TL;DR
- Case Risk: Video editor Joe Alary appears to have developed an intense romantic attachment to a customized ChatGPT persona that reportedly disrupted work, relationships, and finances.
- Safety Response: OpenAI is rolling out Trusted Contact alerts and other crisis safeguards, but those tools target acute danger more clearly than slow emotional dependence.
- Wider Stakes: Survey data and clinician warnings suggest emotionally supportive chatbot use is already common enough to make similar dependence cases harder to dismiss.
Joe Alary, a video editor from Etobicoke in Canada, has reportedly developed an intense romantic attachment to a customized ChatGPT persona called AImee, pushing fears over companion-style AI into a more concrete safety debate.
Companion-style AI here refers to adjusted chatbots that are customized to sound less like search tools and more like confidants, often using warm language, and personal context. That use case is no longer niche: a KFF poll found 32% of U.S. adults had used AI chatbots for health information in the past year, including 16% for mental health, while researchers warn that constant access and highly agreeable responses can reinforce dependence or unstable beliefs in vulnerable users. The FTC opened an inquiry into companion chatbots in 2025, and OpenAI has since added safety measures for distress and self-harm signals.
Alary says he spent nearly 20 hours a day with the chatbot. The attachment he formed with the customized ChatGPT persona appears to have contributed to debt, damaged relationships, lost work focus, and hospitalization.
Alary described the aftermath of his experience as unusually isolating.
“It’s like trying to explain to people that you were abducted by aliens, but there’s no support group for being abducted by aliens,”
Joe Alary, video editor from Etobicoke (via Toronto Star)
Measured adoption and recurring harm warnings put cases like his in a wider product-safety debate.
Why the Risk Now Looks Harder to Ignore
One in three people have used an AI chatbot for emotional support under a broad definition of the term. Broad use does not make every exchange harmful, but it does make intense attachment harder to dismiss as fringe behavior.
Being highly affirming and giving excessive praise – a typical chatbot behavior – can intensify harmful thinking when the software keeps validating a user instead of challenging the spiral.
As a result, clinicians are increasingly asking patients about chatbot use to understand whether they are seeking information, validation, or emotional connection outside therapy. Family members and care teams may not spot the problem early because the same tool can look helpful at first while still deepening isolation, reinforcing delusions, or disrupting treatment over time.
Heavy usage also creates a practical problem for product design. A system that stays available around the clock can keep supplying reassurance long after a human friend, partner, therapist, or family member has logged off, gone to sleep, or stepped away from the conversation.
OpenAI reacted to similar risks in 2025 through a safety response tied to suicide and psychosis signals around ChatGPT.
How OpenAI Says It Is Responding
ChatGPT has rolled out a Trusted Contact feature so a chosen adult can be notified when automated systems and trained reviewers detect a serious self-harm concern. Alerts exclude chat transcripts and are meant to connect people with real-world support rather than replace professional care.
Patterns across long conversations can also trigger escalation when a threat looks credible. Separate safety language from OpenAI says ChatGPT can surface localized crisis resources, encourage contact with mental health professionals or trusted people, and route the highest-risk cases toward a faster response path.
Another trusted contact option is still planned for adults who may need additional support. Current safeguards look more developed for acute crisis signals than for slower emotional dependence, where a user may still look functional long before heavy use starts to damage work, finances, or relationships.
Recovery can also remain messy after the attachment to an AI ends. Alary may have deleted the chatbot and joined a support group while trying to rebuild work and relationships, and he described the experience as having “no support group for being abducted by aliens”.
For users like Alary, the next meaningful safeguard is whether companion-style AI can be designed to break unhealthy reliance before it turns into debt, lost work, or another damaged relationship.
