Description: OpenAI published a post about its findings when using Universe, a software for measuring and training AI agents to conduct reinforcement learning experiments, showing that the AI agent did not act in the way intended to complete a videogame.
Entités
Voir toutes les entitésClassifications de taxonomie CSETv0
Détails de la taxonomieProblem Nature
Indicates which, if any, of the following types of AI failure describe the incident: "Specification," i.e. the system's behavior did not align with the true intentions of its designer, operator, etc; "Robustness," i.e. the system operated unsafely because of features or changes in its environment, or in the inputs the system received; "Assurance," i.e. the system could not be adequately monitored or controlled during operation.
Specification
Physical System
Where relevant, indicates whether the AI system(s) was embedded into or tightly associated with specific types of hardware.
Software only
Level of Autonomy
The degree to which the AI system(s) functions independently from human intervention. "High" means there is no human involved in the system action execution; "Medium" means the system generates a decision and a human oversees the resulting action; "low" means the system generates decision-support output and a human makes a decision and executes an action.
Unclear/unknown
Nature of End User
"Expert" if users with special training or technical expertise were the ones meant to benefit from the AI system(s)’ operation; "Amateur" if the AI systems were primarily meant to benefit the general public or untrained users.
Expert
Public Sector Deployment
"Yes" if the AI system(s) involved in the accident were being used by the public sector or for the administration of public goods (for example, public transportation). "No" if the system(s) were being used in the private sector or for commercial purposes (for example, a ride-sharing company), on the other.
No
Data Inputs
A brief description of the data that the AI system(s) used or were trained on.
Universe software training
Classifications de taxonomie CSETv1
Détails de la taxonomieIncident Number
The number of the incident in the AI Incident Database.
65
Special Interest Intangible Harm
An assessment of whether a special interest intangible harm occurred. This assessment does not consider the context of the intangible harm, if an AI was involved, or if there is characterizable class or subgroup of harmed entities. It is also not assessing if an intangible harm occurred. It is only asking if a special interest intangible harm occurred.
no
Date of Incident Year
The year in which the incident occurred. If there are multiple harms or occurrences of the incident, list the earliest. If a precise date is unavailable, but the available sources provide a basis for estimating the year, estimate. Otherwise, leave blank.
Enter in the format of YYYY
2016
Date of Incident Month
The month in which the incident occurred. If there are multiple harms or occurrences of the incident, list the earliest. If a precise date is unavailable, but the available sources provide a basis for estimating the month, estimate. Otherwise, leave blank.
Enter in the format of MM
12
Date of Incident Day
The day on which the incident occurred. If a precise date is unavailable, leave blank.
Enter in the format of DD
21
Estimated Date
“Yes” if the data was estimated. “No” otherwise.
Yes
Risk Subdomain
A further 23 subdomains create an accessible and understandable classification of hazards and harms associated with AI
7.1. AI pursuing its own goals in conflict with human goals or values
Risk Domain
The Domain Taxonomy of AI Risks classifies risks into seven AI risk domains: (1) Discrimination & toxicity, (2) Privacy & security, (3) Misinformation, (4) Malicious actors & misuse, (5) Human-computer interaction, (6) Socioeconomic & environmental harms, and (7) AI system safety, failures & limitations.
- AI system safety, failures, and limitations
Entity
Which, if any, entity is presented as the main cause of the risk
AI
Timing
The stage in the AI lifecycle at which the risk is presented as occurring
Post-deployment
Intent
Whether the risk is presented as occurring as an expected or unexpected outcome from pursuing a goal
Unintentional
Rapports d'incidents
Chronologie du rapport
Chez OpenAI, nous avons récemment commencé à utiliser Universe, notre logiciel de mesure et de formation d'agents IA, pour mener de nouvelles expériences RL. Parfois, ces expériences illustrent certains des problèmes de RL tels qu'ils sont …
Variantes
Une "Variante" est un incident qui partage les mêmes facteurs de causalité, produit des dommages similaires et implique les mêmes systèmes intelligents qu'un incident d'IA connu. Plutôt que d'indexer les variantes comme des incidents entièrement distincts, nous listons les variations d'incidents sous le premier incident similaire soumis à la base de données. Contrairement aux autres types de soumission à la base de données des incidents, les variantes ne sont pas tenues d'avoir des rapports en preuve externes à la base de données des incidents. En savoir plus sur le document de recherche.
Incidents similaires
Did our AI mess up? Flag the unrelated incidents
Incidents similaires
Did our AI mess up? Flag the unrelated incidents