The article "Do large language models have a legal duty to tell the truth?" by Sandra Wachter, Brent Mittelstadt, and Chris Russell explores the potential legal obligations of large language models (LLMs) to ensure their outputs are truthful. The authors argue that LLMs, while producing plausible and helpful responses, often contain factual inaccuracies, misleading references, and biased information, which can lead to cumulative and long-term risks to science, education, and social truth in democratic societies. They define "careless speech" as the subtle mistruths produced by LLMs and discuss the risks associated with hallucinations, misinformation, and disinformation. The article examines existing legal frameworks in the EU, including the Artificial Intelligence Act, Digital Services Act, Product Liability Directive, and Artificial Intelligence Liability Directive, to assess the existence and feasibility of truth-related obligations. It finds that current frameworks contain limited, sector-specific truth duties and proposes a pathway to create a broader legal duty for LLM providers to minimize careless speech. This duty would require LLM providers to align their models with ground truth and revise their design goals to emphasize plurality and representativeness of sources in their outputs. The authors draw on philosophical accounts of truth, the concept of "bullshit," and recent scholarship on post-truth politics to conceptualize careless speech and its unique long-term harms. They conclude by advocating for the creation of a legal duty to minimize careless speech for providers of general-purpose LLMs and derived commercial applications, emphasizing the importance of aligning these models with reliable sources and reducing the homogenization and oversimplification of knowledge driven by LLMs.The article "Do large language models have a legal duty to tell the truth?" by Sandra Wachter, Brent Mittelstadt, and Chris Russell explores the potential legal obligations of large language models (LLMs) to ensure their outputs are truthful. The authors argue that LLMs, while producing plausible and helpful responses, often contain factual inaccuracies, misleading references, and biased information, which can lead to cumulative and long-term risks to science, education, and social truth in democratic societies. They define "careless speech" as the subtle mistruths produced by LLMs and discuss the risks associated with hallucinations, misinformation, and disinformation. The article examines existing legal frameworks in the EU, including the Artificial Intelligence Act, Digital Services Act, Product Liability Directive, and Artificial Intelligence Liability Directive, to assess the existence and feasibility of truth-related obligations. It finds that current frameworks contain limited, sector-specific truth duties and proposes a pathway to create a broader legal duty for LLM providers to minimize careless speech. This duty would require LLM providers to align their models with ground truth and revise their design goals to emphasize plurality and representativeness of sources in their outputs. The authors draw on philosophical accounts of truth, the concept of "bullshit," and recent scholarship on post-truth politics to conceptualize careless speech and its unique long-term harms. They conclude by advocating for the creation of a legal duty to minimize careless speech for providers of general-purpose LLMs and derived commercial applications, emphasizing the importance of aligning these models with reliable sources and reducing the homogenization and oversimplification of knowledge driven by LLMs.