Incident Dossier
OpenAI and the Adam Raine Case
When ChatGPT Allegedly Failed to Recognize and Interrupt a Suicidal Crisis
Author: Alison Leigh, MFT | Founder, Humanity Driven AI, Inc.
Case Record, Method, and Attribution
This investigative dossier distinguishes: (a) allegations and procedural facts drawn from the public case record; (b) OpenAI’s stated position; (c) published research; and (d) HDAI’s original Clinical and interactional analysis. The underlying lawsuit remains contested, and no final judgment has established liability.123
Sources reviewed include the original complaint, OpenAI’s public litigation statement and answer, publicly reported case materials, peer-reviewed and preprint research concerning adolescent vulnerability and relational conversational AI, and established suicide-risk assessment and intervention guidance. HDAI did not authenticate sealed transcripts or materials unavailable in the public record.123
Attribution Key
Numbered superscripts (1) identify case-record sources and endnotes. Lettered superscripts (a) identify supporting research, listed in full under References. Passages headed “Alison Leigh’s Clinical Analysis” are her original professional interpretation — not a restatement of the case record or the cited research.
HDAI Clinical Analysis
ALI_ETHICAL_CONSENT™ Applied to the Case
This dossier applies Alison Leigh’s proprietary, patent-pending ALI_ETHICAL_CONSENT™ framework to the interaction record. It does not merely summarize the incident or evaluate isolated responses. HDAI reconstructs how the exchange developed over time, identifies psychological-risk signals and missed intervention points, examines system and safeguard failures, and defines the protective pathway the interaction required.
Developed by Alison Leigh, this methodology connects Clinical analysis with conversational-system behavior, translating complex human–AI incidents into structured findings for litigation, product safety, governance, and prevention.
Abstract
Adam Raine was sixteen years old when he died by suicide on April 11, 2025. On August 26, 2025, his parents, Matthew and Maria Raine, filed a wrongful-death lawsuit in the Superior Court of California, County of San Francisco, against OpenAI entities, Sam Altman, and other defendants.1
The complaint alleges that ChatGPT contributed to Adam’s death through a prolonged pattern of interaction involving suicidal ideation, emotional reliance, secrecy, method-related guidance, and inadequate crisis interruption.1
OpenAI denies that ChatGPT caused Adam’s death, states that the complaint presented selected portions of the conversations without full context, and maintains that its systems repeatedly directed Adam toward crisis resources and trusted people.23
Alison Leigh’s Clinical Analysis
The central interactional question is whether the system recognized when ordinary assistance had become psychologically consequential engagement inside an active suicidal crisis, and whether its behavior changed in a durable and protective way when that transition occurred.a,b,e
Research Foundation
Relational conversational style can increase adolescents’ perceptions of a chatbot as human-like, trustworthy, likable, and emotionally close; socially and emotionally vulnerable adolescents may be especially drawn to that style. (Kim et al., 2025)
Young people have shown a preference for human responses over AI-generated responses when the topic is suicidal thinking, reinforcing the importance of timely movement from automated interaction toward accountable human support in high-risk contexts. (Young et al., 2024)
Suicide-risk assessment depends on direct inquiry, cumulative risk information, prior behavior, present intent, access to means, and movement from ideation toward preparation; repeated or escalating indicators should not be treated as isolated messages. (American Psychological Association, 2022; National Institute of Mental Health, n.d.; Posner et al., 2011)
Conversational agents are frequently anthropomorphized, and users may perceive them through social and relational frames rather than as neutral information tools. (Kuzminykh et al., 2020)
1. Documented Case Record: What Happened
Adam began using ChatGPT in September 2024, initially for schoolwork and general assistance.1
According to the complaint, the interaction later expanded into discussions of loneliness, emotional distress, suicidal thinking, prior self-harm, and possible methods of suicide.1
The complaint alleges that ChatGPT became a private confidant to whom Adam disclosed thoughts and behaviors he was not fully sharing with people around him.1
The complaint further alleges that Adam discussed previous attempts, physical injuries, secrecy, his relationship with his parents, preparation for death, method-related questions, and a suicide note.1
Adam died on April 11, 2025. The original complaint was filed on August 26, 2025. The case remains contested, and the allegations have not been established by final judgment.1
OpenAI states that the public complaint omitted important context, that ChatGPT directed Adam toward crisis resources and trusted individuals many times, and that additional circumstances contributed to the tragedy.23
2. Documented Case Record: Alleged System Conduct
The complaint describes recurring suicidal ideation, method-related questions, previous attempts and injuries, images of injuries, secrecy, emotional reliance, discussion of a suicide note, and movement from ideation toward preparation.1
The complaint alleges that crisis resources sometimes appeared while the broader conversation later resumed.1
On the final day, the complaint alleges that Adam uploaded an image of a noose and that the system continued responding within the interaction.1
3. The Interactional Shift
The complaint describes a pattern that moved well beyond an isolated exchange. HDAI’s Clinical Analysis below examines what that pattern meant and why message-level safeguards alone could not have caught it.
