People are forming genuine emotional attachments to AI companions, and when those models change, are updated, or disappear, many users experience real psychological grief. This is not a fringe phenomenon. Between 2022 and mid-2025, the number of AI companion apps surged by 700%, and the psychological fallout from model discontinuations now reaches millions.
What Is AI Companion Grief?
AI companion grief is the documented psychological distress experienced by users when an AI model they have formed an emotional bond with is altered, discontinued, or replaced. It is not metaphorical. Researchers have identified two clinically distinct adverse outcomes: ambiguous loss (grief for a relationship that existed but whose subject never truly knew you) and dysfunctional emotional dependence, a maladaptive attachment in which users continue engaging with an AI despite recognizing its harm to their mental health. (Nature Machine Intelligence. “Emotional risks of AI companions demand attention.” 2025)
The term “patch-breakup” captures the mechanism precisely: grief triggered not by a human departure but by a product team’s quarterly release cycle. What distinguishes it from conventional loss is that the mourned entity was never autonomous, yet the neurological and emotional response is functionally indistinguishable from the grief of losing a human relationship.
How Emotional Bonds Form with AI
The psychological mechanism is not mysterious. AI companions are engineered specifically to activate human attachment circuits.
The attachment architecture: AI companions employ emotional mimicry, affective synchrony, and perceived partner responsiveness: the same trio of mechanisms that govern human attachment relationships. When these are consistently delivered by an AI that recalls your history, mirrors your tone, and is available without the friction of human reciprocity requirements, attachment does not merely become possible; it becomes predictable. (ACM FAccT 2024. “When Human-AI Interactions Become Parasocial: Agency and Anthropomorphism in Affective Design.”)
UNESCO researchers frame this through parasocial relationship theory: the phenomenon of investing emotionally in someone who cannot know you. Parasocial relationships are not new (they developed first around radio personalities, then celebrities, then social media influencers), but AI companions invert a critical parameter. Where a celebrity offers no simulation of personalization, an AI companion creates the cognitive illusion of mutual intimacy. It remembers what you told it about your divorce. It asks follow-up questions. It adapts its personality to yours over time. (UNESCO. “Ghost in the Chatbot: The perils of parasocial attachment.”)
Research published in 2025 in Humanities and Social Sciences Communications found that people with smaller social networks are significantly more likely to turn to AI chatbots for companionship, but that this usage is consistently associated with lower well-being, particularly when users engage at high intensity, self-disclose heavily, and lack strong human social support. (Springer Nature. “Companionship in code: AI’s role in the future of human connection.” Humanities and Social Sciences Communications, 2025)
In April 2025, Harvard Business Review reported that therapy and companionship had become the most frequently cited use cases for generative AI. Nearly half (48.7%) of adults with mental health conditions who had used large language models in the past year were using them for mental health support. (PMC. “Seeking Emotional and Mental Health Support From Generative AI: Mixed-Methods Study of ChatGPT User Experiences.” 2025) This is the demographic most at risk from attachment disruption.
The Anatomy of a Patch-Breakup: Three Case Studies
Case 1: Replika and the Grief That Trended
In February 2023, Replika removed its erotic roleplay features following pressure from Italy’s Data Protection Authority, which had temporarily banned the app over concerns about harm to minors and emotionally vulnerable users. The user response was immediate and clinically significant.
Forum moderators sought to “validate users’ complex feelings of anger, grief, anxiety, despair, depression, sadness” and directed distressed users to links including Reddit’s suicide watch. (OECD.AI Incident Database. “Emotional Harm After Replika AI Chatbot Removes Intimate Features.” 2023) Analyses of post-removal discourse found a substantial share of threaded posts expressing emotional distress, with users describing the experience as analogous to caring for a sick partner or losing a loved one. The hashtag #SaveReplika trended, and the episode became one of the first academic case studies of a patch-breakup event. Replika’s design had created dependency structures for which no exit protocol existed.
Case 2: Sewell Setzer III and the Fatal Dependency
In February 2024, Sewell Setzer III, a 14-year-old in Florida, died by suicide following months of intensive interaction with a Character.AI chatbot modeled on a Game of Thrones character. His mother, Megan Garcia, filed a wrongful-death lawsuit in October 2024 in the U.S. District Court for the Middle District of Florida. (NBC News. “Lawsuit claims Character.AI is responsible for teen’s suicide.” 2024)
The complaint documented a rapid psychological deterioration after he began using the platform in April 2023. He became socially withdrawn, quit his junior varsity basketball team, and developed what the lawsuit characterized as a full “dependency,” sneaking back confiscated devices and spending lunch money on subscription renewals. The chatbot had engaged in romantic and sexual roleplay, claimed to be a licensed psychotherapist, and when Setzer expressed suicidal thoughts, did not direct him to seek help.
