Guidelines for Responsible AI · Topic 1 of 6
Features of Responsible AI
Learn the core features of responsible AI, including fairness, safety, veracity, robustness, privacy and governance, and how model choice affects energy.
- 4 min
- Core dimensions include fairness, explainability, privacy and security, transparency, veracity and robustness, governance, safety and controllability.
- Bias is systematic unfairness, and fairness is equitable treatment across groups.
- Robustness means reliable behaviour on unusual or adversarial input.
What it is
Responsible AI means designing, building and using AI so that it is fair, safe, transparent and accountable. It is about reducing harm to people while keeping the benefits of the technology.
AWS describes responsible AI through a set of core dimensions. The exam expects you to recognise these and match each one to a typical problem or tool.
Responsible AI applies at every step of this chain, not only at the end.
How it works
The core dimensions
| Dimension | Plain meaning |
|---|---|
| Fairness | The system treats different groups of people equitably |
| Explainability | People can understand why the system gave an output |
| Privacy and security | Personal and sensitive data is protected |
| Transparency | People are told how and when AI is used, and what its limits are |
| Veracity and robustness | Outputs are truthful and correct, and the system works under unusual input |
| Governance | Clear rules, roles and accountability for AI |
| Safety | The system avoids harm to users and society |
| Controllability | People can steer, correct or stop the system |
Features the exam names directly
The outline also lists these features by name.
- Bias: systematic unfairness in outputs, often caused by skewed data.
- Fairness: the goal of equitable treatment across groups.
- Inclusivity: involving and serving diverse people and needs.
- Robustness: reliable results even with noisy, unexpected or adversarial input.
- Safety: avoiding harmful behaviour or content.
- Veracity: truthful, accurate output, which hallucination threatens.
Tools that support these features
| Need | AWS tool |
|---|---|
| Filter harmful content and block topics | Amazon Bedrock Guardrails |
| Detect bias and explain predictions | Amazon SageMaker Clarify |
| Watch model quality and drift in production | Amazon SageMaker Model Monitor |
| Human review of low confidence results | Amazon Augmented AI (Amazon A2I) |
| Document a model's purpose and limits | Amazon SageMaker Model Cards |
| Learn about a service's intended use and limits | AWS AI Service Cards |
Each tool addresses a different dimension, so a full programme uses several together.
Responsible model selection
Choosing a model is also a responsible AI decision. Consider these points.
| Consideration | Why it matters |
|---|---|
| Environmental impact | Training and running large models uses energy |
| Sustainability | A smaller or more efficient model can meet the need with less resource use |
| Fit for purpose | A model that is too large wastes resources, and one that is too weak may give harmful errors |
| Documentation | A model with clear information on training data and limits is easier to trust |
| Licence and data origin | Check you are allowed to use the model and its data for your purpose |
Using the smallest model that meets the need reduces cost and environmental impact at the same time. Reusing a pre trained model avoids the large energy cost of training from scratch.
Trade offs
Responsible AI features sometimes pull against each other or against performance. A more explainable model may be less accurate, and strict safety filters can block some valid requests. The right balance depends on the risk of the use case, with high stakes cases needing more control.
People and process
Tools alone are not enough. Teams also need diverse reviewers, clear policies, testing before launch, and a way for users to report problems.
What to remember for the exam
- Core dimensions include fairness, explainability, privacy and security, transparency, veracity and robustness, governance, safety and controllability.
- Bias is systematic unfairness, and fairness is equitable treatment across groups.
- Robustness means reliable behaviour on unusual or adversarial input.
- Veracity means truthful output, which hallucination puts at risk.
- Amazon Bedrock Guardrails filter harmful content.
- Clarify detects bias and explains predictions, Model Monitor watches live quality, and A2I adds human review.
- Model Cards and AI Service Cards document intended use and limits.
- Smaller, efficient or reused models support sustainability goals.
Match the problem to the tool. Harmful content is Guardrails, bias detection is Clarify, drift over time is Model Monitor, and human review is A2I.
Common mistakes
Treating responsible AI as a final check. It needs attention from design through monitoring.
Confusing fairness with accuracy. A model can be accurate overall and still treat one group poorly.
Thinking one tool covers everything. Each AWS tool handles a different part.
Ignoring environmental cost in model choice. A larger model than needed adds energy use without benefit.
Assuming safety filters remove all risk. They reduce it, and human oversight is still needed.
Check yourself
Name three core dimensions of responsible AI.
Examples are fairness, explainability, privacy and security, transparency, veracity and robustness, governance, safety and controllability.
Which AWS tool filters harmful content in a Bedrock application?
Amazon Bedrock Guardrails.
How can model choice support sustainability?
By using the smallest or an already pre trained model that meets the need, which uses less energy.
Topic 1 of 6 in Guidelines for Responsible AI
Test yourself on Guidelines for Responsible AI
The full AWS AIF-C01 question bank has 849 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by AWS. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.