One major difference is that generated content up until recently was pretty obvious. Tons of stuff like finance articles are autogenerated using templates, and SEO spam is obviously not intended for you as a human.
The rest is generally churned out en masse at the cheapest price, so in practice it contains no content and is very poorly written.
ChatGPT can produce decent quality content faster and cheaper than most humans. Despite not being fully accurate, and falling apart in certain domains like math, it has an amazing breadth of topics and things it can do at an acceptable level.
Right now, enough prompt engineering work is required that it still takes handholding to get ChatGPT to churn out content. But given where we are now it seems well within reach for the next gen of models to be able to go from “Write me an article about X that covers Y and Z” to “Write me 100 articles about varying topics in X” to “Take in the information from this corpus and distill it into 50 articles based on the most interesting parts.”
The main thing that should stay safe is detailed technical content like programming guides where you need to actually be able to reason about the material to produce good content, and can’t just paraphrase the ten thousand related sample materials in your training set. ChatGPT is decent about giving mostly-working code snippets (especially if it can use a library, although it may just make one up) but getting it to reason through things will probably require an entirely different approach to how it works. Still, because it’s already capable of producing technical content that passes a basic first glance, it could precipitate a trust crisis. I worry more about what happens when people try to get ChatGPT to generate recipes, or give medical advice, or operate in the support group/personal advice/etc. space.
The rest is generally churned out en masse at the cheapest price, so in practice it contains no content and is very poorly written.
ChatGPT can produce decent quality content faster and cheaper than most humans. Despite not being fully accurate, and falling apart in certain domains like math, it has an amazing breadth of topics and things it can do at an acceptable level.
Right now, enough prompt engineering work is required that it still takes handholding to get ChatGPT to churn out content. But given where we are now it seems well within reach for the next gen of models to be able to go from “Write me an article about X that covers Y and Z” to “Write me 100 articles about varying topics in X” to “Take in the information from this corpus and distill it into 50 articles based on the most interesting parts.”
The main thing that should stay safe is detailed technical content like programming guides where you need to actually be able to reason about the material to produce good content, and can’t just paraphrase the ten thousand related sample materials in your training set. ChatGPT is decent about giving mostly-working code snippets (especially if it can use a library, although it may just make one up) but getting it to reason through things will probably require an entirely different approach to how it works. Still, because it’s already capable of producing technical content that passes a basic first glance, it could precipitate a trust crisis. I worry more about what happens when people try to get ChatGPT to generate recipes, or give medical advice, or operate in the support group/personal advice/etc. space.