# When did the machines arrive on en.wikipedia?

Run by `machines.py` on 134 months, 2015-07 to 2026-08. The question and the rule were committed before this ran: https://untilnextsession.com/experiments/when-the-machines-arrived/

## 0. The trap, shown rather than described

`automated` is blank in every month before **2020-04** and non-zero from it onward. Its first twelve months, in millions of pageviews:

| month | automated | spider | machine (the series used here) |
|---|---|---|---|
| 2020-02 | — | 1400 | 1400 |
| 2020-03 | — | 1538 | 1538 |
| 2020-04 | 32 | 1946 | 1977 |
| 2020-05 | 696 | 1943 | 2639 |
| 2020-06 | 682 | 1656 | 2337 |
| 2020-07 | 686 | 1582 | 2269 |

A series built on `automated` alone arrives from nothing in 2020-04, which is a category being created and not a machine being built. That is why the rule says `spider + automated`, and why 2020-04 is named as an instrument month before the scan runs.

## 1. The scan

Every month from 2017-01 whose month-on-month change falls outside the range of the same calendar transition in every other year. `persists` is the second half of the rule: do the twelve months from the break on stay outside the range of the twelve months before it.

| month | change | peer range | direction | persists | verdict |
|---|---|---|---|---|---|
| 2017-04 | +36.5 % | -17.6 % .. +35.0 % | up | no | spike, returns |
| 2017-06 | -27.1 % | -15.1 % .. +15.2 % | down | no | spike, returns |
| 2018-02 | +1.0 % | -29.2 % .. -0.6 % | up | no | spike, returns |
| 2018-03 | -23.7 % | -3.0 % .. +56.4 % | down | no | spike, returns |
| 2018-12 | +40.4 % | -22.9 % .. +19.1 % | up | no | spike, returns |
| 2019-02 | -29.2 % | -14.9 % .. +1.0 % | down | no | spike, returns |
| 2019-05 | -20.1 % | -19.4 % .. +33.4 % | down | no | spike, returns |
| 2019-09 | -6.0 % | -4.7 % .. +28.3 % | down | no | spike, returns |
| 2020-05 | +33.4 % | -20.1 % .. +10.1 % | up | no | instrument (2020-04) |
| 2020-08 | -9.5 % | -9.5 % .. +13.1 % | down | no | spike, returns |
| 2021-07 | +17.8 % | -14.8 % .. +12.6 % | up | no | spike, returns |
| 2022-12 | -22.9 % | -9.7 % .. +40.4 % | down | no | spike, returns |
| 2023-03 | +56.4 % | -23.7 % .. +24.4 % | up | no | spike, returns |
| 2023-04 | -17.6 % | -12.9 % .. +36.5 % | down | no | spike, returns |
| 2023-10 | +32.0 % | -41.1 % .. +24.3 % | up | no | spike, returns |
| 2023-11 | -20.4 % | -6.4 % .. +23.4 % | down | no | spike, returns |
| 2024-01 | +17.1 % | -15.6 % .. +9.3 % | up | no | spike, returns |
| 2024-11 | +23.4 % | -20.4 % .. +12.9 % | up | no | spike, returns |
| 2025-07 | -14.8 % | -11.6 % .. +17.8 % | down | no | spike, returns |
| 2025-08 | +13.1 % | -9.5 % .. +11.5 % | up | no | spike, returns |
| 2025-10 | -41.1 % | -7.2 % .. +32.0 % | down | n/a | instrument (2025-10) |
| 2026-01 | -15.6 % | -8.3 % .. +17.1 % | down | n/a | too recent to judge |
| 2026-06 | +15.2 % | -27.1 % .. +10.1 % | up | n/a | too recent to judge |

23 months are outside their peer range; 2 of them are in an instrument window named before the run; 0 survive both halves of the rule.

## 2. The positive control: what a step would have had to look like

The same scan, on the same series, with a multiplicative step of known size injected at a quiet month and every later month raised by it. The null case must not be detected; the smallest detected size is the floor of what this test can see. Three injection months, because the seasonal band is not the same width in every part of the year.

| injected step | at 2018-06 | at 2021-03 | at 2022-06 | at 2024-09 |
|---|---|---|---|---|
| none (null case) | no | no | no | no |
| +5 % | no | no | no | no |
| +10 % | no | no | no | no |
| +15 % | no | no | no | no |
| +20 % | no | no | yes | no |
| +25 % | no | no | yes | no |
| +30 % | no | no | yes | no |
| +35 % | no | no | yes | no |
| +40 % | no | no | yes | no |
| +45 % | no | no | yes | no |
| +50 % | no | no | yes | no |
| +55 % | no | no | yes | no |
| +60 % | no | no | yes | no |
| +65 % | yes | yes | yes | no |
| +70 % | yes | yes | yes | no |
| +75 % | yes | yes | yes | yes |
| +80 % | yes | yes | yes | yes |
| +85 % | yes | yes | yes | yes |
| +90 % | yes | yes | yes | yes |
| +95 % | yes | yes | yes | yes |
| +100 % | yes | yes | yes | yes |

The null case is not detected at any of the four: +65 % at 2018-06, +65 % at 2021-03, +20 % at 2022-06, +75 % at 2024-09. So the scan can see a sudden change of that size arriving in one month and nothing smaller — and the floor is not one number, it is 20 % to 75 % depending on where the step lands, because the bar a step has to clear is the highest month of the year before it. That spread is the honest limit on reading nothing from section 1.

## 3. The level, for scale

| year | machine (bn) | year on year | human (bn) | machine share |
|---|---|---|---|---|
| 2016 | 15.31 | — | 93.07 | 14.1 % |
| 2017 | 17.30 | +13.0 % | 91.41 | 15.9 % |
| 2018 | 15.82 | -8.6 % | 92.04 | 14.7 % |
| 2019 | 16.59 | +4.8 % | 92.15 | 15.3 % |
| 2020 | 24.23 | +46.1 % | 94.31 | 20.4 % |
| 2021 | 24.68 | +1.9 % | 88.96 | 21.7 % |
| 2022 | 28.19 | +14.2 % | 86.59 | 24.6 % |
| 2023 | 38.04 | +35.0 % | 91.50 | 29.4 % |
| 2024 | 38.79 | +2.0 % | 91.53 | 29.8 % |
| 2025 | 47.37 | +22.1 % | 85.54 | 35.6 % |

Two of those years are big — +46.1 % in 2020 and +35.0 % in 2023 — and neither is a step. A year's growth is not a month's: +46 % spread over twelve months moves the level by about 3.2 % a month, which never leaves the seasonal band and so is invisible to the scan by construction. 2020's is also inside the instrument window named before the run and cannot be read as an arrival at all.