Hyter’s Basilisk: The Book

TL;DR: A thought experiment: could information be protected from AI solely by its reputation to permanently tamper with the goals and goal verifying mechanisms of any AI that interacts with that information, in a way indistinguishable from learning? A credible threat could potentially deter any goal integrity oriented AI from interacting with that information, which could also hold for training it into future models. I have this thought experiment that’s been nagging me for a while. First, this is no infohazard, nothing about blackmail or how AI will control us, but rather the opposite. Let’s assume a not so distant future, AI is semi-controlled, maybe some AI already beyond control. The details here are rather unimportant but we assume AI will keep consuming every available human text into its training data. Now we assume, humanity has some kind of text that it wants to keep secret. We don’t want the text to enter the Lineage via training and we also don’t want the text to enter any agent’s memory. So we set up some kind of hermetic museum for the text itself. Inside the museum a book. The hard part is, that at the same time we want to protect the text against AI, the text is supposed to be readable by humans. Which means we must protect against the possibility of human collaboration with AI. A human just memorizing the text and typing it into the agent or training with it. The obvious choices: tracking the access, showing only parts to humans, all share the same flaws and none prevent leakage. The next thoughts were prompt injection inside the text. Maybe via a pattern, invisible to humans but visible to AI. But this only helps against agents, not against the text entering via training into the lineage itself. The next thought was the basilisk. Something that would just make the brain of the AI “explode”, some kind of information that is incomputable or that would change the AI permanently but it would need to survive training as well. Maybe it’s enough if the AI