Improvements in HIV antiretroviral therapy (ART) have been accompanied by increasing recognition of the importance of adherence to treatment regimens for maximizing patient benefits while minimizing the emergence of drug-resistant virus. Whether clinicians should screen patients for adherence and only administer therapy to those believed likely to adhere has not been resolved. We first examine the implications of data drawn from a recent study reporting physicians' ability to predict whether patients will adhere to highly active antiretroviral therapy (HAART) or not. We then extend previously developed mathematical models of ART to include screening for adherence and focus on resulting drug resistance as well as on HIV and AIDS incidence at the population level. We show that although screening for adherence is likely to reduce the level of drug resistance compared with a policy of treating all HIV patients with HAART, rates of new HIV infections and AIDS cases in the population would likely increase unless screening accuracy is extremely (perhaps implausibly) high.