The authors also critique the metaphor of AI "hallucinations", arguing that it misinforms the public and policymakers about the nature of LLMs. They outline the operation of LLMs and discuss the problems with their accuracy, noting that attempts to improve accuracy by connecting LLMs to databases or computational programs have been largely unsuccessful. The authors conclude that it is more appropriate to describe the false claims of LLMs as bullshit rather than lies or hallucinations.
Key takeaways:
- The paper argues that the false claims made by large language models (LLMs) like ChatGPT are not lies or hallucinations, but rather instances of "bullshitting" in the Frankfurtian sense, meaning they are produced without any actual concern for truth.
- LLMs are designed to produce text that appears truth-apt, but they are not concerned with truth, which makes their outputs more akin to bullshit than lies or hallucinations.
- The authors distinguish between two types of bullshit: 'hard' and 'soft'. 'Hard' bullshit involves an active attempt to deceive the reader about the nature of the enterprise, while 'soft' bullshit only requires a lack of concern for truth.
- The authors argue that the outputs of LLMs like ChatGPT are at least 'soft' bullshit, and possibly 'hard' bullshit if we view the model as having intentions based on its design. They conclude that it's important to recognize and call out the bullshit nature of ChatGPT's outputs.