Will it lead to some machine learning collapse?
I wouldn't think so. Data is data. From the standpoint of automata, everything is ultimately reduced to a sting of bits. It may even be useful to be able to train AI's using CGI, for instance, in relation to automated vehicles. Not any different from humans using flight simulators.
Creating models and training AIs on them is useful, and a part of the contemporary AI landscape.
Might it lead to some changes in human's perception of the world, because people get a very big part of their knowledge using computers, connected to the Internet?
As nbro astutely points out, it already is. It's not only the false CGI content, but the scope of the search filter that dictates what information a websurfer gets. These results are controlled by algorithms, which evolve. Self evolving algorithms will make the process more opaque. It's definitely seems to be creating social problems already.
Is anyone thinking about this potential problem?
I'm sure there are papers out there on this subject. (Don't have time to search now, but I may do that and come back and amend with some articles and research papers.)
Two authors who are definitely thinking about this are Neal Stephenson and Hannu Rajaneimi. Stephenson addressed the "information unreliability" problem of the internet in Anathem. (It's not a major theme, but his ideas are quite insightful--Stephenson has a hard-science background, with a particular interest in computing.) Rajaniemi extends the ideas in the post-singularity Quantum Thief trilogy, where information and matter are interchangeable, and contains some very interesting ideas. (Rajaneimi holds two advanced mathematics degrees, which is useful in tackling a subject of such great complexity.)