This is definitely a valid concern. OpenAI did extensive data grooming to ensure high-quality inputs were used as training data. They went out of their way, for instance, to attempt to remove auto-translated content.

One interesting thing is that the situation you describe provides a huge moat for the first successful system. For instance, OpenAI can store and fingerprint all of the output from GPT-3 and ChatGPT. They can use these fingerprints to prevent training newer versions of GPT on old outputs. Less popular systems won't be able to sanitize the training data as well.

There must be an equilibrium point at some point. I assume the % of AI content will be far higher than 20% in a few years

I read that GPT adds some “watermarking” to its content to help with detection of pure AI generated content.

Certainly we already need tools for this just like those to detect plagiarism.

Cat and mouse game I know but AIs that can detect AI content seems like a reasonable outcome to me.

Reasonable for a limited amount of time. Eventually down the line someone comes out with an adversarialy trained GPT model that causes a high enough false positive rate that humans trying to use the system get pissed off.

Some would call that thinking…