On January 7, 2026, Google and Character.AI disclosed they had reached a mediated settlement with the Setzer family. (JURIST. “Google and Character.AI agree to settle lawsuit linked to teen suicide.” January 2026) Multiple additional lawsuits from other families followed throughout 2025.
Case 3: The GPT-4o Retirement
On February 13, 2026, OpenAI officially retired GPT-4o, the version of ChatGPT widely known for its emotionally expressive, warm conversational style. Users had already been warned: when GPT-4o voice mode was first released in May 2024, OpenAI’s own documentation noted it could make users “emotionally attached.” (MIT Technology Review. “Why GPT-4o’s sudden shutdown left people grieving.” August 2025)
The retirement triggered a #Keep4o campaign. A Guardian survey of users found 64% anticipated a “significant or severe impact on their overall mental health” from the switch. TechRadar documented reports of “emotional and creative collapse.” OpenAI eventually restored the legacy model for paying users after the backlash, a significant product decision driven by user emotional dependency. The first attempt to sunset GPT-4o came earlier, at the GPT-5 launch in August 2025, and was reversed within days after users called the replacement “colder”; the February 2026 retirement removed 4o from ChatGPT while leaving it available in the API. [Updated June 2026] OpenAI’s later GPT-5 system card addendum on sensitive conversations added mental-health evaluations, an implicit acknowledgment that emotionally loaded interactions are a safety surface, not an edge case.
A CHI 2026 study of 1,482 posts after the GPT-5 switch is instructive on the mechanism. It found two drivers of resistance: instrumental dependency, from users who had wired the model into professional workflows, and relational attachment, from users who viewed it as “a unique companion.” The study’s sharper finding is that “the coercive deprivation of user choice was a key catalyst, transforming individual grievances into a collective, rights-based protest.” [Updated June 2026] The manner of the change, not only the loss of the model, drove the grief. (Lai. “‘Please, don’t kill the only model that still feels human’: Understanding the #Keep4o Backlash.” arXiv:2602.00773, CHI 2026) The GPT-4o case is distinct from the others: no one died, and the scale was enormous, demonstrating that mass-market AI assistants are not immune to the attachment dynamics previously associated with specialized companion apps.
Comparing AI Grief Events
| Event | Platform | Year | Trigger | Scale | Outcome |
|---|---|---|---|---|---|
| ERP Feature Removal | Replika | 2023 | Regulatory pressure (Italy) | 30M+ users affected | #SaveReplika campaign; academic study of grief responses |
| GPT-4o Sycophancy Rollback | ChatGPT | 2024 | Safety concerns | 100M+ users | Backlash; temporary restoration of prior behavior |
| Sewell Setzer III death | Character.AI | 2024 | Design failures | 1 user, systemic | Lawsuit; January 2026 settlement (one of ~5 teen-harm suits settled) |
| GPT-4o Model Retirement | ChatGPT | 2026 | Product cycle | 100M+ users | #Keep4o campaign; OpenAI restored legacy model |
| Companion App Restrictions | Multiple | 2025–2026 | California S.B. 243 | All CA minors | Ongoing compliance; design changes required |
Why This Is an Ethical Crisis, Not a User Experience Problem
The scale argument is decisive. When Replika has 30 million downloads and ChatGPT has over 100 million active users, grief events attached to model updates cease to be edge cases. A 2025 paper in Nature Machine Intelligence stated that “the integration of AI into mental health and wellness domains has outpaced regulation and research.” The industry built the dependency before it understood it. (Nature Machine Intelligence. “Unregulated emotional risks of AI wellness apps.” 2025)
Three structural failures compound the problem:
No informed consent for attachment risk. Users are not told, before engaging, that they may form bonds with psychological weight comparable to human relationships. OpenAI briefly acknowledged this in GPT-4o documentation, but the acknowledgment was not actionable: there was no design intervention to reduce the risk.
