Overview
Domain 7 is the smallest section at 5% and has four objectives, each phrased as describe: the ethical principles of trustworthy AI, the balance between data privacy and data consent, how to use NVIDIA and other technologies to improve trustworthiness, and how to minimize bias in AI systems.
Multimodal generation raises these issues sharply. Faces and voices are personal data, image generators can produce deepfakes or reproduce stereotypes, and training sets scraped from the web may include content used without consent. NVIDIA frames trustworthy AI around privacy, safety and security, transparency and nondiscrimination. Relevant tools include NeMo Guardrails, Model Card++ documentation, federated learning with NVIDIA FLARE, synthetic data, and the NIST AI Risk Management Framework Generative AI Profile listed on the certification page.
How it's tested
Expect short scenario questions about a generative or recognition system that raises an ethical, privacy or bias concern. You select the principle at stake or the mitigation that best addresses it, often choosing between technical and process-based safeguards for images, audio or text.
-
Name the principle
A situation such as undisclosed AI-generated media is described, and you identify the trustworthy AI principle involved.
-
Balance privacy and consent
Use of faces, voices or personal records is proposed, and you choose the step that respects consent and privacy.
-
Reduce the bias
Skewed outputs or uneven accuracy across groups are described, and you choose the mitigation.
An image generator asked for pictures of nurses returns almost only women, and pictures of engineers almost only men. The model reflects imbalances in its training data. Adding a disclaimer does not change outputs, and removing occupation words from prompts avoids the problem rather than fixing it. Rebalancing or augmenting training data and evaluating outputs across demographic groups addresses the bias directly.
Objectives 4
Official objectives from the NVIDIA outline, in outline order.
7.1 Describe the ethical principles of trustworthy AIPrivacy, safety, transparency, fairness and accountability.
- What it means
- Trustworthy AI means systems that respect people's rights and behave as intended. NVIDIA groups its principles under privacy, safety and security, transparency, and nondiscrimination.
- What to understand
- Privacy protects personal data. Safety and security prevent harm and misuse, such as deepfakes or unsafe instructions. Transparency means explaining how a system works and disclosing AI-generated content. Nondiscrimination means fair treatment across groups. Accountability assigns responsibility for outcomes and human oversight. Frameworks such as the NIST AI RMF organize these into practices.
- How it shows up
- A scenario is described, and you identify which principle is affected or upheld.
- Common confusion
- Equating transparency with publishing source code; it is about explaining behavior, data use and AI involvement.
7.2 Describe the balance between data privacy and the importance of data consentUsing personal images, voices and text lawfully and respectfully.
- What it means
- Multimodal models often train on photos, voice recordings and personal text. Privacy limits how such data is collected and used, and consent establishes whether people agreed to that use.
- What to understand
- Consent should be informed and specific to the purpose; consent to post a photo publicly is not consent to train a face model on it. Data minimization collects only what is needed. Anonymization techniques include face blurring and voice de-identification. Federated learning trains without centralizing raw data. Honor removal requests and follow regulations such as GDPR.
- How it shows up
- A proposed data use is described, and you choose the step that respects privacy and consent.
- Common confusion
- Assuming publicly available data is free to use for training without consent or license review.
7.3 Describe how to use NVIDIA and other technologies to improve AI trustworthinessGuardrails, documentation, privacy tech and provenance.
- What it means
- Technology can enforce and demonstrate trustworthiness, from filtering outputs to documenting models and protecting data.
- What to understand
- NeMo Guardrails adds programmable rails that block unsafe topics or responses. Model Card++ documents a model's intended use, limitations and bias evaluation. NVIDIA FLARE supports federated learning, and confidential computing on recent GPUs protects data in use. Synthetic data generation can fill gaps without real personal data. Content provenance standards such as C2PA and watermarking help label AI-generated media.
- How it shows up
- A trust need is described, and you choose the tool or technique that meets it.
- Common confusion
- Expecting guardrails to remove bias from training data; they filter interactions, while data work addresses bias.
7.4 Describe how to minimize bias in AI systemsFinding and reducing unfair outcomes across groups.
- What it means
- Bias arises when data, labels or design lead a system to treat groups unfairly. Multimodal systems show it through uneven recognition accuracy or stereotyped generations.
- What to understand
- Sources include unrepresentative data, biased labels and proxy features. Mitigations include diverse and balanced data collection, augmentation or synthetic data for underrepresented groups, evaluation sliced by demographic group, fairness metrics, and human review. Document known limitations in model cards. Monitor after deployment, since bias can emerge as usage changes.
- How it shows up
- Biased outputs or uneven accuracy are described, and you choose the most direct mitigation.
- Common confusion
- Believing that removing a sensitive attribute removes bias; correlated features can still act as proxies.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
NVIDIA trustworthy AI principles
Privacy, safety and security, transparency, nondiscrimination.
- Remember
- Accountability and human oversight support them.
- Tested as
- Principle identification.
- Trap
- Listing only fairness.
-
Informed consent
Agreement to a specific data use with understanding of it.
- Remember
- Public availability is not consent.
- Tested as
- Privacy and consent scenarios.
- Trap
- Scraping images without review.
-
Data minimization
Collect and keep only the data needed.
- Remember
- Reduces privacy exposure.
- Tested as
- Privacy design.
- Trap
- Collecting everything just in case.
-
Federated learning
Train across sites without moving raw data.
- Remember
- NVIDIA FLARE supports it.
