Overview
Trustworthy AI is the smallest domain, with four objectives that all begin with describe. You should be able to describe the ethical principles of trustworthy AI, the balance between data privacy and the importance of consent, how NVIDIA and other technologies improve trustworthiness, and how to minimize bias in AI systems.
NVIDIA frames trustworthy AI around privacy, safety and security, transparency and nondiscrimination, and the readings for this section include NVIDIA's own trustworthy AI material and its explainer on retrieval-augmented generation. For LLMs, the practical topics are hallucination and grounding, harmful or off-topic output, jailbreaks and prompt injection, personal data in training corpora, consent for data use, and unequal behavior across groups. Tools include NeMo Guardrails for programmable rails, safety classifier models offered as NIM microservices, PII redaction during curation and documentation such as model cards.
How it's tested
Expect concise scenarios about a deployed assistant producing unsafe, biased or private output, or a team planning to use data collected from users. You choose the principle at stake, the control that addresses it, or the NVIDIA tool built for it. Answers favor layered, documented safeguards over single fixes.
-
Name the principle
A harm or concern is described and you identify the trustworthy AI principle it violates, such as privacy or transparency.
-
Choose the safeguard
An LLM risk such as jailbreaks, off-topic answers or leaked personal data is described, and you pick the control or NVIDIA tool that mitigates it.
-
Measure the bias
A model seems to treat groups differently and you pick the evaluation or mitigation step that confirms or reduces it.
A bank's assistant sometimes answers questions about competitors' products and occasionally follows instructions hidden in user messages. Retraining the model is slow and does not guarantee control. Programmable rails can check each input and output, keep conversations on approved topics and block jailbreak patterns. That describes NeMo Guardrails, which sits around the model rather than changing its weights.
Objectives 4
Official objectives from the NVIDIA outline, in outline order.
5.1 Describe the ethical principles of trustworthy AI.Privacy, safety and security, transparency and nondiscrimination.
- What it means
- Trustworthy AI means systems people can rely on to behave as intended, respect rights and avoid harm. NVIDIA groups its principles under privacy, safety and security, transparency, and nondiscrimination, alongside accountability for outcomes.
- What to understand
- Privacy covers lawful, consented and minimal use of personal data. Safety and security mean the system resists misuse and avoids harmful output. Transparency means users know they are dealing with AI and can understand sources and limits, supported by model cards and citations. Nondiscrimination means comparable treatment across groups. Human oversight and clear accountability tie these together.
- How it shows up
- A concern about an AI system is described and you name the principle it affects.
- Common confusion
- Equating transparency with releasing model weights; it means clear disclosure of behavior, data and limitations.
5.2 Describe the balance between data privacy and the importance of data consent.Using data lawfully while respecting individuals' control over it.
- What it means
- Teams need data to build useful models, but individuals have rights over their data. This objective is about weighing those interests through consent, minimization and protection.
- What to understand
- Informed consent means people understand and agree to how their data will be used, and purpose limitation keeps use within that agreement. Data minimization collects only what is needed. Anonymization, pseudonymization and PII redaction lower risk, though re-identification can still occur. Removing data from trained weights is hard, so consent should be settled before training. Federated learning and differential privacy reduce exposure.
- How it shows up
- A team wants to reuse collected data and you decide what privacy or consent step is required.
- Common confusion
- Assuming anonymized data removes every obligation; consent terms and re-identification risk still apply.
5.3 Describe how to use NVIDIA and other technologies to improve AI trustworthiness.Guardrails, safety models, grounding and secure computing.
- What it means
- NVIDIA and others provide tools that make LLM systems safer and more transparent. You should know what each does and where it sits in an application.
- What to understand
- NeMo Guardrails defines input, output, dialog and topical rails that screen requests and responses. NVIDIA also publishes safety models as NIM microservices for content safety, topic control and jailbreak detection. Retrieval-augmented generation grounds answers in sources that can be cited. NeMo Curator removes personal data from training sets. Confidential computing on recent NVIDIA GPUs protects data in use, and NVIDIA FLARE supports federated learning.
- How it shows up
- A safety or transparency need is described and you pick the NVIDIA tool designed for it.
- Common confusion
- Choosing an inference optimizer such as TensorRT-LLM to solve a content safety problem.
