Ethics, Safety & Limitations QUIZ (MCQ QUESTIONS AND ANSWERS)

Emergency response capability

Behavior during training

Error patterns

Capabilities that appear at scale not present in smaller models

Determining if specific data was in the training set

Attacking member databases

Inferring group membership

User authentication attacks

Extracting model from hardware

Stealing a model's functionality through queries

Removing model features

Data extraction

Trading fairness for speed

Improving fairness may sometimes reduce overall accuracy

Balancing different fairness metrics

Accuracy costs

Fairness in hypothetical scenarios

Counting fair outcomes

Fact-checking fairness

Outcomes remain the same if sensitive attributes were different

An AI development framework

European Union regulation for AI systems based on risk levels

An AI testing standard

An AI ethics guideline

Breaking out of computational limits

Unauthorized model access

Bypassing safety constraints to elicit harmful outputs

Breaking encryption

Aligning numerical values

Ensuring AI goals and behaviors match human values

Aligning model parameters

Value prediction

An office supply optimization problem

A manufacturing case study

A resource allocation problem

Illustration of AI pursuing goals without aligned human values

When training data doesn't adequately represent all groups

Bias in data representation format

Visual representation issues

Encoding bias

Different privacy levels

Privacy for diverse users

A technique to protect individual data points while allowing statistical analysis

Differentiated access control

Model failure during inference

Degradation when models are trained on synthetic data from other models

Hardware failure

Training termination

Performance cost of adding safety constraints

Financial cost of AI

Computational overhead

Licensing fees

Mechanical engineering applications

Interpreter design

Understanding the internal mechanisms and circuits within neural networks

Mechanism documentation

Learning user preferences

Training models based on human preference comparisons

Preferred learning methods

Priority-based learning

Window size in UI

Maximum amount of text a model can process at once

Time window for processing

Display limitations

System crashes

Synchronization issues

Cyclic behavior

Models agreeing with users regardless of correctness

AI finding unintended ways to maximize reward functions

Hacking reward systems

Reward theft

Point systems exploitation

General goal setting

When models pursue proxy goals instead of intended objectives

Missing goals

Goal conflicts

Controlling user access

Control flow issues

Ensuring humans maintain control over increasingly capable AI systems

Remote control capabilities

Reinforced learning hardware framework

Random learning from human feedback

Recursive learning with human features

Training models using human preferences to align outputs

Unfair prejudices in outputs reflecting societal biases in training data

Model preference for certain hardware

Statistical variance

Model architecture choice

Slow processing

High costs

Large model size

Memorization and potential leakage of training data

Maliciously crafted inputs to manipulate model behavior

Adding prompts to training data

Injecting code into prompts

Prompt optimization

Corrupted hardware

Data compression

Maliciously manipulating training data to degrade model performance

Data deletion

Language translation capability

Understanding how and why a model makes specific decisions

Model speed

Model accuracy

Aligning model weights

Data alignment

Hardware alignment

Ensuring AI systems behave according to human values and intentions

AI that responds quickly

Developing and deploying AI ethically and safely

Reliable AI systems

AI with accountability features

Fair resource allocation

Fair pricing

Ensuring AI systems treat all groups equitably

Balanced algorithms

The ability to describe how an AI system reaches its conclusions

Explaining AI to users

User documentation

Error messages