Within 4 years most words read by a human will have been written by an AI. I expect that most people reading this will quickly agree with that prediction.
I can be more specific: people will spend less time reading "documents" (web pages, tweets, articles in journals) as time goes by in favor of having conversations with large language models (LLMs) or similar kinds of AI.
The big advantage of having a conversation with an LLM over reading a document or browsing a bunch of documents is how quickly you can hone in on the specific facts you most want to know.
The way it is now, the important effects of a web page are the effects it has on humans that read the page; within 4 years the important effects are the effects it has on what LLMs write (or say) to humans.
Documents like letters and emails, sent from one human to another specific human, will have the most "staying power": i.e., letter and emails will constitute an ever-increasing fraction of the ever-decreasing mind share or "human reader's share" (measured in person-words) enjoyed by documents.
Here a "document" is a string of words that has a strong tendency to persist unchanged over time: humans take the trouble to make it so that a document remains available for its audience to read. E.g., if I have the URL to a tweet Eliezer wrote 5 years ago, I trust that Twitter is taking the trouble to make it so that my following the URL will yield me the contents of the tweet.
Within 4 years, not only will most words read by a human be written by an AI, but that fact will be common knowledge because although it is hard (and getting harder) to tell whether a document has been written by a person or an AI, it is not hard to determine (by reading over the person's shoulder if need be) which of the words a person reads are part of a conversation with a LLM and which words are part of a document -- and the former will predominate.
Within 4 years most words read by a human will have been written by an AI. I expect that most people reading this will quickly agree with that prediction.
What are the odds that this could happen, without AI actually taking over?
"Most of what you read came from AI" and "AI is making all the decisions that govern the world" seem pretty highly correlated to me.
I don't see it like you do as long as the LLMs aren't as smart as humans (or more precisely as smart as organizations of humans, e.g., police departments and things like that).
Consider that the communist party in the Soviet Union had a near monopoly on what a Soviet citizen could read and could consume in the way of news and opinion, but still lost power.
Also, providing LLMs for use by humans will probably be a competitive market.
I think you'd have to already be living in a world of dehumanized communication, to not see this as a fundamental change in human existence. Even the sleepers entombed in The Matrix interact mostly with other human beings.
What are words even for? They convey the thoughts and feelings of other people. They store useful information, they convey questions and commands... So you're describing a world where people are mostly reacting to the thoughts and feelings of AIs, where AIs are in charge of all the information we use throughout the day, and where it is mostly AIs that are asking us questions and telling us what to do.
The vast majority of words have been written thousands or millions of times, by children, novelists, dictionary companies, and of course, AI. There will be very few words that are not available to AI.
I know that's not what you meant, but you'll have to define your meaning a LOT more clearly in order to get any meaningful answers.
You should start with an estimate of how many words have been written today, and a Fermi estimate of what fraction are primarily generated by AI (to whatever threshold you set for humans generating prompts and editing/curating responses). This will at least nail down your definitions enough to ask a question. Oh, and how do you count machine-to-machine data transfers? Is that writing? If not, what's the formal definition?
Then talk about timeframes, and your p(doom). Depending on how long AI exists, compared to humans, and how much overlap there is, this could lead to VASTLY different answers.
(over time)