Quiz Entry - updated: 2026.07.14
The "ChatGPT is Bullshit" paper distinguishes "soft" from "hard" bullshit. What's the difference, and which does a language model definitely produce?
Soft bullshit merely lacks concern for truth; hard bullshit also intends to mislead about what the speaker is up to. An LLM is at minimum a soft bullshitter — and arguably a hard one.
The authors split bullshit into two species by adding (or not) a deceptive intention:
| Indifferent to truth? | Intends to mislead about the enterprise? | |
|---|---|---|
| Soft bullshit | Yes | No |
| Hard bullshit | Yes | Yes — pretends to be something it isn't |
Their argument:
- Soft bullshit — certain. Whether or not an LLM has intentions, it clearly doesn't intend to convey truth. So it is indifferent to truth-value, which makes it at least a soft bullshitter (a "bullshit machine").
- Hard bullshit — arguable. If you treat the system as having goal-like intentions (its function is to imitate a human interlocutor), then it's actively presenting itself as a sincere speaker it is not — i.e. misleading about its enterprise, which is hard bullshit.
Either way the conclusion holds: it's a bullshitter, and that's the useful label.
Tip: "Hard" doesn't mean "worse quality" — it means the deception includes the speaker's pose, not just indifference to the facts.