When students study with AI: after 5pm, and five days before the deadline

Every study guide gives the same advice: start early and spread the work out. It is one of the oldest findings in the psychology of learning, going back to Ebbinghaus in 1885: the same hours of study, spaced over days, beat the same hours crammed into one. We built our study planner on it. Give it a deadline and it spreads your sessions across the days you have left.

Then we looked at the days students gave it. In half the plans, there were five or fewer.

Finding 1: most study plans start a week or less before the deadline

Between 14 July and 9 October 2026, students made 376 study plans. 76 of them, one in five, chose a steady pace with no end date. The rest named a deadline: an exam, a test, a submission.

Across the 297 plans with a usable deadline, made by 253 students, the median gap between making the plan and the deadline was 5 days. 70 plans, 24%, were made for a deadline that same day or the next. 43% were made within three days of it and 58% within a week. Only 18% were made more than a month ahead.

This is not a default pushing students toward short plans. Where the product suggests a date at all, it suggests two weeks or more; the students picked nearer dates themselves. The pattern held before and after we added a study plan offer to the chat on 28 September (25% of plans within a day before, 22% after). It holds in our two largest markets: in the United States, 34% of plans were made a day or less before the deadline and 71% a week or less (40 students); in India, 24% and 55% (42 students).

Finding 2: the three days before the deadline carry most of the work

To see what students did with those last days, we took the 205 plans whose deadline had passed by 8 October and counted the questions their owners asked the tutor in that subject. In the 14 days before the deadline, 140 of these students asked 673 questions.

63% of those questions came in the three days before the deadline (422 questions, from 89 students), and 33% came on the day before alone (221 questions, from 44 students).

Part of this is mechanical, and we want to be clear about it: a student who sets up a plan two days out cannot ask questions ten days out. That is finding 1 seen from the other side. Most of the work starts late, so most of the work happens late.

We do not report the deadline day itself or the week after it. Fewer than 30 students asked anything in that subject then, and 30 is our minimum for publishing any group.

Finding 3: prime time is 5pm to 10pm

Students ask an AI tutor for help when the school or working day is over. Of 11,708 questions from 2,997 students in 94 countries where we can tell the local time, 39% were asked between 5pm and 10pm. 35% came during the day, between 8am and 4pm. 17%, about one in six, came between 10pm and 5am, and 7% between midnight and 5am.

The busiest single hour is 6pm, with 8.3% of all questions, and 9pm is nearly level with it (8.2%). The quietest stretch is 3am to 6am (2.1%).

This is not a handful of night owls. The most active 5% of students account for 28% of questions, and counting each student once per hour instead of every question gives the same picture: 38% in the evening, 17% late at night. Students in Arab League countries keep the same hours as everyone else: 41% of their questions in the evening and 16% late at night, against 39% and 17% elsewhere.

Finding 4: in some countries, more than a third of the studying happens after 10pm

The late-night share is where countries differ most. In Argentina, 38% of questions come between 10pm and 5am, in South Africa 31% and in Pakistan 28%. In the UAE and Spain it is 4%.

We do not know why. School timetables, commutes, dinner hours, part-time jobs and household routines all differ between these countries, and our data cannot separate them. The size of the gap is the finding: in some countries, nearly one question in three is asked after 10pm.

Finding 5: the quietest evening is the first evening of the weekend

Counting active student-days (a student who asks anything on a given day counts once for that day), the weekly rhythm follows the calendar, and the quiet spot lands in the same place in two different calendars: the evening the weekend starts.

In the 72 countries in our data with a Saturday and Sunday weekend, Friday and Saturday are the quietest days and Sunday picks back up. In the 13 countries with a Friday and Saturday weekend, among them Egypt, Saudi Arabia, Jordan, Iraq and Algeria, Thursday is by far the quietest day, and the weekend itself is as busy as any weekday or busier.

What we think this means

The finding we least expected is about our own product. We designed the study planner for spaced practice, and students mostly use it as a rescue plan. That is not a failure of discipline we can lecture away. It is how studying has always worked for most people, and nothing in this data suggests an AI tutor changes it.

Two implications for anyone building in this space. First, design for the last 72 hours, because that is where most of the asking happens. A plan made the day before an exam should decide what to cover tonight, not lay out a four-week schedule the student no longer has. Spacing still matters for the students who plan ahead, and a good tool can serve both. Second, the hours that matter are the ones when nobody else is available. One question in six arrives between 10pm and 5am, and in some countries nearly one in three. Teachers, parents and tutors are mostly asleep then. Whatever help a student gets at that hour comes from a book, a friend who is still awake, or software.

Methodology

• Population. 11,197 accounts on iTutor's public platform, created by 9 October 2026. Of these, 4,523 asked the tutor at least one question, from 154 countries by signup country. Internal, staff, test, demo and app-store review accounts are excluded, and so are accounts that asked to be deleted.
• Window. Questions from 3 November 2025 to 9 October 2026, inclusive. The first counted question was on 4 January 2026, and 87% were asked on or after 1 August 2026. Study plans from 14 July 2026, the first one in the data, to 9 October 2026.
• What is counted. A question is one message a student sent to the tutor in chat, typed or dictated: 18,080 in all. Spoken conversations in the Voice room are not included. No message text was read or used to produce this report: every figure is a count of time stamps.
• Local time. Times are recorded in UTC and converted with the time zone of the country the account signed up from. Only countries that use a single time zone are included in findings 3 to 5 (94 countries). The United States, Canada, Australia, Mexico, Brazil, Indonesia, Russia and other countries with several time zones are left out of those findings, as are accounts with no recorded country. Study plans use the time zone of the student's own device, saved with the plan.
• Study plans. A deadline is the date a student set in the planner ("I have a deadline") or the exam date they gave the tutor in chat. The lead is the deadline minus the date the plan was made. Three plans whose deadline fell before their own creation date (edited later) are dropped. A student can move a deadline after making the plan; we use the current one.
• Finding 2. Questions asked by the plan's owner in conversations about the plan's subject, from 14 days to 1 day before the deadline, for plans whose deadline had passed by 8 October 2026.
• Suppression. No group smaller than 30 students is reported, and no total is reported that would let a smaller group be worked out by subtraction. Single hours of the day are charted as three-hour blocks for this reason.
• Anonymity. Every figure is an aggregate count. No individual, institution, document or message is identified.
• What this cannot tell you. These are iTutor's students, who chose an AI study tool; they are not a random sample of students. Signup country is a proxy for where a student lives, and travel or private networks will misplace some of them. Most of the data comes from August to October 2026, the start of the school year in much of the world, so this report says nothing about exam seasons across a full year. Study plans are a smaller, self-selected group (297 plans from 253 students), and the tutor can offer to build a plan in chat, usually when a student brings up an exam, which may over-represent students with an exam coming up. Read the direction as informative and the percentages as specific to this population and period.

Our earlier reports are at What 3,624 students in 145 countries actually do with an AI tutor and Given the choice, 29% of students asked their AI tutor to teach them, not answer them. We will repeat this one at the same address, so the figures can be compared over time. If you are researching this area and want a cut of the aggregate data we have not published, write to us.