Generative AI and the bullshit singularity (daedtech.com)

A writer revisits their earlier “facadeware” critique to examine generative AI, arguing that while LLMs aren’t facadeware (they’re genuinely impressive engineering feats), they are prolific “bullshit machines.” Using Frankfurt’s definition—bullshitters don’t care about truth, only effect—the piece frames modern LLMs as statistical next-token predictors optimized to please users, not to verify facts. That design produces useful conveniences (contract triage, brainstorming, prototyping, summarization) but also systematic “hallucinations”: confident, hard-to-detect falsehoods that scale with model usage. The author coins a provocative “bullshit singularity” risk as models churn ever more content and potentially learn from their own low-quality outputs, amplifying disinformation and lowering overall signal-to-noise. For AI/ML practitioners the essay is a practical caution: accept LLMs’ strengths while assuming and engineering for their epistemic failures. It distinguishes bullshit-tolerant tasks (social media snacking, boilerplate writing, conversational comfort) from truth-sensitive workflows that require human-in-the-loop verification, testing, and immediate execution checks (e.g., running code, validating contract citations). Key implications include the need for better hallucination mitigation, robust evaluation and provenance, dataset hygiene to avoid feedback loops, and tooling that surfaces confidence and provenance so users can separate persuasive prose from verified fact.