CONCEPT
Soft and Hard Bullshit
Hicks, Humphries, and Slater's 2024 refinement of Frankfurt's framework — soft bullshit is speech produced in a truth-free zone; hard bullshit additionally creates the impression of truth-orientation. Large language models produce at least the first and may be engineered to produce the second.
In their 2024 paper 'ChatGPT is Bullshit,' Michael Townsen Hicks, James Humphries, and Joe Slater drew a useful refinement within Frankfurt's original framework. Soft bullshit is speech produced without concern for its truth — produced in a truth-free zone where the question of correspondence to reality simply does not arise. Hard bullshit is speech that additionally creates the impression of concern for truth — that presents itself as truth-aimed while being nothing of the sort. The distinction matters for evaluating AI systems, because it separates what follows structurally from the architecture from what is added by design choices about how outputs are presented.
In The You On AI Field Guide
Large language models produce soft bullshit by structural necessity. The system has no truth-orientation. It is generating sequences of tokens that are statistically probable given its training data. The outputs may correspond to reality or fail to correspond,