The most unsettling thing at the end of Arthur C. Clarke’s 2001: A Space Odyssey is not that David Bowman has become something other than human. It is that he has acquired dominion before deciding what to do with it. The novel enlarges intelligence without settling the purposes intelligence should serve.

Spoilers for the novel follow.

The first safe night

Moon-Watcher begins as a hungry creature who thinks he might reach the Moon by finding a sufficiently tall tree. Curiosity precedes an adequate account of distance.

Imagine a commuter whose bus route may be cut. At bedtime she checks the same local-news story again. The proposal has not changed; neither has the date of the vote. She is trying to work out whether there is something she must do tomorrow.

Three visits can look better than one. They may also mean that the first two failed to answer her question.

Now imagine an article that makes the situation legible: what has been proposed, what remains undecided, when the decision comes, and what she can do before then. She reads it once and puts the phone down.

Imagine this: a worker reports a threat. The complaint is received, assigned a number and closed because there is not enough evidence.

Four months later another worker reports the same conduct. New number, new reviewer, same conclusion.

A third report arrives after the first file has expired. The institution can now say that its available records show no pattern.

Nobody prohibited a sentence. The system only prevented three true sentences from meeting.

Angela Cerinotti’s Miti greci e di Roma antica ends with an index that accidentally delivers the book’s sharpest argument. Zeus receives 496 references. Near the end come the Vineyard Women with seven, Zagreus with eight, and Tithonus with eight. The numbers map the gravitational field of the pantheon, but they do not measure what remains in the reader. I remember the three Vineyard Women, indexed seven times, more clearly than the 496 appearances of Zeus.

When a language model invents a fact, the first question is obvious: what is true?

The second arrives wearing a lab coat: why did the model do it?

Ask the machine that just failed and it may provide a plausible, technical cause, neatly shaped like closure. It may also be a second hallucination.

We distrust unsupported answers. We are less disciplined with unsupported post-mortems.

A cause is another claim

A model says that it confused two names because they appeared near each other in training. Or that a tool call failed because a rate limit was reached. Or that missing context caused it to infer the wrong event. Each explanation has the texture of mechanism.

Portability is usually described as a moving problem. Export the archive. Transfer the account. Preserve the settings. Let the user leave with their data.

Necessary, and strangely insufficient.

A person can carry every transcript from one machine to another and still lose the part that mattered: which refusal was accepted, which correction changed the relationship, which harm created an obligation. The archive preserves statements. It does not preserve what those statements made binding.