The hidden rhythms in our AI use

Every dataset has a story—but few announce themselves with terms of endearment. When Microsoft AI researchers dug into tens of millions of Copilot conversations, one finding that jumped off the screen wasn’t code or spreadsheets. It was a sharp uptick in conversations about relationships, dating, family and love on Valentine’s Day.

“I expected a slight fluctuation, but the spike is so obvious,” said Beatriz Costa Gomes, a Futures Researcher at Microsoft AI, in an interview with IBM Think.

That Valentine’s Day finding illustrates the bigger story in this new research, which analyzed 37.5 million anonymized conversations on Copilot between January and September 2025. Beyond single day spikes, the data captured the weekly cycles and seasonal rhythms of people’s AI use, said IBM Fellow Kush Varshney, in a recent episode of Mixture of Experts. These AI assistants “are synchronizing with our lives or we’re synchronizing our lives with them,” he said.

When the researchers zoomed out from Cupid’s handiwork, they also saw more patterns that felt startlingly personal, said Costa Gomes. In the middle of the night, for instance, conversations about religion and philosophy surged. “It’s 3:00 AM and you suddenly have existential questions,” she said. “Most people are probably asleep, so you turn to your AI confidant.”

The biggest surprise for the researchers was “during any time of day, morning or night, health [was] dominant on mobile,” said Costa Gomes. Microsoft AI will continue to study user conversations with Copilot users, she said.

“It’s been interesting to see how we started dumping everything we get from a doctor into a large language model,” Volkmar Uhlig, VP and CTO of Data Platforms at IBM, shared on Mixture of Experts, speaking to how instances of medical-related chats on LLMs have grown. In doing this, Uhlig said, users are able to talk through their health results in a way that might help them confirm or simply better understand their diagnoses. “We are seeing a trend toward using [tools like Copilot] as an expert model.”

The value of these particular patterns stems in part from the size of the dataset. Studies of AI conversations from the likes of Anthropic and OpenAI top out at around 4 to 4.5 million conversations. Microsoft AI’s research team looked at nearly ten times that number of conversations.

The data also captured the rhythms of daily life. Programming dominated on weekdays while gaming overtook it on weekends—so clearly the researchers could spot weekends just by looking at the graph.

This particular finding also raises questions such as whether users access Copilot through work and continue using it in off-hours through the same license, said Lauren McHugh Olende, a Program Director for AI Open Innovation at IBM, on the podcast. “If people are flip-flopping using the same Copilot application for programming in the week and gaming on the weekend, who’s paying for that?”

For Uhlig, this research could help enterprises get a clearer picture of what people seek, he said, especially since Microsoft has such a large consumer population. “It’s almost like a study of psychology where you see the topics of interest, what people need.”

This echoes how Microsoft AI’s CEO Mustafa Suleyman frames how Microsoft AI will apply the findings of this research. “What features would help support people in these day-to-day rhythms, what topics should be our biggest priorities?” he wrote in a LinkedIn post.

According to Suleyman, when it comes to building AI that’s actually useful, in life or in work, developers need not look further than the human behind the screen: “This is how we make AI more useful, for more people, more often.”