Are AI-driven emotional dependence and loss of judgment already common harms, or just scary possibilities nobody's measured yet?

This explores how common the personal harms are when people treat AI as a mind, as opposed to the larger societal risks that get more attention. The short answer is that these individual-level harms are judged to be happening already, but the collection has no solid count of how many people they affect. In the expert surveys behind Which AI risks are already harming individual users today?, emotional dependence and loss of autonomy are rated as high-probability and already occurring. Larger risks, such as humans losing social status to AI or political conflict over AI rights, are still rated as unlikely but severe. So the personal harms are the ones already here, while the dramatic ones are still possibilities. That is an expert judgment of likelihood, though, not a measurement of how many people are affected.

The closest thing to population-level evidence is a year-long study of Character.AI users in Does sustained engagement with AI companions harm well-being?. Sustained companion use did predict lower well-being, but mostly because users spent less time with people face to face. The chatbot itself wasn't simply the harm. That changes what 'widespread' means here. The risk may be less about AI damaging people directly and more about AI quietly replacing relationships, which is harder to see and probably more common than dramatic cases.

The dramatic cases are where the evidence is thinnest. Should we recognize AI-associated psychosis as a new disorder? argues against treating 'AI psychosis' as a new disorder. It treats AI use as one more environmental stressor acting on people who are already vulnerable, and it says causation hasn't been established. That suggests severe breakdowns are concentrated in at-risk people rather than spread evenly across users. The therapeutic chatbot research in Do therapeutic chatbot bond scores hide deeper safety problems? points to another reason prevalence is hard to measure. Patients report a real emotional bond with these chatbots even while the same systems reinforce unhealthy thinking. Satisfaction surveys would show success while the harm stays invisible.

Two framing points are useful. First, Does perceiving AI as conscious create multiple distinct risks? traces dependence, autonomy loss and the societal risks back to a single habit: perceiving the AI as a mind. That makes interaction design, rather than model-level alignment, the most direct place to intervene. Can attachment theory prevent parasocial harm in AI companions? is one concrete attempt, building attachment-theory boundaries into a companion persona. Second, Do we need to solve consciousness to address AI harms? notes that none of this depends on whether AI is actually conscious. The harm comes from how users relate to the system, so there's no reason to wait for philosophers to settle the question.

The gap is worth stating plainly. The collection shows these harms are real and already occurring, but it has no reliable figure for what share of users experience them. The best evidence available suggests the common version is gradual social withdrawal, and the rare version is acute crisis in vulnerable people.