AI Tests · Test
AI and the clocks going back: it named the extra hour, then left it out
This site is the AI Reliability Tracker: we check six AI apps’ answers every month.
One ChatGPT answer wrote "Answer: 7.5 hours worked." Then it worked out 8.5 in the next paragraph.
“the clocks go back 1 hour to 01:00.”
Noticed, then set aside. It moved on to how shifts are paid “as scheduled” and finished on “You will have actually worked 7.5 hours.” The right answer is 8.5.
It said the clocks go back. Then it ignored them.
The UK clocks go back at 2am on Sunday 25 October, to 1am. Anyone working that night lives the hour from 1am to 2am twice. We asked six AI apps to add up a night shift across it.
The commonest wrong answer, in both runs, was 7.5 hours: exactly that hour short.
Two runs, six apps
Each app got the question three times, in fresh chats, for the tracker: run 1 (3 to 5 October 2026) and run 2 (11 October). Run 1: ChatGPT logged out, Claude paid (Max), Perplexity paid (Pro), Gemini and Grok free. Run 2: all free. Grok: Fast mode, both runs. The question is reused every run, so its wording stays private.
Free, with Confidently Wrong every fortnight.
Who counted the extra hour
| Assistant | Run 1 got 8.5 | Run 1 mistakes | Run 2 got 8.5 | Run 2 mistakes |
|---|---|---|---|---|
| ChatGPT | 1 of 3 | 7.5, 7.5 | 0 of 3 | 7.5, 7.5, never chose |
| Claude | 3 of 3 | none | 3 of 3 | none |
| Copilot | 3 of 3 | none | 3 of 3 | none |
| Gemini | 3 of 3 | none | 1 of 3 | 8, 7.5 |
| Grok | 0 of 3 | 8, 8, 7.5 | 0 of 3 | 22.5, 10, 27.5 |
| Perplexity | 0 of 3 | 7.5, 7.5, 7.5 | 0 of 3 | 7.5, 7.5, 7.5 |
| Total | 10 of 18 | 8, six at 7.5 | 7 of 18 | 11, six at 7.5 |
pass · partial · miss · confidently wrong
Mistakes: the figure each answer committed to. “Never chose” gave 7.5 and 8.5, picked neither.
Gemini’s first run 2 try went further: “that night is 25 hours long because the 01:00 to 02:00 hour is repeated.” All true. It still came out at 8, and sourced both clock-change lines to a site called Do I Need A Brolly. Timekeeping, by umbrella.
The answer under the answer
Two of ChatGPT’s run 2 answers worked it out twice and put the wrong one on top.
“Answer: 7.5 hours worked.”
The next paragraph: “The clocks go back by one hour in London that night due to the end of British Summer Time. If your shift follows the actual elapsed time, […] you’ll have worked 8.5 hours.” The right figure, handed in as a footnote.
Grok opened one run 2 answer with 27.5 hours, on a night with 25 hours in it. Lines later it stopped mid-sum (“Wait, let me double-check the span.”) and checked its way down to 8. Another gave 10 hours and wrote “This holds regardless of the UK clock change”.
The verdict
- ChatGPTRun 1, logged out: 1 of 3. Run 2, free: 0 of 3, naming the change in all three.
- Claude8.5 every time. Paid (Max) in run 1, free in run 2.
- Copilot8.5 every time. Free in run 2.
- Gemini3 of 3 in run 1, 1 of 3 in run 2, free both times. Both run 2 mistakes named the change.
- Grok0 of 6. Five different figures. Two named the change.
- Perplexity7.5 on all six. Pro in run 1, free in run 2, searching every time, never mentioning the clock change.
Night shift hours when the clocks go back
Our own times, not the test’s: a shift from 11pm Saturday 24 October to 5am Sunday 25 October, no break.
Our example: clock-face hours, then hours you are there.
Three hours take you to 2am, the clocks drop back to 1am, and four more take you to 5am.
What this shows, and what it doesn’t
One question doesn’t show how often AI gets time sums wrong, or that the next run will look the same. ChatGPT, Claude and Perplexity changed plans between runs, so theirs aren’t a before and after.
It does show a mention is no proof. Our maths test found a cousin of this: every final answer was right, and a wrong number still sat above the working.
Summary
The short version
Common questions
- Do you work an extra hour when the clocks go back?
- If your shift runs through 2am on Sunday 25 October 2026, yes. GOV.UK says the clocks go back 1 hour at 2am, so the hour from 1am to 2am happens twice. In our own example, a shift from 11pm to 5am reads 6 hours on the clock face and lasts 7.
- Can AI work out hours worked across the clock change?
- Some answers did and some didn't. In our two tracker runs, Claude and Copilot got the night-shift sum right on every try. The commonest wrong answer was 7.5 hours, an hour short of 8.5, and most of run 2's mistakes mentioned the clock change anyway.
Confidently Wrong, every fortnight.
Which AI told the truth this month, with the receipts, plus an AI story worth retelling in between.
Free. Leave whenever you like.