Richard Stallman on ChatGPT

Bullshit Generator, Truth, and the Grep Analogy

• Some agree with calling LLMs “bullshit generators” in the technical Frankfurt sense: they produce fluent text without caring about truth, optimized for sounding right rather than being right.
• Others argue this is unfair: post-training explicitly tries to align outputs with truth and reduce hallucinations.
• Comparison with grep sparks debate: grep is deterministic and “truthful to its algorithm,” while LLMs are probabilistic and may confidently output falsehoods; critics say probabilistic algorithms are still algorithms and widely accepted elsewhere.

What Counts as “Intelligence” or “AI”?

• One side accepts Stallman’s definition: intelligence requires genuine knowing or understanding; LLMs lack semantics and world models and thus aren’t intelligent.
• Opponents say this is a semantic game (“submarines don’t swim” problem) and note that historically many pattern-recognition systems have been called AI.
• Some point out Stallman elsewhere accepts narrow ML systems as AI if outputs are validated against reality; by that standard, they argue, LLMs also qualify because labs do extensive validation.

Usefulness vs Reliability and Risk

• Many commenters find LLMs extremely useful for coding, shell commands, email drafting, explanations, and “mechanistic” tasks, sometimes outperforming average humans.
• Others stress they remain untrustworthy for high-stakes decisions: they can be confidently wrong, fabricate citations, and don’t “know when they don’t know.”
• A recurring view: they are powerful autocompletion/association engines, great for assistance but dangerous if treated as authoritative.

Free Software, Cloud Dependence, and Transparency

• Strong agreement with Stallman’s critique of closed, server-only deployment: users can’t inspect, run, or verify models; behavior can change or degrade without detectability.
• Some worry about regression, hidden knobs, and opaque incentives (e.g., ad-driven responses) in proprietary models.
• Open-weight models complicate the usual “publish the source” ethic, since training data and pipelines are hard to reproduce.

Meta: Naming and Rhetoric

• Several prefer “LLM” or “associator” over “AI” to avoid overclaiming; others accept “artificial intelligence” as established terminology.
• Opinions split on Stallman’s piece: some see it as accurate but under-argued or dated; others dismiss it as curmudgeonly yet consistent with his long-held freedom-focused philosophy.