Background Opioids are commonly first encountered in medical settings; 59% of individuals with opioid use disorder (OUD) report initial exposure via prescription. While not all individuals prescribed opioids develop OUD, we lack scalable tools to predict who is at risk. The initial subjective effects of drugs are thought to influence use of many substances: positive effects encourage use while negative effects deter use. We extended this idea to opioids by examining the initial positive and negative subjective effects of opioids in 117,508 research participants from 23andMe Inc., using a retrospective survey. We then compared the subjective responses of individuals who did or did not report OUD to identify the optimal questions to predict the risk of OUD. Methods Participants who endorsed prior medical opioid use completed 11 questions about their initial subjective responses to opioids (e.g. “When you first took pain medication, to what extent did you like the way they made you feel overall?”. A similar wording was used for the other questions: “less pain”, “euphoric”, “energized”, “normal”, “relaxed”, “nauseated”, “dizzy”, “tired”, “constipated, and “itchy”. Response options included “Not at all”, “Mildly”, “Moderately”, “Very much”, “Extremely”). After applying quality controls, we examined the responses of 84,296 participants: of these, 5.3% self-reported OUD. We used Logistic Regression, XGBoost, and Random Forest classifiers to predict OUD from subjective responses, combining them into an ensemble model. Model fit was assessed using area under curve (AUC) values for receiver operator characteristic (ROC) curves and precision-recall ROC (PR-ROC), curves as well as precision, recall, and F1 scores. Feature importance was evaluated using SHapley Additive exPlanations (SHAP) and odds-ratios were calculated for select items. A decision tree classifier was also trained to identify a minimal predictive question set. Results Positive subjective effects, particularly “Like Overall”, “Euphoric”, and “Energized” were the strongest predictors of OUD. For example, the individuals responding “Extremely” for “Like Overall” were 36.2 times more likely to have OUD. Analgesia and negative effects were much less predictive. The ensemble model of combined features was strongly predictive (ROC-AUC = 0.91); “Like Overall” alone achieved a ROC-AUC of 0.87. We present a two-step decision tree that uses “Like Overall”, “Itchy”, and “Energized”, which can identify a small high-risk subset with 77.4% prevalence of OUD and a much larger low-risk subset with 1.7% prevalence of OUD. Discussion Our results demonstrate that positive subjective responses are predictive of future misuse and suggest that vulnerable individuals may be identified and targeted for preventative interventions.