The bucket and the searchlight are metaphors for opposing theories of the growth of scientific knowledge. The bucket theory proposes that knowledge is gained by observing the world without preconceptions, and that knowledge emerges from the accumulation of observations that support a hypothesis. There are many problems with this theory, the most serious of which is that it does not appear to offer a means to distinguish between the many hypotheses that could explain a particular set of observations. The searchlight theory proposes that preconceptions are unavoidable and that knowledge advances through the improvement of our preconceptions – our hypotheses – by continuous criticism and revision. A hypothesis is a searchlight that illuminates observations that test the hypothesis and reveal its flaws, and knowledge thereby increases through the elimination of false hypotheses. Research into the risks posed by the cultivation of transgenic crops often appears to apply the bucket theory; many data are produced, but knowledge of risk is not advanced. Application of the searchlight theory, whereby risk assessments test hypotheses that transgenic crops will not be harmful, seems to offer a better way to characterise risk. The effectiveness of an environmental risk assessment should not be measured by the size of the bucket of observations on a transgenic crop, but by the power of the risk hypothesis searchlights to clarify the risks that may arise from cultivation of that crop. These points are illustrated by examples of hypotheses that could be tested to assess the risks from transgenic crops and their hybrids becoming weeds or invading non-agricultural habitats.

Environmental risk assessments can provide high confidence of minimal risk by testing theories, “risk hypotheses”, that predict the likelihood of unacceptable harmful events. The creation of risk hypotheses and a plan to test them is called problem formulation. Effective problem formulation seeks to maximize the possibility of detecting effects that indicate potential risk; if such effects are not detected, minimal risk is indicated with high confidence. Two important implications are that artificial test conditions can increase confidence, whereas prescriptive data requirements can reduce confidence (increase uncertainty) if they constrain problem formulation. Poor problem formulation can increase environmental risk because it leads to the collection of superfluous data that may delay or prevent the introduction of environmentally beneficial products.