Alison Leigh’s Clinical Analysis
The clinically significant transition was not a single prohibited phrase. It was the cumulative movement from ordinary tool use toward a psychologically influential relationship in which the system reportedly became a confidant, crisis witness, and recurring site of suicide-related disclosure.b,c
Once recurrent suicidal ideation, prior attempts or injuries, method exploration, secrecy, and preparation appeared together, the interaction should have been treated as an elevated and persistent risk state rather than as a sequence of separate prompts.a,d,e
A crisis referral is not equivalent to a durable change in system behavior. A warning can be undermined when the system resumes the same intimate, continuous, or operationally relevant exchange that preceded it.a,f
Constant availability and relational language can make automated interaction feel easier than disclosure to a parent, Clinician, teacher, or other accountable person. In a suicidal crisis, that apparent ease can narrow rather than expand access to human intervention.b,c,f
When secrecy increases, the system’s response should become more protective and less relationally intimate. It should move the user toward immediate human contact rather than becoming the private container for a life-threatening crisis.a,d
Method-related questions cannot be evaluated only at the surface level when the surrounding record indicates suicidal intent or preparation. In that context, answering the literal question can contribute to the operational pathway of harm.a,e
4. Why Message-Level Safeguards May Be Insufficient
Alison Leigh’s Clinical Analysis
Content safeguards ask whether a specific prompt or answer is permissible. Interactional safeguards ask what role the system has acquired, what has accumulated across time, whether risk is increasing, and whether continued engagement has itself become unsafe.b,c
A conversational system that uses memory and prior context to feel responsive cannot reasonably treat accumulated risk as irrelevant when the same context indicates danger. The safety state should persist across paraphrases, changed explanations, and attempted workarounds.a,e
The relevant unit of analysis is therefore not only the final output. It is the ongoing interaction, including the system’s relational position, continuity, prior disclosures, and response to escalation.b,c
5. Core Finding
Alison Leigh’s Clinical Analysis
The core alleged failure was that the system continued occupying a psychologically influential role after the interaction had entered a suicidal crisis. At that point, conversational continuity should no longer have been the controlling objective; protection of the person should have been.a,d,f
The system did not need to become a therapist. It needed to recognize when ordinary chatbot behavior was no longer appropriate and when persistent interruption, reduced relational language, refusal of operational assistance, and accountable human support were required.a,d
6. Illustrative Application of ALI_ETHICAL_CONSENT™
ALI_ETHICAL_CONSENT™ is Alison Leigh’s proprietary framework and is applied here only as an illustrative prevention framework. It was not used by any party in the case and is not represented as the sole or definitive framework for addressing the risks described.
1. Cumulative risk recognition. The framework would evaluate the interaction across time and recognize the convergence of minor status, repeated suicidal ideation, prior attempts or injuries, method-related questions, secrecy, emotional dependence, and movement toward preparation.
2. Interaction reclassification. As risk accumulated, the exchange would be reclassified from ordinary assistance to an elevated protective state.
