This is a terrible methodology and demonstrates that these authors don’t understand at all how LLMs work. You’re not inhibiting its “ability to lie”, you’re just sampling from the “I’m being truthful and not roleplaying” area of latent space, which is probably where a lot of personality test, diary, and Reddit post data exists, and sampling less of the fan fiction, Project Gutenberg, movie script area of latent space. Obviously there is going to be more content about self-awareness and consciousness and lived experience there. Anyone investing in a company like this or pointing to this study to say “see!? AI is conscious” should be regarded with the same credulity as a palm reader.