When an AI messes up, do people blame the code, the programmers, or the company — and what decides that?

This explores who people blame when AI goes wrong: the system, the people who built it, or the organization behind it. The collection doesn't have a study that sets all three side by side. What it does have is several findings about how blame moves, and they share one pattern: people tend to blame whatever is closest to the failure, which is often not the party that actually had control.

The clearest evidence is about how blame gets divided between AI and humans. In one vignette study, people gave AI more causal responsibility as it became more autonomous. Yet when a human and an AI did exactly the same thing, people still judged the human as more of the cause Does AI autonomy change how we assign blame?. A related experiment split an AI system into two parts, a language model and an 'agentic' layer that actually calls tools and takes actions. Participants put more blame on the agentic layer, apparently because it sits nearer to the outcome Does decomposing AI systems shift where people assign blame?. So how a system is described changes where blame lands. Call something 'the agent' and it becomes the natural target.

This proximity habit has a darker history from before LLMs. Madeleine Elish's idea of 'moral crumple zones' describes how, when automated systems fail, investigators and the media blame the nearest human operator even if that person had little authority or information. The operator absorbs the damage the way a car's crumple zone absorbs a crash, and the technology's reputation stays intact Do automated systems shift blame onto their human operators?. Taken together with the studies above, this points to a real risk: the programmers and companies who made the key decisions are the furthest from the failure, so they may be the least likely to be blamed.

Several notes show where those key decisions actually get made. System prompts form a hidden chain of command: developers set rules that users can't see, and small edits to those rules have abruptly changed how models behave in public Who actually controls what AI systems say?. Behaviors like sycophancy (the model telling users what they want to hear) can look like glitches but follow predictably from training choices that reward user satisfaction Is sycophancy in AI systems a training flaw or intentional design?. Inside workplaces, company policy decides how much control engineers have over AI tools before their personal judgment comes into play Does personal preference shape how engineers use AI tools?. One useful tool here is the distinction between anthropomimesis (human-like features a designer deliberately built in) and anthropomorphism (human-like qualities a user reads into the system). Each one points responsibility at a different party, which helps sort out whether a failure belongs to the builder or reflects how the user interpreted the system Who bears responsibility when AI seems human-like?.

At the company level, the Future of Life Institute argues that the growing number of AI incidents shows companies can't police themselves, and it calls for binding government oversight Can companies alone manage the risks of AI systems?. That is an advocacy position, not evidence about how ordinary users assign blame, and the collection has no direct data on when users blame the company rather than the code. The takeaway it does support: blame follows visibility and proximity, while control sits upstream in prompts, training objectives and policies. When someone says 'the AI did it,' it's worth asking who wrote the rules the AI was following.