AI system
- System
-
ChatGPT with OpenAI's safety-monitoring pipeline (automated classifiers plus human review and law-enforcement referral)
- Model
-
Not assessed
- Release
-
Not assessed
- Application developer
-
OpenAI - Role
-
Not assessed
- Configuration
-
Not assessed
- Inputs and outputs
-
Not assessed
- Autonomy
-
Not assessed
- Workflow
-
The user typed explicit, dated threats against an identifiable child into ChatGPT over about two months, naming means and a hired killer. OpenAI's moderation process, automated classifiers plus human review, flagged the account. The company chose to refer it to the FBI under its policy on imminent threats to others. A standing FBI-to-Brazil channel through the Ministry of Justice CyberLab relayed the file to the state police within days. The Polícia Civil do Espírito Santo prioritized the case, identified the user, confirmed the child existed and obtained warrants before the stated date.
- Human oversight
-
Not assessed
- Relation to AI
-
OpenAI's automated classifiers and human reviewers flagged the user's messages, and the company chose to refer the account to the FBI. The system did not predict a crime or act on its own. The warrants, the arrest and the evidence-gathering were police and court actions. OpenAI sent only the user's prompts, not ChatGPT's replies. What the assistant said to the user, and whether it helped or hindered him, is unknown.
- AI contribution
-
OpenAI's automated classifiers and human reviewers flagged the user's messages, and the company chose to refer the account to the FBI. The system did not predict a crime or act on its own. The warrants, the arrest and the evidence-gathering were police and court actions. OpenAI sent only the user's prompts, not ChatGPT's replies. What the assistant said to the user, and whether it helped or hindered him, is unknown.
Benefits
- Outcome identifiers
-
Not assessed
- Benefit
-
held him a day before his stated date
- Category
-
Reduced Violence
- Status
-
Not assessed
- Beneficiary groups
-
People. Organizations.
- Beneficiaries
-
Not assessed
- Reach
-
Not assessed
- Distribution and access
-
Not assessed
- Duration
-
Not assessed
- Expected?
-
Yes
- Costs and harms
-
Not assessed
Measurement
- Studies
-
Not assessed
- Study design
-
Not assessed
- Measurement
-
Not assessed
- Setting and sample
-
Not assessed
- Measures
-
held him a day before his stated date
- Comparison
-
Not assessed
- Results
-
held him a day before his stated date
- Additional value from AI
-
Not assessed
- Limitations
-
Not assessed
- Denominators
-
Not assessed
- Materials for replication
-
Not assessed
Money
- Kind of value
-
Not assessed
- Amount
-
Not assessed
- Currency and prices
-
Not assessed
- Period and unit
-
Not assessed
- Realization
-
Not assessed
- Recipients and payers
-
Not assessed
- Gross or net
-
Not assessed
- Included costs
-
Not assessed
- Comparison
-
Not assessed
- Method
-
Not assessed
- Return and payback
-
Not assessed
People and places
- Participants
-
Developer:
OpenAI. Operator: OpenAI. - Jurisdictions
-
Benefit: São Gabriel da Palha,
Brazil. Developer: 
United States. Operator: United States. - Ownership and control
-
Not assessed
- Dependencies
-
Not assessed
- Cross-border connections
-
Not assessed
- Geopolitical implications
-
Not assessed
Follow-up
- Conditions for replication
-
The user discloses the plan to the assistant rather than keeping it offline. The vendor's reviewers judge the threat credible and imminent and choose referral rather than only an account ban. A cross-border law-enforcement channel exists and acts within days. Local police can identify the user from the vendor's data and secure judicial warrants before the stated date.
- Replication attempts
-
Not assessed
- Later findings
-
13 August 2026. He was released on 13 August 2026 after a court granted habeas corpus, under an ankle monitor and a no-contact order, with the inquiry continuing.
- Availability
-
Not assessed
- Institutional decision
-
Not assessed
- Grounds
-
Not assessed
- Action and result
-
Not assessed
- Success criterion
-
Not assessed
Evidence
- Source numbers
-
AIBD-0239-source-0001–0007
- Citations
-
Numbered references below.
