Hull contrasted purpose-built and general-purpose models: training on clinically relevant data lowers the base rate of harmful responses — so guardrails start from a safer baseline — and shifts the model from an information-giving "assistant" stance toward information-gathering and appropriate referral (for example, in OCD or bipolar disorder). On outcomes, he presented proof-of-concept real-world data using latent growth mixture modeling, goal-attainment scaling, behavioral activation, and several social-health measures (perceived social support, social interaction, and outings), noting that interaction volume predicted improvement rather than harm.

He framed the work as early, proof-of-concept evidence and emphasized open questions — appropriate control groups, longer follow-up, and which therapeutic mechanisms AI is best suited to deliver — while disclosing that he is compensated by Slingshot AI.