Soft and Hard Bullshit

In the AI Story

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, but the system has no mechanism for caring which. Output that happens to be true and output that happens to be false have identical status from the system's perspective: both are successful generations.

Whether the same outputs constitute hard bullshit depends on presentation. A system that produces statistically probable text and labels it as such — 'here is a statistically probable continuation of your prompt' — is producing soft bullshit without the additional pretense of truth-orientation. A system that produces the same text and presents it in the rhetorical register of knowledgeable human discourse — with confident assertions, apparent citations, the texture of authoritative communication — is producing hard bullshit. The design choice to adopt this register is what converts the structural soft bullshit into the more dangerous hard variety.

The distinction clarifies what is at stake in AI interface design. A system architecturally incapable of truth-orientation can nonetheless be more or less honest about its limitations. Disclaimers, uncertainty quantification, and interface features that remind users of the system's nature reduce the hard-bullshit quotient without changing the underlying soft-bullshit architecture. The responsibility for the framing falls on the builders of the interface, not the architecture of the model.

The hard-vs-soft distinction also illuminates the specific danger of fluent AI output. The more polished the presentation, the more the soft bullshit acquires the markers of hard bullshit — and the harder it becomes for consumers to maintain the evaluative stance that would distinguish plausible from accurate. The Deleuze passage in The Orange Pill is a textbook case of hard bullshit generated without intent: the system produced rhetorically confident prose that created the impression of informed philosophical engagement, and the impression was convincing enough to nearly survive into the published text.

Origin

Hicks, Humphries, and Slater introduced the distinction in 'ChatGPT is Bullshit,' published in Ethics and Information Technology in 2024. The paper argued that 'hallucination' misdescribes AI output and proposed Frankfurt's bullshit framework as a more accurate characterization.

The distinction has been taken up in subsequent work on AI ethics, interface design, and education, where the difference between architectural bullshit and rhetorical bullshit has practical consequences for how systems should be designed and how users should be trained to interact with them.

Key Ideas

Soft bullshit is architectural. It follows from the system's structural indifference to truth. No design choice can eliminate it without changing the architecture.

Hard bullshit is presentational. It is produced by design choices about how outputs are framed — confident register, apparent authority, rhetorical markers of truth-orientation.

The categories are cumulative. Hard bullshit includes soft bullshit. The question is whether additional presentational choices make the structural bullshit more dangerous.

Interface design has moral stakes. The choice to present architecturally truth-indifferent output in a register that mimics truth-oriented discourse is an ethical choice with consequences for what users are trained to expect and evaluate.