AWA AIF-C01 - Domain 4 - Guidelines for Responsible AI

Developing and using AI in ways that promote fairness, transparency, accountability, safety, privacy, and appropriate human oversight.

The principle of ensuring AI systems do not produce unjustified or inappropriate differences in outcomes for individuals or groups.

A systematic tendency in data, models, or processes that can lead to unfair or inaccurate outcomes.

Bias introduced when training or evaluation data is incomplete, unrepresentative, imbalanced, or otherwise systematically skewed.

Systematic patterns in model behavior that can cause certain groups, inputs, or outcomes to be treated differently or inaccurately.

The degree to which a dataset appropriately reflects the population or situations the AI system is intended to serve.

The degree to which data is accurate, complete, relevant, consistent, timely, and appropriate for its intended purpose.

Making information about an AI system, its purpose, capabilities, limitations, and use sufficiently understandable to relevant stakeholders.

The ability to provide understandable reasons or explanations for a model's predictions or outputs.

The degree to which humans can understand how a model or its components contribute to its outputs.

The assignment of responsibility for AI system decisions, development, deployment, monitoring, and outcomes.

Human involvement in reviewing, approving, monitoring, or intervening in AI system decisions or actions.

A system design in which humans actively review or approve AI outputs or decisions before consequential actions occur.

A system design in which humans monitor an AI system and can intervene when necessary rather than reviewing every individual decision.

A governance approach where humans retain ultimate authority over an AI system and its use.

The ability of an AI system to continue performing reliably when inputs vary, conditions change, or unexpected situations occur.

The ability of an AI system to maintain or recover appropriate operation when failures, disruptions, or unexpected events occur.

Designing and operating AI systems to reduce the likelihood and impact of harmful or unintended outcomes.

Protecting personal or sensitive information from inappropriate collection, use, disclosure, or exposure.

Collecting and using only the data necessary for a defined purpose rather than unnecessarily gathering additional information.

Information that can identify an individual directly or indirectly, such as a name, email address, or identification number.

Data requiring additional protection because unauthorized access or disclosure could cause harm, such as financial, health, or confidential information.

Harmful, offensive, abusive, or otherwise inappropriate content produced by or supplied to an AI system.

Detecting, filtering, blocking, or otherwise managing inappropriate, harmful, or prohibited content.

An attempt to manipulate an AI system by placing malicious or conflicting instructions in input data or prompts.

Attempts to bypass an AI model's safety controls or behavioral restrictions through specially crafted prompts or inputs.

A model output that appears plausible but contains inaccurate, fabricated, or unsupported information.

Providing a model with reliable external information so responses are based on relevant, verifiable sources rather than solely on learned model knowledge.

Controls that constrain AI system behavior, inputs, or outputs to help enforce safety, security, privacy, and organizational requirements.

An Amazon Bedrock capability for applying configurable safeguards to model inputs and outputs, including controls for topics and harmful content.

Continuously observing a deployed AI/ML system for changes in performance, data, behavior, or other indicators of problems.

A change in the relationship between input data and the target outcome over time that can cause model performance to degrade.

A change over time in the statistical characteristics or distribution of input data used by a model.

The process of identifying systematic disparities or problematic patterns in data, model behavior, or outcomes.

Techniques used to reduce unwanted bias in training data, models, predictions, or AI system processes.

An AWS capability that helps detect potential bias in datasets and models and provides tools for explaining model predictions.

Documentation describing important information about a model, such as intended use, limitations, evaluation results, and potential risks.

Documentation describing a dataset's characteristics, motivation, composition, collection process, and intended uses.

The ability to examine and trace AI system activities, decisions, data, and processes for review or investigation.

Ongoing evaluation, monitoring, feedback, and refinement used to improve an AI system's performance, safety, fairness, and reliability over time.