It'll also be worse than before, since:
1. "Ads" through LLMs are so much easier to personalize to the viewer and can be deployed at just the right moment to influence them in a way that might pass as organic.
2. Companies/politicians/wealthy people can simply pay for a model to be generically favorable to them, where positive words/events are mentioned and negative aspects are given dynamically-generated excuses or buried. (Or unfavorable to opponents, if the model-company thinks they can avoid lawsuits over it.)
3. Broad bias is easier to purchase, where a consortium or political party can pay to favor one entire class of thing over another, without developing individual ads for specific products or candidates.
In the end, the risk-profile is similar to hiring a human "personal assistant"... who takes secret bribes to manipulate you.
You can't download a Google or Meta but you can download a local model. This time is different