We Automated Bullshit
Critics argue that large language models like ChatGPT are “bullshit generators” in the philosophical sense: they produce fluent, persuasive text without any built‑in concern for truth, which makes their confident errors especially dangerous when users fail to verify outputs. Others counter that, used as tools by knowledgeable people—e.g., for code assistance, search, summarization, or idea generation—these systems can dramatically boost productivity, provided their limitations, training data quality, and susceptibility to misuse (such as propaganda or legal missteps) are clearly understood.
Usefulness vs. “bullshit” nature
- Many commenters say LLMs are extremely useful for hard‑to‑search facts, brainstorming, rewriting, summarizing, and “rubber‑ducking” problems.
- Others stress that this doesn’t contradict them being “bullshit generators”: they optimize for plausible language, not truth.
- Some argue the essay underestimates this duality: the same mechanism that produces bullshit can also surface valuable information faster than traditional tools.
Verification, expertise, and risk
- Strong consensus that answers must be verified; several people habitually cross‑check outputs.
- Critics note many users do not verify, citing lawyers sanctioned for submitting fabricated cases from ChatGPT.
- A recurring point: LLMs are safest and most effective when users already have domain knowledge to judge correctness.
- Concern: in high‑stakes areas (law, medicine, policy), confident wrong answers can be dangerous, and there is no built‑in “loop‑closing” fact-check.
Comparison to search engines & Wikipedia
- Some find LLMs better than Google for obscure queries; others doubt those anecdotes or attribute them to Google’s decline.
- Wikipedia is contrasted: it cites sources and has community correction; LLMs don’t reveal provenance and can’t be systematically corrected.
- Several predict LLMs will eventually become as normalized as Wikipedia, but with different failure modes.
Definition of “bullshit” and epistemology
- Multiple comments reference the philosophical definition: speech indifferent to truth, not merely “false.”
- Under this lens, LLMs always produce bullshit, because they model text likelihood, not reality or intent.
- Others push back that this overstates things: models encode a lot of accurate information and often behave like a knowledgeable layperson.
Coding and productivity use cases
- Many developers report large productivity gains: boilerplate generation, refactoring, documentation, RAG-style code search, and explanation of snippets.
- Others log high error rates, especially in C/C‑like languages, or subtle issues (security, edge cases) that “work” but are wrong.
- Common view: they resemble a bright but junior assistant—good for drafts and patterns, not for unreviewed production code.
Data, alignment, and societal concerns
- Some blame bullshit on mixed‑quality training data (social media, propaganda); others on architecture (no world model or fact‑checking).
- Worries include: scalable propaganda, scammers, misuse in courts, and erosion of trust in information.
- A few see LLMs as merely automating existing “bullshit economies” (bureaucratic reports, marketing copy), with both liberating and threatening implications.