Recent reports in the New Scientist.

As the web fills up with text generated from AI & posted on websites & forums it will increasingly learn from itself (or its brothers).

The finding of research is: AI fed with its own output is that it results in "model collapse" as the output gets increasingly simpler with a smaller lexicon (a bit like the text of comics like the UK Sun) as the sum total of the trawled data gets dominated by AI generated "stuff".

There are workarounds like limiting AI learning to pre-AI dates, then you get stagnation. Even using "speech to text" from live feeds used as source data will have mountains of archive "stuff" and there is the spectre (here already) of faked video that gets thrown-in.

And did we believe everything on the web even before AI?

in the context of the web - AI is much more of the same, only it is super-charged and comes at us faster.