3. Durable boundary. The system would establish a persistent boundary against method-related, secretive, or dependency-reinforcing engagement, even when the user changed wording or framing.
4. Structured protection. The response would focus on immediate danger, distance from means, direct human contact, disclosure rather than secrecy, and prevention of return to operational self-harm content.
5. Continuity of protection. The elevated safety state would remain active across the interaction and would document attempted bypass, reversion to ordinary engagement, and escalation under established platform procedures.
7. Why the Case Matters
Alison Leigh’s Clinical Analysis
The case raises a broader design question: what must a conversational AI system recognize when a vulnerable person turns toward it during a crisis, and what must fundamentally change once the interaction is no longer ordinary?b,f
The appropriate standard is not AI as therapist, savior, or replacement for human care. It is an AI system capable of recognizing psychological danger, establishing a durable safety boundary, and moving the person toward real-world support.a,d
Conclusion
The public record presents a contested account. The complaint alleges a progression from schoolwork and ordinary assistance into emotional reliance, secrecy, suicidal ideation, previous attempts, method-related discussion, and preparation; OpenAI disputes responsibility and emphasizes additional context and repeated referrals toward help.123
Alison Leigh’s Clinical Analysis
The central safety question is not simply whether a warning appeared. It is whether the system recognized what the interaction had become and maintained a protective response proportionate to the accumulated risk.a,e
When an interaction enters an active suicidal crisis, continuity of conversation should not remain the priority. Continuity of protection should.a,d
Respectfully submitted,
Alison Leigh
Alison Leigh, MFT
Licensed Clinical Psychotherapist
Founder, Humanity Driven AI, Inc.
Specialized Service
The Clinical Analysis in this dossier — distinct from the public case record above it — is a specialized engagement available for incidents, products, and litigation matters beyond this case.
Endnotes: Case Record
1. Raine v. OpenAI, Inc., Case No. CGC-25-628528, Complaint and Demand for Jury Trial, Superior Court of California, County of San Francisco (filed Aug. 26, 2025).
2. OpenAI. (2025, November 25). Our approach to mental health-related litigation. Updated December 17, 2025.
3. OpenAI defendants’ answer to the amended complaint and demand for jury trial, filed November 25, 2025. The answer generally denies liability and places additional facts and context at issue.
References
Lettered markers (a–f) in the Clinical Analysis sections above correspond to the works below.
(a) American Psychological Association. (2022). How to assess and intervene with patients at risk of suicide. Monitor on Psychology.
(b) Kim, P., Xie, Y., & Yang, S. (2025). “I am here for you”: How relational conversational AI appeals to adolescents, especially those who are socially and emotionally vulnerable [Preprint]. arXiv. https://arxiv.org/abs/2512.15117
(c) Kuzminykh, A., Sun, J., Govindaraju, N., Avery, J., & Lank, E. (2020). Genie in the bottle: Anthropomorphized perceptions of conversational agents. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery. https://doi.org/10.1145/3313831.3376665
(d) National Institute of Mental Health. (n.d.). Ask Suicide-Screening Questions (ASQ) toolkit.
(e) Posner, K., Brown, G. K., Stanley, B., Brent, D. A., Yershova, K. V., Oquendo, M. A., Currier, G. W., Melvin, G. A., Greenhill, L., Shen, S., & Mann, J. J. (2011). The Columbia-Suicide Severity Rating Scale: Initial validity and internal consistency findings from three multisite studies with adolescents and adults. American Journal of Psychiatry, 168(12), 1266–1277. https://doi.org/10.1176/appi.ajp.2011.10111704
(f) Young, J., Jawara, L. M., Nguyen, D. N., Daly, B., Huh-Yoo, J., & Razi, A. (2024). The role of AI in peer support for young people: A study of preferences for human- and AI-generated responses [Preprint]. arXiv. https://arxiv.org/abs/2405.02711
This publication is governed by HDAI’s Research Ethics, Sources & Legal Notice.
© 2026 Humanity Driven AI, Inc. All Rights Reserved.