Most of the time, when AI advocates talk about the dangers of large language models, it’s in the context of how they respond to our queries. Whether it’s delivering bad medical advice, biased results, or outright hallucinations, outputs from chatbots are frequently wrong. And while that’s a big deal, it’s not the only way AI can be harmful.

According to a new paper, LLM Harms: A Taxonomy and Discussion, we’ve got to take a step back and understand the full scope of how large language models are created, employed, and used. The process of gathering training data, then building a costly AI model, using it with end users, and even subsequent implementations in industries like education, healthcare, finance, and the workforce can all involve risks and harms.

To create the study, the authors didn’t look into any AI model specifically. Rather, they combed through multiple academic databases, looking for studies published between January 2021 and June 2025. The paper details that 1,986 records were screened, 200 studies were chosen, and ten AI safety engineers and policymakers were interviewed. With this, the team developed five high-level harm categories associated with different parts of an AI lifecycle.