AI proofread a class email. It missed a wrong weekday.
The usual way to proofread
Tested in GPT-5.5 (OpenAI API), Oct 2026
Most of us paste the email in and type one short request. This fictional email has two wrong weekdays on purpose: Saturday 8 November 2026 and Wednesday 5 November.
It fixed one and skipped one
Real output · GPT-5.5 (OpenAI API) · Oct 2026
In our test, the AI noticed 8 November is a Sunday and changed the class to Sunday. It left Wednesday 5 November as written.
It guessed which half was wrong
When a weekday and a date disagree, either one could be the mistake. The AI chose for Rosa, and only you know which half is wrong.
Source: SlideModel, November 2026 Calendar
Ask for a date check instead
Tested in GPT-5.5 (OpenAI API), Oct 2026
This version asks for a date list with the real weekdays and tells the AI to flag mismatches, not change them. Change 2026 to the year your dates fall in.
Both wrong weekdays, flagged
Real output · GPT-5.5 (OpenAI API) · Oct 2026
In our test, the AI listed all three dates and flagged both mismatches. It left every date and weekday as written, so the fix stays with you.
Then check your own calendar
This was one run on one model, and weekday maths can still go wrong. We checked: 1 November 2026 is a Sunday, so the 5th is a Thursday and the 8th is a Sunday.
Try it: Use the date-check prompt on your next invite before you send it.
Source: SlideModel, November 2026 Calendar
Sources and assumptions
- SlideModel, November 2026 Calendar: November 1, 2026 is a Sunday; 5 Nov 2026 is a Thursday, 7 Nov a Saturday, 8 Nov a Sunday, 15 Nov a Sunday (checked 2026-10-24)
- Our own test run, engine/ai-test.js (GPT-5.5 via the OpenAI API), record.tests ids date-before-gpt and date-after-gpt: The before and after outputs quoted on slides 3 and 6 (run 2026-10-25T02:01Z UTC, i.e. 2026-10-24 local) (checked 2026-10-24)
Assumptions:
- The sample email is fictional (Rosa's Bakery, Mill Lane). The two wrong weekdays were put there on purpose: 'Saturday 8 November 2026' and 'Wednesday 5 November'. 'Wednesday 5 November' has no year in the email; the prompt tells the model to assume 2026.
- Use only test ids date-before-gpt and date-after-gpt. The other entries in record.tests are a dropped subject-line test (before-gpt, after-gpt) and failed Claude runs (output null). Do not use or mention them.
- Only GPT-5.5 through the OpenAI API was run. Claude runs failed in this environment, so the slides must not claim the prompt was tested in Claude or in the ChatGPT app.
- One run per prompt. Results can differ on another run or another model, so the slides say 'in our test'.
- Weekdays were checked by us against the SlideModel calendar and a Python calendar calculation.
The short version
- A plain proofread can miss a wrong weekday.
- It may guess which half of the pair is wrong.
- Ask for every date with its real weekday.
- Tell it to flag, not change.
- Check flagged dates on your calendar.


