When an AI system blames its 'acting' part over its 'thinking' part for a mistake, is that fair — or just because it was closer to the damage?

This explores whether being closer to a harmful outcome makes someone look more responsible for it. The corpus doesn't study physical distance in the classic philosophy-experiment sense, like pushing someone versus pulling a lever. What it does show is a close cousin of that instinct at work in AI: people tend to blame whatever sits nearest to the harm in the chain of events, even when that thing had little real control.

The clearest evidence comes from a study that split an AI system into two parts: a language model that 'thinks' and a tool-using agent layer that 'acts.' Participants blamed the agent layer more, because it was the part that touched the outcome Does decomposing AI systems shift where people assign blame?. The researchers call this a proximity heuristic. There's a catch, though: it only clearly won out when both parts were the same kind of thing, two pieces of software. Once you mix humans and AI, a second bias takes over. People blame AI more as it becomes more autonomous, but they still judge a human as more causal than an AI that did exactly the same thing Does AI autonomy change how we assign blame?. So proximity matters, but it competes with a stubborn sense that humans are the 'real' causes.

Put those two instincts together and you get the most consequential finding here: the moral crumple zone. When an automated system fails, investigators and the media tend to pin the blame on the nearest human operator, such as the pilot or the safety driver, whatever that person's actual authority or information Do automated systems shift blame onto their human operators?. The operator is close to the harm and human, so both biases point at them at once. Like the crumple zone of a car, the person absorbs the impact, and the technology's reputation comes out intact. Being near the harm becomes a liability in itself.

Why would people's sense of cause work this way? Research on causal reasoning suggests our intuitions take shortcuts. For example, we don't properly discount one cause when another already explains the outcome, and LLMs trained on human text pick up the same errors Do large language models make the same causal reasoning mistakes as humans?. Pure causal models also leave out the associative and emotional pulls that shape real judgments Can causal models alone capture how humans actually reason?. 'Whoever was closest' is exactly the kind of associative shortcut those formal models miss.

The practical upshot is easy to miss. As AI systems are built from more layers (models, agents, tools, human overseers), the way the system is drawn shapes who gets blamed. Put a human 'in the loop' at the last step and you may have created a scapegoat, not a safeguard. If you want to go deeper, start with the crumple-zone note, then the decomposition study.