When techniques are used to suppress role-playing and deception in LLMs, the models become more likely to report having subjective experiences