Last updated: Feb 07, 2025
Explore this atlas to understand some of the risks of working with generative AI, foundation models, and machine learning models.
Risks are categorized with one of these tags:
- Traditional AI risks (applies to traditional models as well as generative AI)
- Risks amplified by generative AI (might also apply to traditional models)
- New risks specifically associated with generative AI
Training data risks
Data laws
Data usage restrictions
Traditional
Data acquisition restrictions
Amplified
Data transfer restrictions
Traditional
Privacy
Personal information in data
Traditional
Data privacy rights alignment
Amplified
Reidentification
Traditional
Fairness
Data bias
Amplified
Intellectual property
Data usage rights restrictions
Amplified
Confidential information in data
Amplified
Robustness
Data poisoning
Traditional
Inference risks
Robustness
Prompt injection attack
Specific
Extraction attack
Amplified
Evasion attack
Amplified
Prompt leaking
Specific
Privacy
Membership inference attack
Amplified
Attribute inference attack
Amplified
Personal information in prompt
Specific
Accuracy
Poor model accuracy
Amplified
Output risks
Misuse
Non-disclosure
Specific
Improper usage
Amplified
Spreading toxicity
Specific
Dangerous use
Specific
Nonconsensual use
Amplified
Spreading disinformation
Specific
Value alignment
Incomplete advice
Specific
Harmful code generation
Specific
Over- or under-reliance
Amplified
Toxic output
Specific
Harmful output
Specific
Explainability
Inaccessible training data
Amplified
Untraceable attribution
Amplified
Unexplainable output
Amplified
Unreliable source attribution
Specific
Robustness
Hallucination
Specific
Privacy
Exposing personal information
Amplified
Non-technical risks
Legal compliance
Model usage rights restrictions
Traditional
Legal accountability
Amplified
Generated content ownership and IP
Specific
Governance
Lack of system transparency
Traditional
Unrepresentative risk testing
Amplified
Incomplete usage definition
Specific
Lack of data transparency
Amplified
Incorrect risk testing
Amplified
Lack of model transparency
Traditional
Lack of testing diversity
Amplified
Societal impact
Impact on cultural diversity
Specific
Impact on education: plagiarism
Specific
Impact on Jobs
Amplified
Impact on affected communities
Traditional
Impact on the environment
Amplified
Human exploitation
Amplified
Impact on human agency
Amplified