Marvin Minsky (1980), MIT Artificial Intelligence Laboratory, A.I. Memo No. 603, November 1980, 25 pp. — https://hdl.handle.net/1721.1/5701. Published as a chapter in L. Vaina and J. Hintikka (eds.), Cognitive Constraints on Communication, Dordrecht: Reidel, 1984, pp. 175–200 — DOI-10.1007/978-94-010-9188-6_10. This filing was read from the memo.
Key takeaways
- Thesis. “Common sense logic is too unreliable for practical use. It cannot be repaired, so we must learn to avoid its most common malfunctions.” Humour has a special role in learning and passing on how to avoid them.
- Censors. These are learned suppressors, in a “B-brain” that watches the “A-brain”. They recognise the precursors of an unproductive mental state and prevent it, quickly and without phenomenology, which is why they are unconscious. A joke heard before is less funny because a censor has since been built.
- The common element of humour is “unexpected frame-substitution”: a scene is described from one viewpoint and then, often by one word, re-viewed in another.
- Positive and negative learning. General methods cannot be patched without losing generality, so their failures must be handled subtractively, by censors. Particular analogies are handled additively, by links.
- Intellect and affect. The “affect-intellect ‘dumbbell’” does more harm than good. Cognitive traumas such as the nearness paradox, the liar or Zeno create censors much as Freudian taboos do, without needing an authority figure.
- Laughter disrupts the absurd line of thought and focuses on it, so that the censor-learning agency can absorb it. Sexual humour stays funny because those censors are “slow learners”.
Standing in the record
Filed on 2026-10-01 at the owner’s request, as part of Minsky’s Society of Mind programme. Its Note 1, headed “The Society of Mind”, places it with three earlier works: the frame memo (LIT-410), Plain Talk (LIT-399) and K-lines (LIT-402). “The key idea is to reject the conventional view of the mind as a Single Agent… rather, the mind is composed of many smaller minds.”
It extends the frame memo. Its section III restates frames, terminals, default assignment and frame-shifts, and builds its account of commonsense logic and humour on them. Commonsense inference is “learning to make shifts between frames that have terminals in common”, and a joke is an improper or unexpected shift. It also uses K-lines’ partial mental states (its Note 2) and Plain Talk’s ambiguity of thoughts (its Note 7), but it does not build on them in the same way. The 1986 book and The Emotion Machine (2006) are filed alongside by another pass. How they take up censors, and Note 12’s sketch of pleasure as a “constructed illusion” of coherence, is for those filings to say. I have not read them.
NOTE-361 is the close reading of 2026-10-01, from the memo, and it placed the work: Active. It is worth reading for the idea of negative, anecdotal meta-knowledge and for a computational answer to a hole in Freud’s theory of jokes. As evidence for any of it, it has none.
Within this record:
- Somatic markers. The somatic-marker hypothesis (LIT-384) has learned situation–feeling links that covertly prune the options reasoning considers. The gambling-task study (LIT-383) reports nonconscious biases steering choice before the reason is known. Minsky’s censors are the cognitive mirror image. They are learned, unconscious and anticipatory, and they prune thought by its precursors, not by feeling. Both reject a clean separation of affect from intellect. The pairing is mine.
- Logic. In the logic papers held here, paradox is a property of a formal system. In this memo it is an event in a developing mind.
The memo carries no instruction for machine-learning practice, and no anthology topic holds it.