Algorithmic Review Changes Which Mistakes Tennis Officials Make

When correction changes the original call

Hawk-Eye review gives tennis players a way to challenge officials’ decisions. But its influence may begin before anyone requests a replay. A preprint by David Almog, Romain Gauriot, Lionel Page and Daniel Martin reports that officials became more inclined to call the closest balls “in” after review was introduced.

The article, Human Responses to Algorithmic Correction: Evidence from Centre Court, was first posted on January 30, 2024. This account concerns arXiv version 2401.16754v5, updated November 4, 2026.

The central finding concerns the composition of initial mistakes, not simply their total. Among balls landing within 20 millimetres outside the line, the adjusted rate of mistaken “in” calls was 8.6 percentage points higher under review. Among balls landing within 20 millimetres inside it, mistaken “out” calls were 4.1 points less frequent.

Why the two mistakes differ

Officials make their initial decisions without seeing Hawk-Eye’s assessment. Players can then challenge, and a successful challenge publicly reverses the call. Because review depends on players invoking it, some mistakes remain uncorrected.

The consequences of reversal differ. During a rally, correcting a mistaken “in” call ordinarily establishes who won the point. Correcting a mistaken “out” call is more complicated: the call has stopped play, so the chair umpire must decide whether to award the point or replay it. A replay cannot recreate the interrupted rally.

The shift toward calling close balls in is consistent with officials trying to avoid mistakes whose correction requires that additional, potentially contentious judgment. The study does not establish this explanation as the sole mechanism.

Reconstructing decisions before review

The researchers combine ball-location records, challenge records and video audits to reconstruct initial calls, rather than examining only decisions players challenged. Their consolidated sample covers 698 men’s singles matches across 35 tournament editions: 109 matches without review and 589 with it.

Hawk-Eye recorded ball locations at some events before player challenges became available, allowing the researchers to use the same accuracy benchmark under both regimes. The paper treats its readings as ground truth, while noting the system’s reported average error of 3.6 millimetres.

The analysis accounts for factors including distance from the line, speed, score, tournament tier and match round. However, review adoption was not randomized. A causal interpretation requires assuming that other changes would not independently have shifted close calls toward “in.”

The pattern persists when comparisons are restricted to named events observed under both regimes, although that smaller sample produces less precise estimates. Statistical uncertainty also matters for the headline results: the individual 8.6-point increase and 4.1-point decrease do not meet a conventional 5% threshold under the reported wild-cluster-bootstrap tests. A contrast comparing the closest outside bin with the average of the first two inside bins does meet that threshold.

What happens after an overturn

Within matches, subsequent close calls move toward avoiding the kind of mistake previously reversed. After a mistaken “in” call is overturned, later close balls become less likely to be called in; after a mistaken “out” call is overturned, they become more likely to be called in.

These associations provide complementary evidence of adjustment, but challenges and their outcomes are not randomly assigned. The data also do not identify individual officials, so the findings concern officiating crews rather than demonstrated changes within particular judges.

For balls within 100 millimetres of the line, the adjusted initial mistake rate was 1.1 percentage points lower with review, with a standard error of 0.8 points. That difference was not statistically significant. The finding is therefore no evidence of average deterioration, not proof that accuracy improved or effort stayed unchanged. For serves within 20 millimetres, initial accuracy instead worsened.

Valuing review means valuing human responses

The authors compare an investment calculation that holds human decisions fixed with one incorporating the initial error rates observed under review. Both hold correction fractions and unit costs constant, and distinguish losses from uncorrected errors from the costs of remedying corrected ones.

Using descriptive sample rates, equal losses for both error types and costless remedies, incorporating human responses supports approximately 32% more expenditure than the fixed-decision calculation. But with costless remedies, the ordering reverses when an uncorrected mistaken “in” call costs roughly four times as much as a mistaken “out” call.

These are conditional investment illustrations, not measured monetary benefits. The broader lesson is that aggregate accuracy alone can miss consequential changes in error composition. The evidence does not distinguish responses specific to algorithms from responses to review more generally, or establish that less reliable review systems would produce the same behavior.