5.4 Describe how to minimize bias in AI systems.Sources of bias and ways to measure and reduce it.
- What it means
- Bias is systematic unfair difference in how a system treats people or groups. This objective covers where it comes from and how to detect and reduce it across the lifecycle.
- What to understand
- Bias enters through unrepresentative or historical data, skewed labeling, model design and evaluation that ignores some groups. Detect it with metrics broken down by group, counterfactual tests that swap demographic terms, and red-teaming. Reduce it with balanced, curated data, diverse annotators, fine-tuning or preference training aimed at fair responses, output filters and ongoing human review after deployment.
- How it shows up
- A biased outcome is described and you pick the step that measures or reduces it.
- Common confusion
- Checking only overall accuracy, which can hide poor performance for smaller groups.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
NVIDIA trustworthy AI principles
Privacy, safety and security, transparency, and nondiscrimination.
- Remember
- Accountability and human oversight support all four.
- Tested as
- Principle identification questions.
- Trap
- Treating accuracy as one of the principles.
-
Informed consent
People knowingly agree to a specific use of their data.
- Remember
- New uses may need new consent.
- Tested as
- Privacy questions.
- Trap
- Assuming one consent covers any future use.
-
Data minimization
Collecting and keeping only the data a purpose requires.
- Remember
- Less data means less exposure.
- Tested as
- Privacy design questions.
- Trap
- Hoarding data in case it is useful later.
-
PII redaction
Detecting and removing personal identifiers from datasets.
- Remember
- Do it before training, since removing data from weights is hard.
- Tested as
- Data curation questions.
- Trap
- Relying on the model not to repeat personal data.
-
NeMo Guardrails
Open-source toolkit for programmable rails around LLM conversations.
- Remember
- Controls topics, blocks jailbreaks and checks outputs without retraining.
- Tested as
- NVIDIA safety tooling questions.
- Trap
- Confusing it with fine-tuning.
-
Safety classifier models
Models that flag unsafe content, off-topic requests or jailbreak attempts.
- Remember
- NVIDIA offers them as NIM microservices.
- Tested as
- Safety tooling questions.
- Trap
- Expecting the main LLM to police itself reliably.
-
Hallucination and grounding
Confident but unsupported output; grounding ties answers to sources.
- Remember
- RAG with citations improves transparency.
- Tested as
- Reliability questions.
- Trap
- Thinking a lower temperature eliminates hallucination.
-
Prompt injection
Hidden instructions in input that override intended behavior.
- Remember
- Separate untrusted content and screen inputs.
- Tested as
- Security questions.
- Trap
- Treating it as a training data issue only.
-
Disaggregated evaluation
Reporting metrics separately for each group.
- Remember
- Reveals gaps that averages hide.
- Tested as
- Bias measurement questions.
- Trap
- Relying on a single overall score.
-
Counterfactual testing
Swapping attributes such as names or genders to compare outputs.
- Remember
- Differences point to bias.
- Tested as
- Bias detection questions.
- Trap
- Testing only one demographic variant.
-
Model card
Documentation of a model's intended use, data, performance and limits.
- Remember
- Supports transparency and accountability.
- Tested as
- Transparency questions.
- Trap
- Thinking it is marketing material.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Privacy and consent controls
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Informed consent | Agreement to a specified use | Governs whether data may be used | Assuming it extends to new purposes |
| Anonymization | Removing identifying links | Lowers risk, not obligations | Ignoring re-identification |
| Data minimization | Collect only what is needed | Limits exposure from the start | Collecting everything just in case |
| Federated learning | Train where data lives, share updates | Raw data stays local | Thinking it needs no other safeguards |
Ways to improve LLM safety
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Guardrails | Rules and checks around the model | No retraining, quick to update | Expecting them to fix the model's knowledge |
| Safety classifiers | Separate models flag unsafe content | Specialized detection | Running them on outputs only, skipping inputs |
| Alignment fine-tuning | Train the model toward safe responses | Changes model behavior itself | Treating it as the only safeguard |
| Grounding with RAG | Answers tied to retrieved sources | Improves accuracy and citations | Assuming it blocks harmful requests |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Anonymized means unrestricted
- What candidates pick
- Reuse anonymized customer data for any new model.
- Why it looks right
- Removing names feels like removing risk.
- Why it's wrong
- Consent and purpose limits still apply, and combined fields can re-identify people.
Clue to watch for The scenario introduces a new purpose for existing data.
