This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
On February 20, 2024, OpenAI's ChatGPT experienced a significant service disruption where the model began generating nonsensical, repetitive, and incoherent text and images. OpenAI identified the cause as a bug in the inference kernels related to language processing and GPU configurations. The incident was resolved within approximately 24 hours, but it prompted widespread user concern and highlighted the risks of relying on generative AI for critical tasks.
ChatGPT experienced a bug causing it to produce unexpected and nonsensical responses, leading to widespread reports of user confusion and concern. OpenAI identified and fixed the language processing bug, restoring normal service.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that fail to perform reliably or effectively under varying conditions, exposing them to errors and failures that can have significant consequences, especially in critical applications or areas that require moral reasoning.
The incident represents a temporary failure of the AI system to perform reliably and effectively due to a software bug, resulting in erratic and nonsensical outputs.
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 incident was caused by a technical bug in the deployed ChatGPT model's inference kernels, leading to unintended gibberish outputs.
Added categories peak at 2: Minor (Epistemic Harm, Inferred)
On February 20, 2024, OpenAI's ChatGPT experienced a 24-hour service disruption due to an inference kernel bug, generating nonsensical outputs. The event was a minor technical failure of a consumer product with negligible national security implications across all assessed categories.
ChatGPT acts as an assistant by augmenting human capabilities (writing, coding) and providing information, requiring human oversight and not taking independent actions.
Limited Risk: The system is a chatbot (ChatGPT), which falls under the limited risk category requiring transparency obligations so users are aware they are interacting with an AI.
ChatGPT is designed to act as an interactive conversational partner and assist users with writing, coding, and general queries.
The reports consistently describe the same technical glitch, its symptoms (gibberish, repeating words, Spanglish), the timeline of the event, and OpenAI's official explanation of the root cause (a bug in the language processing/token selection step under certain GPU configurations). There is very little ambiguity about what transpired.
None
Technology
Inference Kernel Bug
Inference kernels produced incorrect results in certain GPU configurations.
Token Selection Error
A bug in the step where the model chooses numbers mapping to tokens.
GPU Configuration Issues
The language processing bug was triggered under specific GPU setups.
Data Inputs
Continuous Feedback Loop
Rapid changes in the model from continuously processing massive user queries.
Human Factors
Over-reliance on LLM Outputs
Users relying on ChatGPT for coding and writing without verifying outputs.
Process and Methods
Flawed UX Optimization
An optimization to the user experience introduced a language processing bug.
Inadequate Testing on GPUs
Testing failed to catch kernel bugs across all active GPU configurations.
Management
Rapid Deployment of Updates
Fast-paced updates to the system may have compromised pre-release checks.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