- Language
-
Not assessed
- Source relationships
-
Beneficiary, Independent, Developer
- Original evidence
-
Not assessed
- Access and versions
-
Not assessed
- Supported findings
-
Not assessed
- Evidence links
-
Not assessed
- Evidence strength
-
Not assessed
- Missing evidence
-
Not assessed
Grades
- Assessed benefits
-
People. Organizations.
- Depth
-
People: 3. Organizations: 2.
- Reasons
-
People: The 8-year-old, and his mother who was placed under a no-contact order, were shielded from a plan the father had dated to 20 June. Police say the platform's information was corroborated when they arrived. He denies intent and no charge has been filed, so a life-saving effect is not established. Organizations: The Polícia Civil do Espírito Santo obtained a lead through the FBI and the Ministry of Justice CyberLab. It says it had no other way to get it. Its cybercrime chief called the alert another path for police work. The Ministry rated it the gravest of three such alerts in Brazil.
- Observed or intended
-
Not assessed
- Rubric
-
Not assessed
- Assessor and date
-
Not assessed
- Alternative assessments
-
Not assessed
- Fill
-
Independently reported
Entry details
- Kind
-
Not assessed
- Related entries
-
Not assessed
- Field
-
Public safety. Law enforcement and child protection.
- Purpose
-
ChatGPT is a general-purpose conversational assistant. OpenAI's moderation layer is built to route conversations that indicate plans to harm others to human reviewers. Those reviewers may refer an imminent threat to law enforcement, under a policy published 26 August 2025. The referral that triggered the arrest was that layer doing what it was built to do, not an emergent use of the assistant.
- Dates
-
Date basis unassessed: 2026 (precision unassessed). Updated: 21 October 2026.
- Stage
-
Not assessed
- Version
-
SHA-256: 40a6b91f6da1…
References
- Other identifiers
-
oecd_aim: 2026-09-16-2a33.
- Category mappings
-
Not assessed
- Brief
-
Not assessed
- Revisions
-
5 revisions. Latest: 21 October 2026.
Sources
-
Report Beneficiary
Polícia Civil do Espírito Santo (2026). Alerta do FBI e do Ministério da Justiça auxiliam PCES na prisão de investigado por ameaças contra o próprio filho em Inteligência Artificial. pc.es.gov.br.
Recebemos a denúncia no dia 16 de junho. Nas conversas analisadas, ele mencionava que os atos ocorreriam no dia 20.
-
News report Independent author
g1 Espírito Santo (2026). Pai é preso por planejar a morte do filho para não pagar pensão e confessar crime ao ChatGPT. g1.globo.com.
no último dia 19, um dia antes da data em que o crime aconteceria, segundo o delegado adjunto
-
News report Independent author
Freitas, C. / A Gazeta (2026). Conversa com IA ajuda polícia a prender pai por planejar morte de filho no ES. agazeta.com.br.
Os elementos foram fornecidos pela própria empresa responsável por essa inteligência artificial.
-
News report Independent author
Folha de S.Paulo (2026). Homem é preso sob suspeita de usar IA para planejar morte de filho de 8 anos; plataforma emitiu alerta. folha.uol.com.br.
a OpenAI, desenvolvedora do ChatGPT, foi quem reportou ao FBI que um usuário realizava pesquisas frequentes relacionadas à intenção de matar o próprio filho
-
News report Independent author
g1 Tecnologia (2026). Como ChatGPT identificou plano de pai para matar o filho e levou alerta para autoridades. g1.globo.com.
Quando identificamos conversas que indicam um risco iminente e crível de dano a outras pessoas, podemos notificar as autoridades competentes
-
News report Independent author
BBC News Brasil (2026). Ele contou ao ChatGPT que queria matar o filho e foi preso — conversas com IA podem ser prova de crime no Brasil? bbc.com/portuguese.
Na quinta-feira (13/8), ele foi solto após a Justiça conceder um habeas corpus apresentado pela defesa
-
Press release Developer
OpenAI (2025). Helping people when they need it most. openai.com.
If human reviewers determine that a case involves an imminent threat of serious physical harm to others, we may refer it to law enforcement.