that any powerful person decides is some sort of threat, even if it's merely a threat to their ego.
Not dangerous at all! An analogy would be comparing a pea-shooter to an automatic rifle, or a thimble full of shandy when compared to a gallon of vodka. There is not a dangerous word in my local library!
The questions and info you ask friends doesn't end up in a massive data profile on you stored in somebody's cloud to be used for future manipulation/marketing/profiling...
> And as someone who's into spy stories, I know that a big part of tradecraft is of formulating your questions in a way that divulges the least about your actual intentions and current information.
Not necessarily disagreeing with you, but if everyone started doing this, we will be in XY problem city.
Just as a PSA - there's nothing unique to AIs here - whenever you ask a question of anyone, in any way, they then have the memory of you having asked it. A lot of sitcoms and comedic plays have the plot premise build upon such questions that a person voiced then eventually reaching (either accurately or inaccurately) the person they were hiding the question from.
And as someone who's into spy stories, I know that a big part of tradecraft is of formulating your questions in a way that divulges the least about your actual intentions and current information.
If anything, LLM-driven AIs are the first technology that in principle allow you to ask a complex question that would be immediately forgotten. The thing is that you need to be running the AI yourself; if you ask an AI controlled by another entity, then you're trusting that entity with your question, regardless of whether there's an AI on the way.