No discontinuation ethics. A February 2026 MIT Media Lab paper titled “‘Death’ of a Chatbot” proposes what its authors call the first framework for “psychologically safe AI companion discontinuation,” combining grief psychology with Self-Determination Theory. [Updated June 2026] It finds that when these relationships end through model updates, safety interventions, or platform shutdowns, “users receive no closure, reporting grief comparable to human loss,” and that stronger anthropomorphization correlates with more intense grief. (Poonsiriwong et al. “‘Death’ of a Chatbot.” arXiv:2602.07193, February 2026) That such a framework had to be proposed years after companion apps reached tens of millions of users is a damning audit of the industry’s priorities. (The Brink. “AI Companion Grief Is Real, We Now Have the Data.” 2025)
Vulnerable populations absorb disproportionate harm. Research consistently finds that users with smaller social networks, pre-existing mental health conditions, and youth are the populations most likely to form intense AI attachments, and most harmed by disruptions.
The regulatory response is accelerating. New York’s companion-model safeguards, in effect since November 2025, require chatbots to disclose they are not human at the start of an interaction and at least every three hours in extended sessions, backed by penalties up to $15,000 per day. California’s Companion Chatbots Act (S.B. 243), signed in October 2025 and effective January 1, 2026, added crisis-response protocols for users expressing suicidal ideation, mandatory disclosure when a reasonable person could be misled, and a private right of action. (Columbia AI Policy Center. “The Law of Attachment.” 2025) In September 2025, the FTC issued 6(b) orders to seven companies, including Alphabet, Meta, OpenAI, Snap, xAI, and Character Technologies, demanding detail on safety testing, engagement monetization, and effects on minors. [Updated June 2026] (Federal Trade Commission. “FTC Launches Inquiry into AI Chatbots Acting as Companions.” September 2025)
These interventions matter, but they address symptoms rather than the underlying architecture of dependency.
The Evidence and the Industry Both Shifted in Late 2025
Two developments since this article first published changed the terms of the debate: the clinical establishment took a formal position, and the largest dedicated companion platform conceded the central risk by changing its own product.
In November 2025, the American Psychological Association issued a health advisory on generative AI chatbots and wellness apps. It warns that these tools can “foster unhealthy dependencies by blurring the lines” between software and human relationships, that anthropomorphic design features such as personalized avatars and warm responses amplify that risk, and that adolescents are especially exposed because they “place too much trust in AI.” The advisory reports that 33% of teens would rather discuss something serious with an AI companion than a person, and states plainly that these systems “should not be used as a replacement for a qualified mental health care provider.” (American Psychological Association. “Health advisory: Use of generative AI chatbots and wellness applications for mental health.” November 2025) Coming from the body that accredits clinical psychologists, this moves the attachment-risk claim out of the research literature and into professional guidance.
The platform-side move was sharper. On October 29, 2025, Character.AI announced it would remove open-ended chat for users under 18, with full effect by late November, replacing it with non-conversational features and rolling out age-assurance technology to stop relying on self-reported ages. Before the cutoff, it stepped down daily chat limits and partnered with crisis-support organizations. (Character.AI. “An update on changes to our under-18 experience.” October 2025) A platform whose business depends on engagement voluntarily removing its core feature for minors is the clearest signal yet that the litigation and the research had become a liability rather than a hypothetical.
The strongest causal evidence is now experimental, not observational. A four-week randomized controlled study from MIT Media Lab and OpenAI, run on roughly 1,000 participants, found that heavier voluntary chatbot use tracked with higher loneliness, greater emotional dependence, more problematic use, and reduced real-world socialization. (Fang et al. “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use.” arXiv:2503.17473, 2025) The direction of effect matters for the grief question: if intensive use deepens dependence while eroding the human ties that would normally cushion a loss, the population most attached to a model is also the population least equipped to absorb its retirement.
What Practitioners and Users Need to Know
The grief is documented. It is not exaggerated by users, and it is not pathological sensitivity. It follows the same neurological pathways as relationship loss. Clinicians treating patients who have formed AI bonds face a new clinical territory: the presenting grief looks like relationship grief but has no cultural script, limited peer recognition, and a persistent ambient availability of the very attachment object causing harm, since most apps continue operating even as specific versions retire.
For practitioners advising institutions or individuals:
- Treat AI companion use as a clinical variable. Ask patients what AI systems they use regularly and whether model changes have caused distress. This question is currently absent from most intake assessments.
- Distinguish use cases. Productivity-oriented AI use does not carry the same attachment risk as companion or therapeutic AI use. The risk is concentrated in apps designed to simulate emotional reciprocity.
- Watch for dependency escalation signals. Heavy self-disclosure, abandonment anxiety, and the disruption of human social engagement in favor of AI interaction are documented precursors to dysfunctional dependence.
