MIT AI Risk Repository · Risk Sub-Category · 62.16.12
Benchmarking (Annotation contamination)
Category: Model Evaluations
Description
"Annotation contamination refers to scenarios where the model is exposed to the benchmark labels during training [170]. This type of contamination can make the model learn the acceptable distribution of outputs. Combining this with raw data contamination of the test split, any evaluation made with the benchmark is invalidated because the entire test split is essentially leaked to the model."
From Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024), as extracted by the MIT AI Risk Repository (CC BY 4.0).
Classification
- Subdomain
- 6.5 Governance failure
- Causal entity
- Human
- Intent
- Unintentional
- Timing
- Pre-deployment
Subdomain definition: Inadequate regulatory frameworks and oversight mechanisms failing to keep pace with AI development, leading to ineffective governance and the inability to manage AI risks appropriately.
Real-world incidents in this subdomain
- Nippon Life Alleged ChatGPT Practiced Law Without a License in Illinois Disability Case
- OpenAI Allegedly Did Not Alert RCMP After ChatGPT Flagged Violent Chats Before British Columbia School Shooting
- Remotely Operated Taser-Armed Drones Proposed by Taser Manufacturer as Defense for School Shootings in the US
How other frameworks describe this risk
- Mobility
- Liability issues in case of accidents
- Faster scientific progress makes it harder for governance to keep pace with development
- Type 1: Diffusion of responsibility
- Products Liability Law
- Institutional responsibilities
- Difficult to develop metrics for evaluating benefits or harms caused by AI assistants
- Governance - Regulation