By Pascale Caron — for EntrepreneurIA
Analysis of the HBR study « Making Sense of Research on How People Use AI », published in November 2025
A quiet but profound turning point
What are users actually doing with AI? That’s the question Marc Zao-Sanders raises in a recent op-ed published by the Harvard Business Review. Going against the grain of technical promises or futuristic visions, the author proposes a shift in focus: observing concrete practices, as close to the ground as possible.
In just a few months, we’ve moved from technological enthusiasm to rapid normalization. But what does this normalization mean? What are the actual usage behaviors? And how do these behaviors quietly transform our relationship to knowledge, work, and others?
Three lenses to shed light on usage
The article offers an unprecedented triangulation: cross-referencing massive quantitative data (OpenAI, Anthropic) with qualitative field analyses from the “AI in the Wild” project. This approach invites us to deconstruct preconceived ideas about AI’s promises.
- OpenAI: quantifying the scale
OpenAI’s report on ChatGPT usage provides an initial snapshot. In 2025, millions of daily interactions are recorded. What’s striking is not so much the volume as the diversity. Behind the façade of a conversational assistant lies a multiplicity of expectations: writing faster, phrasing better, generating ideas, understanding a topic, organizing one’s thoughts.
Usage isn’t limited to developers or analysts. Students, teachers, managers, therapists, and content creators are all embracing it. We’re witnessing a hybridization of intellectual practices: AI is becoming the facilitating third party of a new grammar of thought.
But this report remains centered on the platform itself. It doesn’t deeply examine the motivations or the social contexts of use. What are people really looking for when they consult an AI? Clarity? Comfort? Productivity? Legitimacy?
- Anthropic: measuring inequalities
The second perspective comes from the Anthropic Economic Index, which draws on analysis of Claude, ChatGPT’s competitor. The indicator reveals a major geographic and social divide: usage is heavily concentrated in high-income countries, with significant gaps in interaction levels.
This finding highlights a strategic question: does AI reinforce inequalities in access to cognition? In developed countries, it becomes a lever for intellectual acceleration, decision support, and synthesis. Elsewhere, its impact is more limited, due to lack of bandwidth, reliable translation, or digital literacy.
The map of adoption mirrors that of structural inequalities. It directly challenges the promises of shared progress. Can we really speak of “AI democratization” if access remains conditional on a high socio-educational level?
- AI in the Wild: observing practices
Finally, Marc Zao-Sanders draws on a source often missing from technical discussions: ethnographic observation of usage. Through his AI in the Wild project, he analyzes thousands of real interactions on forums, public platforms, YouTube channels, and shared transcripts.
What emerges is less expected. Users aren’t just looking to produce more. They explore, test, dialogue, and sometimes confide. AI becomes a tool for self-reflection. It supports emotional dynamics, not just cognitive ones.
Several dominant patterns emerge: personal organization, emotional support, help with focus, articulating unclear thoughts, or seeking moral clarity in difficult decisions.
Far from productivity fantasies
Contrary to the image of a purely technical tool, the data suggests that AI is largely used for non-productive purposes. Zao-Sanders emphasizes this: the three most frequent uses in 2025 are not automation, translation, or code generation. They are:
- Emotional support (companion- or therapist-like)
- Personal life management (organization, prioritization, life choices)
- Help articulating abstract or intuitive ideas
This shift is fundamental. It shows that AI, far from being confined to the role of functional assistant, is making its way into the most intimate zones of human thought.
What does this reveal about our times? A growing need for clarity in a saturated world? A loss of bearings in the face of information overload? Or a new way of dialoguing with oneself, through an algorithmic mirror?
A relational technology?
Generative AI is perceived as a tool. But in practice, it sometimes comes closer to a confidant. The work of Eliot and Osler (Ethics & Information Technology, 2026) on “AI gossip” confirms this dimension: users talk to their AI as if to a person.
This anthropomorphization is not trivial. It creates an illusion of closeness. It facilitates projection. But it also blurs the boundary between technical assistance and social interaction.
Who is talking to whom? And what do we really want to say? Does technology act as an amplifier of speech, or as a substitute for human listening?
A cognitive tool… or a lever for dependency?
Another recent MIT study warns of an emerging risk: cognitive debt. The more AI produces content in our place, the more our own capacity to argue, synthesize, or structure declines.
Zao-Sanders also touches on this tension. AI is a support. But it can become a crutch. How can we preserve intellectual autonomy? What safeguards should we put in place to prevent the most vulnerable users (students, young professionals, isolated individuals) from using it passively?
This question is central for companies, universities, and public actors. It’s no longer enough to teach how to use AI. We must now learn to question it, to step back from it, to make it a critical partner — not a cognitive replacement.
The blind spots of research
The three sources cited by HBR are rich. But they still leave several blind spots:
- The voice of the excluded: little data comes from low-income countries or from people without structured access to technology. Yet that’s where the stakes of social transformation are most acute.
- The role of gender, age, and cultural contexts: who uses AI, in what settings, with what implicit expectations? Are usages gendered? Culturally situated? No firm answers to date.
- The quality of interactions: behind the usage statistics, what about the subjective experience? Do people feel understood, inspired, guided, manipulated? The data remains silent on the qualitative impact of these person-machine dialogues.
A question of design… but also of society
Zao-Sanders’s analysis raises a political question: do we want an AI that accelerates, enlightens, reassures, entertains… or all of these at once? The design of interfaces, models, and possible uses must reflect a vision. Yet today, this vision is shaped by market logic, more than by a collective project.
What AI do we want? An AI that helps each person think for themselves, or an AI that anticipates our needs to satisfy them automatically? An AI of autonomy or an AI of dependency?
The educational stakes: learning to dialogue with AI
In the face of this transformation, education becomes central. It’s not enough to learn to “use” an AI. We must train people to dialogue with it: identify its biases, understand its limits, structure their own thinking so as not to become dependent on it.
This requires a new educational contract. Who will carry it forward? Schools, businesses, the media? And with what resources?
Conclusion: observe before regulating
The Harvard Business Review article marks a methodological turning point. It doesn’t seek to define what AI should be, but what it is becoming in the hands of its users. This modest and empirical stance opens up a field of analysis too often neglected: real-world usage as political terrain.
In a world where AI discourse saturates the media landscape, this bottom-up approach, attentive to concrete practices, reminds us of the essential: it’s not technology that decides. It’s the uses, the intentions, the frameworks of action.
We just need to understand them, document them, and question them.
