adding emphasis to words in unformatted text draws the reader's attention in ways that are not implied by the text itself.
you cannot figure out which word to highlight with the raw text "I didn't say he stole the money", but the semantics are completely different depending on if/where they exist.
you already have a mental model of surprise. adding visual noise to destroy what is already part of reading just makes things bad. the reason why syntax highlighting is successful is that the semantics are rigid. natural language is explicitly not that.