Abstract We discuss the possibility that causal confusion will be a significant alignment and/or capabilities limitation for current approaches based on "the scaling paradigm": unsupervised offline training of increasingly large neural nets with empirical risk minimization on a large diverse dataset. In particular, this approach may produce a model which uses unreliable (“spurious”) correlations to make predictions, and so fails on “out-of-distribution” data taken from situations where these correlations don’t exist or are reversed. We argue that such failures are particularly likely to be problematic for alignment and/or safety in the case when a system trained to do prediction is then applied in a control or decision-making setting. We discuss: * Arguments for this position * Counterarguments * Possible approaches to solving the problem * Key Cruxes for this position and possible fixes * Practical implications for capability and alignment * Relevant research directions We believe this topic is important because many researchers seem to view scaling as a path toward AI systems that 1) are highly competent (e.g. human-level or superhuman), 2) understand human concepts, and 3) reason with human concepts. We believe the issues we present here are likely to prevent (3), somewhat less likely to prevent (2), and even less likely to prevent (1) (but still likely enough to be worth considering). Note that (1) and (2) have to do with systems’ capabilities, and (3) with their alignment; thus this issue seems likely to be differentially bad from an alignment point of view. Our goal in writing this document is to clearly elaborate our thoughts, attempt to correct what we believe may be common misunderstandings, and surface disagreements and topics for further discussion and research. A DALL-E 2 generation for "a green stop sign in a field of red flowers". Current foundation models still fail on examples that seem simple for humans, and causal confusion and spurio