"I don't really understand this hypothesis as it assumes that information quality of AI generated content on the internet will drop as a result of ChatGPT, not increase."
It has to drop. ChatGPT can not source new truths except by rare accident.
I bet a lot of you are choking on that. So, I'd say this: Can you just "source" new truths? If you just sit and type plausible things, will some of them be right? Yes, but not very many. Truth is exponentially exclusive. That's not a metaphor; it's information theory. It's why we measure statements in bits, an exponential measure, and not some linear measure. ChatGPT's ability to spin truth is not exponentially good.
A confabulation engine becoming a major contributor to the "facts" on the internet can not help but drop the average quality of facts on the internet on its own terms.
When it starts consuming its own facts, it will iteratively "fuzz" the "facts" it puts out even more. ChatGPT is no more immune to "garbage in garbage out" than any other process.
"Finally, I think what you'll see happen in future iterations of ChatGPT to improve quality and accuracy is that content will be fed in weighted by how authoritative the source is"
Even if authority is perfect, that just slows the process. And personally I see no particularly strong correlation between "authority" and "truth". If you do, expand your vision; there are other "authorities" in the world than the ones you are thinking of.
> It has to drop. ChatGPT can not source new truths except by rare accident.
How are we defining a "truth" here? For example, if I want to find specific SQL query, which will work for my specific database schema and my specific version of MySQL I won't find that online. Traditionally I'd need to come up with the new query for this novel scenario, or I'd need to ask someone to do it for me (perhaps on Stackoverflow). Now ChatGPT can come up with these new novel queries instead. You're right that it can't do it's own research and come up with fundamentally new information, but it can come up with answers to questions never before asked based on what it can infer from existing knowledge.
I'd argue most of the useful stuff people do isn't coming up with things that are fundamentally new, but applying things that are known in new and interesting ways. If you're a developer this is what you probably do every day of the week. And ChatGPT can absolutely do this.
Secondly, I'd also argue regurgitation of known facts is not necessarily without value either. A good example of this is your typical non-fiction / text book. If you write a text book about mathematics, you don't necessary have to include new information for it to be useful. Sometimes the value comes from the explanations, the presentation or a focus on lesser covered topics. Again, ChatGPT can absolutely do this. It already explains a lot of things to me better than humans can, so in that sense it is an increase in quality over what I'd already able to find online.
As for your point on authority I do agree with you somewhat there. I suppose the point I was trying to make is that this isn't a blind process. There are content sources which you can under weight, or simply non include, if it has a negative impact on the quality of the results. You can also improve algorithms to help the AI make better use of the information it's trained on. For example, if I asked you to read Buzzfeed for a week you wouldn't necessarily get any stupider because you're able to understand what's useful information and what's not.
I think all you really need to ask here is whether the next iteration of ChatGPT is likely to provide better results than the prior iteration, and will the iteration after that produce better results again? If your answer is yes, then it suggests the trend in quality would be higher, not lower, as a function of time.
Finally, wherever AI is applied the trend is always: sub human ability -> rivals that of the average human -> rivals that of the average elite human -> super human ability. Is language fundamentally different? Maybe? I think you can argue that generative AI is very different to the AI used for something like Chess, but it would at least be an expectation if future iterations of this AI got progressively worse. Maybe this the best ChatGPT will ever be at writing code. I guess I just think that is unlikely.
----
Btw, this is just how I see things likely playing out. Given how new the technology is my certainty isn't very high. I initially agreed with your point of view, but the more I thought about my reasoning the more my position shifted.
Honestly, I'm not really impressed with "but what is truth anyhow?" as an argument method.
But in this case it really doesn't matter because regardless of your definition of truth, unless you use a really degenerate one like "The definition of truth is 'ChatGPT said it'", ChatGPT will not be reliably sourcing statements of truth.
"Finally, wherever AI is applied the trend is always:"
My statement is not about AI. My statement is about ChatGPT and the confabulation engine it is based on. It does not matter how you tune the confabulation engine, it will always confabulate. It is what the architecture does.
AI is not ChatGPT, or the transformer architecture. I can not in general disprove the idea of the AI that turns on one day, is fed Wikipedia, and by the end of the day has derived the Theory of Everything and a viable theory of FTL travel. What I will guarantee is that such an AI will not be fundamentally a transformer-based large language model. It might have some in it, but it won't be the top-level architecture. No matter how well equipped the confabulation engine gets, it won't be able to do that. It is fundamentally incapable of it, at the architectural level. This is a statement about that exact architecture, not all possible AIs.
It has to drop. ChatGPT can not source new truths except by rare accident.
I bet a lot of you are choking on that. So, I'd say this: Can you just "source" new truths? If you just sit and type plausible things, will some of them be right? Yes, but not very many. Truth is exponentially exclusive. That's not a metaphor; it's information theory. It's why we measure statements in bits, an exponential measure, and not some linear measure. ChatGPT's ability to spin truth is not exponentially good.
A confabulation engine becoming a major contributor to the "facts" on the internet can not help but drop the average quality of facts on the internet on its own terms.
When it starts consuming its own facts, it will iteratively "fuzz" the "facts" it puts out even more. ChatGPT is no more immune to "garbage in garbage out" than any other process.
"Finally, I think what you'll see happen in future iterations of ChatGPT to improve quality and accuracy is that content will be fed in weighted by how authoritative the source is"
Even if authority is perfect, that just slows the process. And personally I see no particularly strong correlation between "authority" and "truth". If you do, expand your vision; there are other "authorities" in the world than the ones you are thinking of.