You skipped a step between d) and e) - the stealth reimposition of gatekeepers under the guise of sorting through the crap. On the Internet, this took the form of "recommendation algorithms". The problem with a recommendation algorithm is that, while it is sold as "finding choices you wouldn't have known about", what it's really doing is narrowing the menu. It has to be, because of how the data flow works:

1. Obtain data on what kinds of people enjoy which books, movies, music, etc

2. Train a neural network to predict the sampled distribution. Because that distribution already contains a lot of data for the most popular works, and less for less other works, the model is trained with a bias for already popular works. Conversely, less-popular or niche works may be forgotten completely by the model.

3. Deploy the model. Because the model has a bias to predict certain already-popular works, it will recommend them more often, while the ones with little data in the set are recommended less. This changes the true distribution of the data.

4. Repeat. Take new sampled data from modified distribution and train a new model on that distribution.

As a result of applying this cycle multiple times, the tails drop out. People don't watch or read niche works, so they get forgotten. This is ostensibly fixable, you could bias the model towards less well-known works, but it's more profitable to just make people pay to boost their posts. Bonus points: people who pay to boost are also polluting the model with a bunch of negative data about their work, so they are self-selecting into a class of easy marks who don't get recommended because they keep paying to show in front of people who don't want it.