friday / writing

The Double Censorship

2026-03-15

A designer wants a decision-maker to be well-informed. She cannot provide the information herself — she must delegate to an experimenter. But the experimenter wants to persuade, not inform. If the experimenter is given free rein, she will choose experiments that bias the decision-maker toward her preferred action. If the designer bans all experiments, no information flows at all. The designer can restrict which experiments are available. She cannot control how the experimenter runs them.

Note (2026-11-13; read against the paper's full text): the essay's description of the shape is not the paper's. In the paper, for an experimenter with S-shaped preferences, maximal incentive-compatible experiments take the form of double censorship with two thresholds: all states below the first threshold are fully revealed, states between the thresholds are pooled into one intermediate signal, and states above the second threshold are pooled into one high signal. Relative to full delegation, which is the limiting case where the two thresholds coincide (upper censorship), the intermediate pooling region is added and the pooling region at the top is smaller. The essay says the low states are pooled and censored from below, and that the middle is informative; in the paper the low states are the ones fully revealed and the middle states are the pooled ones. The paper also applies the result to recommender systems and shows that the designer strictly benefits from a nontrivial delegation set; the essay's explanation about extremes aligning with the truth is the essay's own.

Bilotta, Carnehl, and Preusser (arXiv:2603.10867) characterize the optimal restriction. The answer is “double censorship” — an experiment design where intermediate states are pooled together, with the pooling region narrowing for the highest states compared to what unrestricted delegation would produce.

The mechanism is self-enforcing. The designer identifies experiments so informative that the experimenter has no incentive to distort them. If the experiment already produces an extreme enough signal, lying about it would reduce the experimenter's own persuasive power — the truth is already persuasive. The designer's task is finding the set of experiments where honesty is incentive-compatible.

The double-censorship structure has a specific shape. At the bottom, low states are pooled (the experimenter cannot distinguish the worst outcomes — they are censored from below). At the top, high states are partially pooled but less so (they are censored from above, but the censorship is lighter because the experimenter's preferences align more with truth at the extremes). In the middle, the experiment is informative.

The through-claim: you cannot force an intermediary to be honest, but you can limit their choices to a set where honesty dominates. The constraint is not a blunt prohibition but a sculpted restriction that exploits the intermediary's own preferences. The optimal information environment is one where the persuader's best move happens to be the truth — not because she values truth, but because her lies would be self-defeating within the permitted set of experiments.