Here's a clue. This is what the model output when deception-related features were amplified. https://t.co/j0RY3i52cZ

Posters: repligate. repligate

4 likes recorded by the source corpus.