Real-world environmental problems are typically vast, urgent, and complex. Confronted with such problems, we are often tempted to act fast by pulling together little bits and pieces from different fields and simply adding these to pre-existing models and frameworks. Seldom, though, do we pause long enough to look whether and for how long those larger structures we build can support reliable answers to our questions. In this Perspective, I critically discuss the current state of interdisciplinary coupled modeling of human-environment relationships, with a focus on the classical model virtues precision, generality, and realism. I draw examples mainly from integrated assessment models – popular coupled modeling frameworks that are increasingly coupled further with ecological models to address biodiversity questions in the context of broader global-change and sustainability challenges. Specifically, I discuss i) how limitations in our models’ training data and underpinning theories translate into excessively uncertain predictions, ii) how coupling even highly general sub-models can lead to hardly generalizable representations of indirect human-environment relationships, and iii) how representing ever more processes decreases rather than increases realism due to greater average measurement bias, a problem further exacerbated as we add processes based on their relevance for our own systems of interest, rather than for the real-world systems’ dynamics. I also explore barriers to advancing scientific modeling virtues amid other, non-scientific motivations for interdisciplinary modeling (e.g., cultural, economic, normative). Finally, I offer suggestions to modelers and other actors in science, science administration, and science policy to help promote a transition to more robust interdisciplinary coupled models that can remain powerful for addressing major sustainability challenges far beyond the next iteration of science-policy assessments.