#59 | Your AI at 10am ≠ your AI at 10pm

TL;DR: Today: Microsoft watched 37.5 million conversations with Copilot and noticed something obvious once you see it: you’re not using one AI. You’re using two. Same tool, two completely different roles. Your desk has one job. Your pocket has another.

👋 Happy Friday,

You’re not the same person at 10 a.m. and 10 p.m.

And your AI has figured this out.

​Microsoft​ spent nine months analyzing 37.5 million de-identified Copilot conversations between January and September 2025, and the pattern is hard to miss.

When you move from desk to pocket and from daytime to evening, the questions you ask shift with you.

Turns out we’re more predictable than we’d like to admit.

The workday has a shape

Between 8 a.m. and 5 p.m., desktop conversations are dominated by the topic “Work and Career”.

You can almost read the clock from the queries alone.

On weekdays, “Programming” climbs in rank; on weekends, it slides down while “Games” climbs instead. Coding during the week, gaming on Saturday and Sunday.

The system simply reflects the way your workday already runs.

We built machines to be flexible and adaptive, then used them on the exact same schedule every single week.

The rhythm is so reliable that the AI could set its watch by us.

Your phone ignores office hours

Mobile usage looks different.

On phones, “Health and Fitness” sits at the top of the topic ranking every single hour of the day, across every month in the dataset.

Middle of the day, middle of the night, beginning of the year, end of summer—it doesn’t move.

People come back to the same concern: their bodies, their health, and what to do about both.

There’s something quietly revealing about this. Health becomes urgent at inconvenient times.

It’s a constant background hum—the thing you think about at 3 p.m. and again at 3 a.m. when you can’t sleep. The phone became the place where that worry goes.

The intent data gives this more texture. On mobile, health questions are often tagged as “Getting Feedback or Advice”, not only “Searching for Information”.

Instead of just asking what something is, people are asking what they should do next. The conversations tilt from facts toward judgment.

As in: “I have these symptoms” becomes “Should I be worried about these symptoms?” which is a completely different kind of question, psychologically speaking.

The phone stays available when the workday ends and the questions feel more personal. Office hours don’t apply.

February told its own story

In February, another pattern appeared.

As Valentine’s Day approached, conversations about “Personal Growth and Wellness” slowly rose in rank. On the day itself, “Relationships” spiked.

People brought Copilot questions about connection, timing, and how to handle situations that matter to them.

Those conversations clustered in the evening and late-night hours, on phones rather than desktops.

Think about that for a second.

On the day designed for human connection, millions of people asked a machine for relationship advice.

There’s something both sweet and slightly absurd about it. The questions were real.

The machine just happened to be there.

And also happened to be trained on essentially every romance novel, advice column, and Reddit relationship thread ever written, which makes it either the perfect advisor or the worst possible one, depending on your view of crowd-sourced romantic wisdom.

The same model underneath gets pulled into different roles depending on when and where you reach for it.

The device is the frame

The ​Copilot Usage Report​ points out something the product teams have now started to act on.

A desktop assistant works best when it focuses on information density and task execution.

A mobile assistant becomes more useful when it feels concise, responsive, and tuned to personal context.

​Axios​ reported that Microsoft is already experimenting with these differences in how Copilot appears on each device.

The underlying model stays the same. The way it meets you changes.

At the desk, it behaves like a colleague that helps you gather information, draft text, review code, and get through “Work and Career” tasks.

In your pocket, it behaves more like a thinking partner—one that listens to questions about health, relationships, and how you’re doing.

Not that it actually listens in any meaningful sense, but the experience feels enough like listening that the distinction stops mattering after a while.

Well, you didn’t sit down and plan this distinction. It formed gradually as you kept reaching for the same device in the same situations.

We trained the AI to have two jobs without realizing that’s what we were doing.

2 a.m. has its own kind of honesty

Late at night, another topic rises: “Religion and Philosophy”.

