Botshit

A critical term for generative-AI output produced without adequate concern for truth and allowed to enter a decision or communication without sufficient epistemic controls.

The term adapts Harry Frankfurt's account of bullshit to generative systems and organizational use. A model can produce plausible market claims without a truth-tracking process, and a workflow can pass them into a client document without source support. The failure can occur in model generation, retrieval, evaluation, automation, or organizational incentives; it is not defined by one person failing to read every line.

Builder example

Risk scales with two factors: how much truth matters in the domain, and how difficult verification is. An AI drafting social media captions operates in a low-stakes zone. An AI generating legal citations, medical dosage information, or financial projections operates where a single unverified claim can cause real harm. Match the depth of your verification to the stakes.

An assistant invents a plausible study title with an author name and year. The user pastes it into a client memo without verifying.

Use consequence-based source checks and constrain publication so unsupported claims cannot enter a decision path.

Common confusion: Hallucination describes an unsupported generated claim. Botshit emphasizes the wider production and distribution process that lets truth-indifferent output influence work; automated publication can create it without a manual handoff.