AI system
- System
-
ChatGPT (OpenAI), with OpenAI's automated detection and human review pipeline for users planning to harm others
- Application developer
-
OpenAI
- Configuration
-
Not assessed
- Workflow
-
OpenAI reportedly disclosed threatening messages to the FBI. Local police compared them with messages that the threatened person had preserved, investigated, and brought charges. A guilty plea led to court restrictions. [1] [2] [3] [4]
- Human oversight
-
Not assessed
- Relation to AI
-
The reporting identifies an OpenAI disclosure as the route that brought Zhou’s messages to police attention. OpenAI’s general policy describes detection followed by human review before possible referral to law enforcement. That policy does not document the exact classifier, trigger, reviewer decisions, or ChatGPT replies in this case. [1] [2] [3] [4]
- AI contribution
-
Reporting identifies OpenAI’s disclosure as the route through which police learned of the threats. OpenAI’s general policy describes automated detection followed by human review, but the sources do not identify the classifier or review process used here. [1] [2] [3] [4]
Benefits
- Outcome identifiers
-
Not assessed
- Benefit
-
Zhou was arrested on 29 May 2026 and pleaded guilty on 13 August. [1] [2] [3] [4]
- Category
-
Reduced Violence
- Beneficiaries
-
Not assessed
- Distribution and access
-
Not assessed
- Costs and harms
-
Not assessed
Measurement
- Study design
-
Not assessed
- Setting and sample
-
Not assessed
- Measures
-
arrest on May 29, 2026; guilty plea on August 13
- Results
-
arrest on May 29, 2026; guilty plea on August 13
- Additional value from AI
-
Not assessed
- Denominators
-
Not assessed
- Materials for replication
-
Not assessed
Money
- Kind of value
-
Not assessed
- Currency and prices
-
Not assessed
- Period and unit
-
Not assessed
- Recipients and payers
-
Not assessed
- Gross or net
-
Not assessed
- Included costs
-
Not assessed
- Return and payback
-
Not assessed
People and places
- Participants
-
Developer:
OpenAI. Operator:
OpenAI.
- Ownership and control
-
Not assessed
- Dependencies
-
Not assessed
- Cross-border connections
-
Not assessed
- Geopolitical implications
-
Not assessed
Follow-up
- Conditions for replication
-
A disclosure must reach an agency able to investigate. Investigators must establish identity and assess the threat using supporting evidence. A general detection-and-review policy does not establish the details of a particular referral. Arrest and court restrictions do not prove that an attack would otherwise have occurred. [1] [2] [3] [4]
- Replication attempts
-
Not assessed
- Later findings
-
13 August 2026. Zhou pleaded guilty on 13 August 2026 to written threats to kill, aggravated stalking and misuse of a two-way device, with adjudication withheld and eight years of probation. [1] [2] [3] [4]
- Availability
-
Not assessed
- Institutional decision
-
Not assessed
- Action and result
-
Not assessed
- Success criterion
-
Not assessed
Evidence
- Original evidence
-
The accessible reporting describes the disclosure, arrest, and guilty plea. The affidavit, chats, emergency disclosure, and court order were inaccessible, so the account does not independently establish their full contents. [1] [2] [3] [4]
- Access and versions
-
BocaNewsNow’s report and OpenAI’s policy article are available. The police blotter and court portal rejected access, so the underlying filings were not inspected. [1] [2] [3] [4]
- Supported findings
-
Not assessed
- Evidence links
-
Not assessed
- Evidence strength
-
Not assessed
- Missing evidence
-
Not assessed
Grades
- Reasons
-
The referral led to investigation, arrest, and restrictions on contact and weapons. These measures protected the threatened person, although the sources cannot establish whether an attack would otherwise have occurred. [1] [2] [3] [4]
- Observed or intended
-
Not assessed
- Assessor and date
-
Not assessed
- Alternative assessments
-
Not assessed
- Fill
-
Independently reported
Entry details
- Related entries
-
Not assessed
- Field
-
Public safety - stalking and threat investigation. Consumer chatbot trust-and-safety referral to law enforcement.
- Purpose
-
OpenAI says that when it detects users planning to harm others, it routes the conversations to a small team trained on its usage policies. It says that if human reviewers find an imminent threat of serious physical harm to others, it may refer the case to law enforcement. This referral is that designed use. ChatGPT itself is a general-purpose assistant, which Zhou used to rehearse his plans.
- Dates
-
Observed: 29 May 2026. Updated: 24 October 2026. [1] [2] [3] [4]
- Version
-
SHA-256: 1305768f24ac…
References
- Other identifiers
-
oecd_aim: 2026-08-15-e71a.
- Category mappings
-
Not assessed
- Revisions
-
6 revisions. Latest: 24 October 2026. [1] [2] [3] [4]
Collections: Scrutiny and findings, Safeguards, Security and threat-watching.