As the report shows, these conversations grow more frequent in the small hours, well after the workday is over.

The questions shift from “How do I do this?” to “What does this mean?” and “Where does this leave me?”

These questions rarely show up on a work laptop at noon. They appear when the room is quiet, the screen is small, and there aren’t many other places to put them.

There is something very ordinary about this. People have always had late-night questions and few places to take them.

An AI system trained on large amounts of text has become one of the places where those questions now land.

The machine doesn’t judge. It doesn’t get tired. It’s awake when you are. So you ask it things you might not ask anyone else.

The quiet shift from search to guidance

One of the most interesting parts of the report is the change in intent.

Across topics, “Getting Feedback or Advice” has grown faster than “Searching for Information”. People are moving from “tell me about this” toward “help me think this through.”

On a desktop, that often looks like asking Copilot to review drafts, suggest improvements, or check reasoning.

On mobile, it often looks like advice on habits, health choices, and interpersonal situations. The same intent label covers both, but the subject matter shifts with the device.

This maps to the kinds of questions people are willing to test first with a system that is always available and never tired.

Also: never offended, never impatient, never keeping a mental tally of how many times you’ve asked the same basic question in slightly different ways. Which is honestly liberating, even if it’s also slightly depressing when you think about it.

The gap that remains

Inside ​Microsoft​, people working on Copilot are aware of the tension here.

Sarah Bird, chief product officer for responsible AI, has spoken about the fact that people bring sensitive topics into these chats even though the systems lack the legal and professional safeguards of, say, a doctor or a therapist.

The underlying models still hallucinate, of course. They still mix accurate information with confident mistakes. They don’t hold responsibility for outcomes.

Yet, as the report shows, millions of people are using them to think through matters of health, work, and relationships.

Here’s the strange part: we know the machines get things wrong. We know they sound more certain than they should.

And we keep asking them anyway. The behavior is ahead of the guardrails. That’s the gap.

It’s also, arguably, very human—we’ve been seeking guidance from imperfect sources since the invention of advice columns, horoscopes, and self-help books.

The difference is that this particular imperfect source sounds remarkably authoritative and is available at 3 a.m. when everything else is closed.

What the pattern really shows

Much of the AI conversation focuses on “one assistant everywhere”—the idea that a single system follows you across contexts and devices.

The Copilot usage data points to something more human and less uniform.

You behave one way when you sit at a keyboard between 8 a.m. and 5 p.m. You behave another way when you are lying in bed at midnight with a phone in your hand.

The same model answers both, but your relationship to it shifts with context.

The report simply shows that people have already folded AI into the daily rhythm of their lives: as a work aid at the desk and a quiet companion in the pocket.

Which sounds either comforting or vaguely dystopian, depending on your general disposition toward technology and loneliness. (Probably both, if we’re being honest.)

At your desk during work hours, Copilot looks like a colleague who helps you move projects forward.

On your phone at night, it feels closer to a confidant—one that listens while you work through whatever is on your mind.

The interesting part: you already taught yourself what role you needed, just by using it.

We didn’t need instructions. We just reached for the phone when we had a certain kind of question, and the desk when we had another.

The machine learned the pattern by watching what we did. Isn’t that awesome?

Cheers,

Mark
The AI Learning Guy
👋⚡😎

Interesting Sources

• ​Microsoft AI – It’s About Time: The Copilot Usage Report 2025​
• ​Microsoft Source – Overview of Copilot usage patterns​ – Detailed breakdown of usage across devices and times of day
• ​Axios – Exclusive: Microsoft Copilot is getting personal​ – Discussion of interface and experience differences on desktop vs. mobile
• ​Microsoft Responsible AI Leadership​ – Commentary by Sarah Bird on user behavior and sensitive uses of AI

Note: No single website has all the answers. This list serves as a starting point for those who want to explore or satisfy their curiosity about AI. Links: Links with * are affiliate links. See disclosure below.