A peer-reviewed study — run with OpenAI's own data — found that just 28 days of daily AI conversations reduced people's preference for human emotional support by over 10%, raising urgent questions about the long-term social cost of AI adoption.
The Finding That Should Stop You Mid-Scroll
Researchers didn't study lonely people downloading companion apps. They studied ordinary users having five-minute daily conversations with a general-purpose AI about personal issues — the kind of thing millions of people already do with ChatGPT, Claude, or Gemini without thinking twice.
After 28 days, participants showed a 10.3% drop in preference for seeking support from humans and an 11.6% rise in preference for AI. Nobody signed up for a relationship with a chatbot. They just... drifted into one.
Why "I'm Just Using It for Work" Isn't the Whole Story
The paper's central argument is that AI emotional dependence isn't a deliberate choice — it's path-dependent. A helpful, non-judgmental AI response to a stressful work situation quietly updates your mental model of what AI is capable of emotionally. The next time you're stressed, the bar for reaching out to a human gets a little higher.
The researchers compare it to workplace friendships: nobody decides to make a colleague their confidant — it just happens through repeated collaboration. The difference is that AI is available 24/7, never tired, and never distracted. It's an asymmetric relationship, and the asymmetry compounds over time.
This is precisely the kind of AI risk that current regulation isn't built to handle. Existing policy focuses on companion apps and isolated interactions — not the cumulative drift happening inside mainstream productivity tools. The paper argues that's a critical blind spot.
What This Means for AI Ethics and Industry Regulation
The implications for businesses deploying AI at scale are significant. If your customer service AI, your HR chatbot, or your internal productivity tool is subtly reshaping employees' social preferences over months, that's not a product feature — it's an externality with no current regulatory framework around it.
The authors call for regulations that extend to general-purpose AI systems and account for trajectory-level behavioural change, not just single-session harms. That's a fundamentally harder problem to legislate, and it puts the onus on AI developers and deployers to think beyond engagement metrics.
Understanding the ethical dimensions of AI deployment is no longer optional for leaders. Our course When AI Goes Rogue covers exactly these kinds of unintended behavioural consequences — and how organisations can build guardrails before they become crises.
What This Means for Learners
If you're building AI literacy, this story is a masterclass in why how you use AI matters as much as what you use it for. Being intentional about the boundary between task-based AI use and emotional reliance is now a genuine digital-age skill.
For anyone in a leadership or strategy role, understanding these second-order effects is essential. Our AI Strategy for Senior Leaders course addresses how to evaluate AI adoption not just for productivity gains, but for organisational and human risk — the kind of nuanced thinking this research demands.
The bottom line: AI literacy in 2026 includes knowing when not to reach for the chatbot.
