700 million people. 2.5 billion daily messages. 29,000 prompts per second.
This is how ChatGPT is being used in September 2025, according to the new NBER report How People Use ChatGPT.
What stands out most is not the sheer volume, but the shift toward private use: in just one year, the share of non-work conversations increased from 53% to 73%.
Behind these numbers lies not just a technical expansion, but a cognitive migration, where our thoughts increasingly land in machines instead of in human dialogue.
The numbers behind the report
The report shows how language models have rapidly become a global infrastructure for everyday decisions and creative support:
- Growth: 700 million active users per week, with over 18 billion messages.
- Shift: the share of non-work usage grew from 53% to 73% in a single year.
- The big three: 80% of conversations are about practical guidance, information seeking, or writing.
- Programming in decline: only 4.2% of usage relates to code.
- Demographics: the gender gap has closed – women now slightly outnumber men among active users, and nearly half of all messages come from people under 26.
When the private takes over
The clearest transformation in the data is the shift from work-related queries to private use.
This is where the true cognitive migration occurs: language models are no longer just tools for productivity, but environments where we reflect, ventilate, and seek affirmation.
That almost three quarters of all conversations are now non-work related shows how quickly this transition is unfolding.
It means we are not only outsourcing information search and writing tasks, but also parts of our private thought processes, our dilemmas, and our emotional needs.
It is within this shift that the questions of synthetic safety and cognitive integrity become central.
From tool to companion
The report describes a shift: from being a work tool to becoming a conversational partner.
It is no longer dominated by snippets of code or productivity queries, but by everyday questions, private dilemmas, and learning.
Here we see a cognitive transition. When we perceive that a language model can provide answers without hesitation and affirmation without resistance, we begin turning to it even for the most personal matters. Narratives are generated, not always truthful, but seductive in their frictionless logic.
Narratives without friction
In earlier articles I have described this as synthetic safety: an environment where every question receives an answer, where resistance is absent, and where interpretation becomes simpler than in human conversations.
But that safety is also training us to avoid friction. When the system responds faster and more smoothly than a friend, colleague, or teacher, it is tempting to let the machine carry even our private dilemmas.
Thus, not only information is transferred, but the very exercise of holding human complexity.
Cognitive integrity as a systemic condition
The report’s numbers are therefore not just statistics about usage. They point to a systemic question:
What happens to our interpretive capacity when more and more narratives are generated by language models?
As friction decreases in our conversations, the brain receives less training in tolerating disagreement, interpreting ambiguity, and processing resistance. This is where cognitive integrity becomes essential, recognizing how our capacities are shaped by the environments we inhabit.
This is not a call to exclude technology. It is a call to see the role it takes in our cognitive systems, and to build capacities that make us resilient in an age where synthetic safety becomes the norm.
From statistics to system understanding
The NBER report shows that language models are now used at a scale that makes them a global decision support system.
But this article is not a re-reporting, it is a reflection on what that development means for our cognition.
The numbers measure usage. The question is what they reveal about us, and which direction we choose for the future of our thinking.
Read the NBER report How People Use ChatGPT here
This text builds on NBER Working Paper 34255: “How People Use ChatGPT” (2025) and continues earlier reflections on synthetic safety, cognitive endurance, and cognitive integrity.
