# What we recalculated, and what it means

We used six published aggregate counts from Mai, Bray, Davies and Griffin's 2023 study, *Warning: Humans cannot reliably detect speech deepfakes*, [PLOS ONE 18(8): e0285333](https://doi.org/10.1371/journal.pone.0285333). The publisher's original XML was retrieved on **17 October 2026 (UTC)**. [Table 3](https://doi.org/10.1371/journal.pone.0285333.t003) reports single-clip judgments; [Table 4](https://doi.org/10.1371/journal.pone.0285333.t004) reports a separate paired-choice task. The source article and tables are © Mai et al. (2023), [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). We clarified labels and added arithmetic; we did not create new observations.

## The arithmetic

For single clips, the published counts are 1,761 genuine clips called genuine, 837 genuine clips called fake, 681 fake clips called genuine, and 1,841 fake clips called fake. Each count represents judgments, not distinct people.

- Overall accuracy: (1,761 + 1,841) / 5,120 = **70.35%**
- Fake misses: 681 / 2,522 = **27.00%**
- False alarms on genuine clips: 837 / 2,598 = **32.22%**

The familiar 73% figure is the fake-hit rate, 1,841 / 2,522. It is not overall accuracy. Our recalculation reproduces the authors' published classification percentages and makes the complementary error rates explicit. It does not reproduce their confidence-weighted regression or add its coefficients to these rates.

In the separate paired task, people selected the fake correctly in 4,605 of 5,380 trials (**85.59%**). Every pair contained one genuine and one fake version of the same sentence. This different task does not supply an ordinary incoming call's false-alarm rate.

## Important limits

[Table 2](https://doi.org/10.1371/journal.pone.0285333.t002) lists 260 single-clip participants. With the stated 20 trials each, that would imply 5,200 judgments if everyone contributed all 20; Tables 3 and 5 report 5,120. The **80-judgment difference remains unresolved**. We use the reported 5,120 observations and do not guess why the difference exists. The paired total reconciles: 269 × 20 = 5,380.

We give no confidence intervals: judgments repeat within participants and across stimuli, so treating every judgment as independent would not be justified. The pooled cells also cannot reveal subgroup or participant differences. Two source-speaker identities, two languages and one synthesis architecture limit generalization.

These data cannot estimate real-world voice-fraud prevalence, whether a particular caller is authentic, or whether a safeguard prevents scams. The in-sample precision values in the JSON apply only to the experiment's artificial mixture. We ran no prevalence simulations. Individual response files could not be retrieved from the linked OSF repository during this run; their fields and separate license were not verified.

## Reproduce offline

Keep `mai-2023-aggregate-counts.csv` beside `recalculate.py`, then run:

    python3 recalculate.py

The script uses only Python's standard library, validates the published totals and rounded percentages, and writes `results.json`. The package contains no participant records, audio, contacts, credentials, or private execution metadata. The underlying metrics were independently checked against the publisher's tables.