Last updated: Dec 12, 2024
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
Risks associated with input
Training and tuning phase
Robustness
Data poisoning
Traditional
Intellectual property
Confidential information in data
Amplified
Data usage rights restrictions
Amplified
Privacy
Data privacy rights alignment
Amplified
Personal information in data
Traditional
Reidentification
Traditional
Data laws
Data usage restrictions
Traditional
Data acquisition restrictions
Amplified
Data transfer restrictions
Traditional
Fairness
Data bias
Amplified
Inference phase
Privacy
Personal information in prompt
Specific
Attribute inference attack
Amplified
Membership inference attack
Amplified
Robustness
Prompt leaking
Specific
Prompt injection attack
Specific
Extraction attack
Amplified
Evasion attack
Amplified
Accuracy
Poor model accuracy
Amplified
Risks associated with output
Value alignment
Toxic output
Specific
Harmful output
Specific
Incomplete advice
Specific
Over- or Under-reliance
Amplified
Explainability
Unreliable source attribution
Specific
Unexplainable output
Amplified
Inaccessible training data
Amplified
Untraceable attribution
Amplified
Privacy
Exposing personal information
Amplified
Misuse
Nonconsensual use
Amplified
Non-disclosure
Specific
Improper usage
Amplified
Dangerous use
Specific
Spreading disinformation
Specific
Spreading toxicity
Specific
Robustness
Hallucination
Specific
Harmful code generation
Harmful code generation
Specific
Non-technical risks
Governance
Lack of model transparency
Traditional
Lack of testing diversity
Amplified
Lack of system transparency
Traditional
Unrepresentative risk testing
Amplified
Incomplete usage definition
Specific
Lack of data transparency
Amplified
Incorrect risk testing
Amplified
Societal impact
Impact on human agency
Amplified
Impact on cultural diversity
Specific
Impact on education: plagiarism
Specific
Impact on Jobs
Amplified
Human exploitation
Amplified
Impact on affected communities
Traditional
Impact on the environment
Amplified
Legal compliance
Legal accountability
Amplified
Model usage rights restrictions
Traditional
Generated content ownership and IP
Specific