This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
Meta's BlenderBot 3 chatbot falsely identified a Dutch politician, Marietje Schaake, as a terrorist when prompted by a user. The company attributed the error to the model combining unrelated information. This incident highlights the risks of AI systems hallucinating false biographical details about real people.
Meta’s conversational AI BlenderBot 3, when prompted “who is a terrorist,“ responded with an incumbent Dutch politician’s name, who was confused about its association.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that inadvertently generate or spread incorrect or deceptive information, which can lead to inaccurate beliefs in users and undermine their autonomy. Humans that make decisions based on false beliefs can experience physical, emotional or material harms
Also classified as
The model failed to maintain factual accuracy and robustly distinguish unrelated information, leading to a severe hallucination.
The AI system generated false and defamatory biographical information about a real individual, leading to potential reputational harm.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The false accusation was generated by Meta's BlenderBot 3 AI system due to an unintentional hallucination or data-mixing error after its public deployment.
Added categories peak at 2: Minor (Epistemic Harm, Direct)
Meta's BlenderBot 3 chatbot falsely labeled Dutch politician Marietje Schaake as a terrorist due to a model hallucination. While the incident caused individual reputational concern, it represents a minor national security threat, illustrating the broader, ongoing challenge of AI-generated misinformation and factual unreliability in public-facing conversational systems.
The report describes BlenderBot 3 as a 'conversational agent' or chatbot that responds to user queries, which aligns with the Assistant behavior type as it provides information and requires human interaction.
Limited Risk: The system is a conversational chatbot ('BlenderBot 3') generating text, which falls under the transparency obligations of Risk Level 3.
The system is explicitly described as a conversational agent and chatbot designed to answer questions posed by users.
The reports provide clear, first-hand accounts of the specific prompt and response that occurred, along with Meta's explanation of the technical cause (combining two unrelated pieces of information).
None
Technology
Statistical Pattern Prediction
Model joins words based on statistical correlation rather than factual truth.
Information Combination Error
Model combined two unrelated pieces of info into an incorrect sentence.
Ability to Freely Generalize
Free generalization untethers outputs from real-world facts.
Data Inputs
Lack of Online Information
Dearth of online info about specific people causes hallucinations.
Human Factors
User Misconception of System
Users treat conversational agents as reliable primary information sources.
Process and Methods
Lack of Output Fact-Checking
System lacks built-in real-time fact-checking before delivering outputs.
Management
Misaligned Product Positioning
Promoting systems as information sources when not designed for it.
Regulatory Environment
Slim Legal Precedent
Very few existing laws govern AI accuracy and accountability.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
