Understanding how individuals conceptualize addiction is an important approach to the study of substance use etiology. We asked participants in a large free-response study of of intuitive conceptualizations of addiction among alcohol and cannabis users and co-users to share their beliefs about the benefits and harms of alcohol and cannabis, and to explain in simple terms what it means to be addicted. Using a frontier language model (ChatGPT-4o) we extracted structured representations of people's beliefs and explanations, assessing the extent to which responses represented 11 clinically relevant diagnostic symptoms from the DSM-5 section on Substance Use Disorders. People's beliefs showed clear substance-specific biases, attributing more clinically relevant symptoms to alcohol than cannabis. A prompt-context manipulation that contextualized participants' substance-neutral explanations as relevant to either cannabis or alcohol revealed evidence sometimes for similar, and for other times opposite direction, substance-specific biases imposed by the ChatGPT annotation process itself.