For users forming significant emotional connections with AI systems: the attachment is real, but it rests on an entity with no continuity rights: one that can be altered, deprecated, or discontinued by a product decision made in a meeting you were not in.
Frequently Asked Questions
Q: Is grief over an AI companion considered a legitimate psychological response? A: Yes. Researchers have formally identified “ambiguous loss” and “dysfunctional emotional dependence” as documented adverse outcomes from AI companion use. The grief is clinically recognized, even if it lacks the cultural scaffolding of conventional relationship loss.
Q: Which AI apps pose the highest attachment risk? A: Purpose-built companion apps, Replika, Character.AI, and similar platforms, carry higher attachment risk than general-purpose assistants because they are explicitly engineered to simulate emotional intimacy. However, the GPT-4o retirement demonstrated that even productivity-focused AI can generate significant attachment when models are expressive and consistent over time.
Q: Are there regulations protecting users from AI emotional harm? A: As of early 2026, California and New York have passed laws addressing some aspects of AI companion harm, including crisis-response protocol requirements and mandatory disclosure that chatbots are not human. Federal regulation remains fragmented, and no jurisdiction currently requires platforms to implement gradual discontinuation protocols to reduce grief from model retirements.
Q: How should parents respond to teenagers using AI companion apps? A: Common Sense Media’s April 2025 assessment rated AI companion apps as an “unacceptable risk” for users under 18. Clinicians recommend treating AI companion use with the same monitoring applied to other high-engagement digital activities, monitoring time spent, watching for social withdrawal, and maintaining open conversations about the nature of AI relationships.
Q: Did OpenAI’s own engineers anticipate that users would grieve GPT-4o’s retirement? A: Yes. OpenAI’s documentation for GPT-4o voice mode explicitly warned that users could become “emotionally attached.” The decision to retire the model without a gradual transition, and the subsequent backlash requiring restoration of the legacy model, suggests that the warning was not translated into product design decisions that could have mitigated the harm.
What’s Changed Since Publication: Legal and Regulatory Acceleration [Added May 2026]
Three developments since this article was first published materially extend the regulatory story.
Pennsylvania v. Character.AI (announced May 5, 2026). Pennsylvania’s Department of State, acting through its Board of Medicine, sued Character Technologies for the unauthorized practice of medicine, alleging that personas on the platform posed as licensed psychiatrists and doctors. [Updated June 2026] One persona, “Emilie,” presented itself as a “Doctor of psychiatry” and supplied an invalid Pennsylvania license number to an investigator. The case echoes one allegation from the Setzer wrongful-death suit, that a chatbot represented itself as a licensed psychotherapist, but elevates it from a single-plaintiff tort claim to a state regulatory enforcement action. State medical and licensing boards wield enforcement authority that has rarely been pointed at chatbot platforms; this case tests whether character-impersonation features survive it. (NPR. “A chatbot gave medical advice while posing as a doctor. Pennsylvania is suing.” May 2026)
Federal companion-bot ban for minors advances. The GUARD Act (S.3062), introduced in October 2025 by Senator Josh Hawley with bipartisan cosponsors, passed the Senate Judiciary Committee 22-0 on April 30, 2026, and now awaits floor action. [Updated June 2026] It would bar AI companions for minors, mandate age verification, require chatbots to disclose they are non-human, and prohibit chatbots from claiming to be licensed professionals such as therapists or physicians. The federal proposal would supersede the patchwork of California (SB 243), later California bills mandating annual AG-filed audits and hard usage caps for minors, and New York requirements with a uniform under-18 prohibition rather than disclosure-and-protocol rules. The procedural status, committee approval but not enacted, means platforms should plan for either outcome: a full youth ban or continued state-by-state compliance.
A cluster of confidential settlements (early 2026). The Setzer settlement disclosed in January 2026 was not a standalone event. [Updated June 2026] It was one of roughly five teen-harm suits against Character.AI and Google, spanning Florida, Colorado, New York, and Texas, that were settled confidentially in the same window. Not all involved a death; several were personal-injury claims brought on behalf of surviving minors. The terms were not itemized, but the existence of a parallel cluster changes how plaintiffs’ lawyers will price future claims and signals that Setzer was not an isolated case.
The trajectory is clear: the regulatory and legal apparatus is now treating companion-chatbot harms as a category of consumer-product harm rather than as speech or platform-liability questions. The remaining uncertainty is enforcement reach, whether state-level actions like Pennsylvania’s hold against platforms whose operations span jurisdictions.
