One purpose of the science of ecotoxicology is to provide information for protecting ecosystems. Understanding the hazards of chemicals is essential to wise decision making, and it is now clear that community structure changes are closely linked to altered ecosystem function. Uncertainty is high when decisions are made from a small biological (toxicological) database. Individual bioassays provide little insight into biological interactions that are important in sustaining ecosystems. Artificial ecosystem experiments with herbicides demonstrate the limited predictive power of bioassays and ecological risk assessment methods that are heavily dependent on animal testing. Many herbicides interfere with unique pathways in photosynthetic organisms but are not very toxic to animals. For example, the herbicide atrazine is not considered toxic to fishes, because atrazine interferes with electron transport in photosystem II. But, adding atrazine at low levels (3–100 μg/L) to aquatic microcosms demonstrated significant increases in algal biomass, concurrent enhancement of nutrient recovery systems, and increases in the detectable number of heterotrophic microbial species. Higher levels of atrazine (100–300 μg/L) produced general collapse of these laboratory ecosystems. Low levels of atrazine capable of producing ecosystem‐level effects can occur from days to weeks in streams of midwestern agricultural areas. Conversely, the herbicide diquat is rapidly immobilized in the field if fine sediments are present. Laboratory bioassays tend to overestimate diquat toxicity if sediments are not present because the material rarely persists in the water column. A variety of measures of ecosystem condition are available for assessment of chemical effects. Community structure changes (especially of nontarget groups) and changes in ecosystem process variables have technical importance and are not assessed in current risk assessment paradigms. Regulators need to draw on a more comprehensive data set than is presently used to make risk assessment decisions. Sometimes, this may require using methods other than those considered standard for data development.