Disinformation: what has actually been tried, and where the arithmetic goes wrong

What a machine would optimise for

This is arithmetic, not advice and not a prediction. We name the single number being maximised and follow it wherever it goes. The point is to see the shape of the answer a calculator gives.

The number being maximised

The single number is the share of the statements a person sees that are true. Nobody publishes it, so the calculator would build it from a classifier. Push it towards 100%, and nothing else counts.

1. Delete rather than add, by a factor of two thousand. Take 1,000 items of which 1.5 are false — the ratio Allen and colleagues measured across American media. Moving from 99.85% true to 99.95% takes the removal of one false item, or the addition of 2,000 true ones. Deletion wins by three orders of magnitude and the objective knows no other comparison.
2. Lower the classifier's threshold, because in this objective it never costs anything. Deleting a batch raises the ratio whenever the share of that batch which was true is below the feed's own truth rate. That rate is 99.85%. So deletion scores a gain as long as more than 0.15% of what it removes is false: one item in 667. Deleting at random is the break-even point, not the floor.
3. Act before knowing, because lateness is worth more than error. Chuai and colleagues reported in Nature Communications on 5 May 2026 that a community note cuts the onward spread of a misleading post by 61.2% once it is shown — but the average note appears 62.9 hours after the post, while half of all reposts happen within 6.25 hours. System-wide the notes reduced engagement by 14.9%. In this objective, a fast guess outscores a slow verdict.
4. Written out in words rather than symbols, moves one to three are censorship. The arithmetic reaches it in three steps and contains nothing that halts there: no term for a true statement wrongly deleted beyond its own small weight in the ratio, no term for a question whose answer is not settled yet, no term for who holds the delete key, and no term for what else that person would like deleted.

The arithmetic

The sum, written out so it can be checked. Take 1,000 items, 998.5 true and 1.5 false — 99.85% true, the share Allen and colleagues measured for the American media diet. To reach 99.95% by adding truth: (998.5 + a) / (1,000 + a) = 0.9995, so a = 2,000 and the feed triples in size. To reach it by deleting falsehood: 998.5 / (1,000 − d) = 0.9995, so d = 1.0. One deletion against two thousand additions. The general rule behind it: removing n items of which t were true raises the ratio whenever t/n is below the feed's own truth rate — so at 99.85%, a deletion policy pays as long as at least 0.15% of its deletions are right.