-
Retrain to block topics
- What candidates pick
- Fine-tune the model to stop off-topic or jailbreak responses.
- Why it looks right
- Changing the model seems thorough.
- Why it's wrong
- Retraining is slow and leaves gaps; programmable guardrails and safety classifiers enforce policy at runtime.
Clue to watch for The need is quick, enforceable control over topics or attacks.
-
Average accuracy shows fairness
- What candidates pick
- Declare a model fair because overall accuracy is high.
- Why it looks right
- A single number is easy to report.
- Why it's wrong
- Averages can hide much worse results for smaller groups; per-group metrics reveal disparities.
Clue to watch for The scenario mentions different user groups.
-
Low temperature ends hallucination
- What candidates pick
- Set temperature to zero to stop false statements.
- Why it looks right
- Deterministic output sounds reliable.
- Why it's wrong
- A model can be consistently wrong; grounding answers in retrieved sources and citing them addresses hallucination.
Clue to watch for Answers must be factually supported.
-
Transparency means open weights
- What candidates pick
- Publish model weights to meet a transparency requirement.
- Why it looks right
- Openness sounds like transparency.
- Why it's wrong
- Transparency is about disclosing AI use, data, limits and sources to the people affected, for example through model cards.
Clue to watch for Users need to understand how an answer was produced.
Answer framework
Work through these steps for any Domain 5 scenario involving risk, privacy or fairness in an LLM system.
- Name the principle at stake first, whether privacy, safety and security, transparency or nondiscrimination, since it narrows the answer set.
- For data questions, check that consent covers the intended use, then apply minimization, redaction or anonymization as further protection.
- For unsafe or off-topic output, prefer runtime controls such as NeMo Guardrails or safety classifiers before retraining the model, using input rails to block jailbreak attempts and output rails to filter responses.
- For accuracy and transparency concerns, ground answers with retrieval, cite sources and document limits in a model card.
- For fairness, measure results per group and with counterfactual tests, such as swapping a name or gender in the prompt, then address the data, training or filtering stage responsible.
Practice question
Want more like this? The NCA-GENL practice test has questions for every domain.
Checklist
0 of 11
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Because this domain is small and descriptive, a focused few days are enough. Start by reading NVIDIA's trustworthy AI overview and its blog post on what trustworthy AI is, then write one sentence and one LLM example for each principle: privacy, safety and security, transparency and nondiscrimination.
Next, study privacy and consent as a pair. Make a short list of controls, including consent, purpose limitation, minimization, anonymization, PII redaction, federated learning and differential privacy, and note what each does and does not solve.
Then learn the tools. Read the NeMo Guardrails documentation overview to understand input, output, dialog and topical rails, and look at NVIDIA's safety models for content safety, topic control and jailbreak detection. Connect RAG to transparency through citations.
Finish with bias: where it enters, how per-group metrics and counterfactual prompts reveal it, and which lifecycle stage to fix. Practice with scenario questions and, for each, name the principle first and the control second. That two-step habit matches how this domain's questions are written.
FAQ
How much of NCA-GENL is the Trustworthy AI domain?
Trustworthy AI is 10% of the exam, the smallest domain, which is around five or six questions on a typical form.
What principles does NVIDIA use for trustworthy AI?
NVIDIA describes trustworthy AI in terms of privacy, safety and security, transparency, and nondiscrimination, supported by accountability and human oversight.
What is NeMo Guardrails?
An open-source NVIDIA toolkit that adds programmable rails to LLM applications, screening inputs and outputs, keeping dialogs on approved topics and blocking jailbreak attempts.
Does anonymizing data remove the need for consent?
No. Consent and purpose limits still govern use, and anonymized records can sometimes be re-identified by combining fields.
How can bias in an LLM be detected?
Compare metrics across groups, run counterfactual prompts that change only a demographic detail, and use red-teaming to probe for unfair or harmful responses.
How does RAG relate to trustworthy AI?
Grounding answers in retrieved documents reduces hallucination and lets the system cite sources, which supports transparency and user trust.
Are the questions in this domain technical?
Mostly descriptive. You need to recognize principles, controls and tools in scenarios rather than configure them.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 5 knowledge
You have covered trustworthy AI principles, privacy and consent, NVIDIA safety tools and bias reduction. See whether you can match each risk to its safeguard.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the NCA-GENL AI tutor Get any concept explained your way