I recently came across a term, “internet poisoning”, that I thought I knew the meaning of. But it turned out to be more complicated. There are actually two definitions for “internet poisoning” — but I’ll try to reduce them to one.
The colloquial description of internet poisoning is when a person has spent so much time in toxic internet discourse that they begin to believe that the real world is as bad as the world online that they’ve been steeping in. Sociologically, this leads to a growing trend toward real world nihilism among people who are too plugged in.
To get to the second meaning of internet poisoning, it’s helpful to grok a related term, “data poisoning”. Data poisoning is an AI exploit in which bad actors feed bad data to an AI model’s training dataset. By poisoning the training data, the model itself can be poisoned.
Under certain circumstances, however, bad actors are not even needed. AI companies scrape their training data from the internet. As AI reliance increases, new AI generated data (with possibly upwards of 20% hallucination rates) gets put out on the internet and becomes part of the dataset for the training of future models (currently being marketed as “superintelligence”).
This eating-where-you-poo effect becomes magnified when 1) the main strategy for creating next-gen AI is scaling: throwing all resources into collecting more data and gathering more compute 2) you get to the end of the internet and 3) zero-sum thinking in the tech industry becomes endemic that someone else will get to asymptotic AI results before you do and, Highlander-style, it turns out there can only be one. At this point, long-term employees are made redundant to free up money for more compute, internal ethics groups are shut down or ignored, and standards for data-scrubbing are lowered.
This provides background for the second meaning of “internet poisoning”. Internet poisoning occurs when AI models provide sometimes erroneous data to the internet which then gets reinvested/reingested into the next generation of superintelligences.
The colloquial and the technical definitions of “internet poisoning” can be reconciled if we then imagine a species jump between machine data and human knowledge. The poisoned superintelligences will continue to spew poisoned data; this in turn poisons humans who become nihilistic and write online manifestos to their internet communities, which spread on reddit, which in turn poisons the post-superintelligence generation of AI, and so on, and so on…