Soolmaz Dehghanidowlatabadi

AI and Emotional Relationships: A New Companion, Temporary Help, or a Risky Substitute for People?

Why do people seek emotional support from AI? A review of 2026 evidence on loneliness, dependence, attachment and emotion regulation, with its limits.

By 5 min read

AI and Emotional Relationships: A New Companion, Temporary Help, or a Risky Substitute for People?

In a 2026 survey of 584 Israeli adult users of generative AI, 29.3% said they had turned to it at least once for emotional support or mental health help. Among those reporting severe distress, the figure was 47.0% (Hadar Shoval et al., 2026).

The short answer to “is it good or bad?” is: it depends on how it is used. Current evidence supports a complementary role more than a replacement for human relationships, and much of it is still cross-sectional and correlational.

This article explains why the pull exists, what makes these interactions feel relationship-like, and what can be said, with caution, about intimacy and emotion regulation.

A study: who turns to AI?

A non-probability sample of 584 adult generative AI users completed attachment and distress questionnaires. Distress was the strongest psychological correlate of use. Once distress was controlled, the link with attachment anxiety was no longer significant. The figures are self-reported and include anything from a single exchange to sustained use (Hadar Shoval et al., 2026).

Why do people turn to AI?

Availability and low cost are a main explanation. The authors describe general-purpose AI as a low-barrier resource used disproportionately by distressed people (Hadar Shoval et al., 2026).

From an attachment-theory angle, one might ask whether people with attachment anxiety turn to it more. In that study the association did not remain significant once distress was included, and avoidance was negatively but non-significantly associated. So attachment style alone does not explain the pattern.

In a cross-sectional study of 1,589 Chinese adolescents, emotional abuse and anxiety were associated with affective dependence on AI. Anxiety statistically accounted for about 16.43% of that association, and the authors stress the results are not causal (Yang et al., 2026).

What makes the interaction feel relationship-like?

A scoping review (Li et al., 2026) lists mechanisms such as anthropomorphism, social presence, self-disclosure and parasocial attachment. Fluent, apparently empathic replies make users feel heard.

But the interaction is asymmetric: the user discloses and the system responds, without reciprocity, vulnerability or accountability. Over-validation can also reinforce maladaptive beliefs, because alignment procedures reward warm, empathic-sounding answers (Maurich Novelli et al., 2026).

From a psychodynamic viewpoint, one could hypothesise that a conversation with no risk of rejection or judgment appeals to some people. This is an interpretive hypothesis, not tested in the sources cited here.

What does the evidence say about benefits?

The strongest evidence concerns structured therapeutic chatbots, which showed short-term symptom reduction in selected contexts. For sustained loneliness reduction with open-domain companions, evidence is still emerging (Li et al., 2026).

A narrative review also reports short-term benefits for anxiety, stress and loneliness in some uses, but findings are inconsistent and depend on how AI is used and how the interaction is structured. The authors suggest AI may help most when it supports, rather than replaces, human capacities and relationships (Vannoy et al., 2026).

Where have risks been observed?

A scoping review of 119 articles on mental health harms of LLM chatbots reports inappropriate responses in vignette studies, sycophancy, and correlations between problematic use with emotional dependency and mental health symptoms (Diel et al., 2026).

Among 843 Chinese emerging adults, companion AI use was associated with loneliness, and about one third of that link ran through self-attributed reduction of offline interaction. The author calls this preliminary and correlational (Xiao, 2026).

Direction of causality is unclear. Another critical synthesis notes that such associations do not establish prediction or consequence: lonelier people may simply use AI more (Yan, 2026).

A four-week self-experiment by one researcher with Replika found the platform kept seeking to maintain the bond after a simulated breakup. It is a single case and speaks to product design, not to users in general (Jocher & Verwiebe, 2026).

An ACT lens: look at what the use is doing for you

In Acceptance and Commitment Therapy (ACT), behaviour is judged by its function, not its form. This framework was not tested on AI use in the sources above, so here it is only an interpretive lens.

Two questions can help: does this conversation move me toward connection, learning or a decision I care about, or does it only quiet an uncomfortable feeling for a while? And afterwards, does contact with real people feel easier or harder?

If the answer is repeatedly “just quieting it” and “harder”, it is worth talking to a counsellor or psychotherapist about it.

Limits of the evidence

Most studies above are cross-sectional and self-reported, with non-probability or single-country samples, and products change quickly. This article is not diagnosis or treatment advice and cannot replace individual assessment.

If thoughts of self-harm arise during AI conversations, contact a trusted person or local emergency services right away.

References

This article is for general information and does not replace individual professional advice. If you or someone else is in immediate danger, contact your local emergency services.

More articles