- Tested as
- Privacy-preserving technology.
- Trap
- Thinking it shares raw data.
-
NeMo Guardrails
Programmable rails that constrain model conversations.
- Remember
- Filters topics, inputs and outputs.
- Tested as
- NVIDIA safety tools.
- Trap
- Expecting it to fix training data bias.
-
Model Card++
NVIDIA's expanded model documentation.
- Remember
- Covers bias, safety, privacy and explainability.
- Tested as
- Transparency questions.
- Trap
- Confusing it with a license.
-
Content provenance and watermarking
Labels and traces AI-generated media.
- Remember
- Supports transparency about synthetic content.
- Tested as
- Deepfake scenarios.
- Trap
- Assuming watermarks cannot be removed.
-
Proxy features
Inputs correlated with a protected attribute.
- Remember
- They carry bias after the attribute is removed.
- Tested as
- Bias mitigation questions.
- Trap
- Fairness through unawareness.
-
Disaggregated evaluation
Measuring performance separately for each group.
- Remember
- Reveals hidden disparities.
- Tested as
- Bias detection.
- Trap
- Relying on overall accuracy.
-
Synthetic data
Generated data used to fill gaps or protect privacy.
- Remember
- Can rebalance underrepresented groups.
- Tested as
- Mitigation options.
- Trap
- Assuming it carries no bias.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Trustworthy AI principles
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Privacy | Protect personal data and honor consent | About data subjects' rights | Treating public data as free to use |
| Safety and security | Prevent harm and misuse | About outputs and attacks | Ignoring deepfake misuse |
| Transparency | Explain behavior and disclose AI use | About understanding and disclosure | Equating it with open source |
| Nondiscrimination | Fair treatment across groups | About equitable outcomes | Assuming removing attributes suffices |
Trust technologies
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| NeMo Guardrails | Constrain model interactions | Runtime control | Expecting data debiasing |
| Model Card++ | Document model behavior and limits | Transparency artifact | Treating it as enforcement |
| Federated learning (FLARE) | Train without centralizing data | Privacy during training | Assuming raw data moves |
| Provenance and watermarking | Label generated media | Disclosure of synthetic content | Assuming tamper-proof labels |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Public means permitted
- What candidates pick
- Train a voice model on podcast audio because it is publicly available.
- Why it looks right
- The data is easy to access.
- Why it's wrong
- Availability is not consent; voices are personal data and need consent and license review for training or cloning.
Clue to watch for The data involves identifiable people.
-
Disclaimer fixes bias
- What candidates pick
- Add a disclaimer to a generator that produces stereotyped images.
- Why it looks right
- It acknowledges the issue cheaply.
- Why it's wrong
- A disclaimer leaves outputs unchanged; rebalancing data, adjusting the model and evaluating across groups reduce the bias.
Clue to watch for The question asks how to reduce or minimize bias.
-
Drop the attribute
- What candidates pick
- Remove race or gender labels to make a model fair.
- Why it looks right
- The model cannot see the attribute.
- Why it's wrong
- Correlated features such as appearance, names or location act as proxies, so bias can persist.
Clue to watch for Other features correlate with the protected attribute.
-
Guardrails clean the data
- What candidates pick
- Use NeMo Guardrails to remove bias learned in training.
- Why it looks right
- Guardrails are an NVIDIA trust tool.
- Why it's wrong
- Guardrails constrain interactions at runtime; learned bias needs data and model-level mitigation.
Clue to watch for The problem originates in training data.
-
Anonymized forever
- What candidates pick
- Assume blurring faces makes any image dataset fully anonymous.
- Why it looks right
- Faces are the obvious identifier.
- Why it's wrong
- Tattoos, locations, license plates or metadata can still identify people, so anonymization needs a broader review.
Clue to watch for Images include context, metadata or other identifiers.
Answer framework
Use these steps when a Domain 7 question raises an ethical or safety concern.
- Name the principle at stake first: privacy, safety and security, transparency or nondiscrimination, since the right mitigation depends on which one is threatened.
- For privacy questions, check whether informed, purpose-specific consent exists and whether data can be minimized, anonymized or kept local with federated learning.
- For safety questions, consider runtime controls such as guardrails and content filters, plus provenance labels for generated images, audio or video.
- For bias questions, look for causes in data and labels, then choose rebalancing, augmentation, per-group evaluation and documented limitations over cosmetic fixes.
- Prefer answers that combine a technical safeguard with human oversight and documentation, since trustworthy AI is a process rather than a single tool.
Practice question
Want more like this? The NCA-GENM practice test has questions for every domain.
Checklist
0 of 10
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
This domain is small, so budget a few focused hours and aim for full marks. Start by memorizing the principles NVIDIA uses: privacy, safety and security, transparency and nondiscrimination, and attach one multimodal example to each, such as voice cloning for privacy or deepfakes for safety.
Then read the two NVIDIA articles listed in the study guide, Trustworthy AI for a Better World and What Is Trustworthy AI?, and skim the NIST Generative AI Profile linked from the certification page for vocabulary. Note the NVIDIA technologies they mention, especially NeMo Guardrails, Model Card++, federated learning and synthetic data.
For bias, build a short list of causes and matching fixes, and practice rejecting cosmetic answers such as disclaimers or simply deleting a sensitive column. For privacy and consent, rehearse why scraped faces and voices need review. Finish with a handful of practice questions; the patterns here repeat, so a small set is usually enough to confirm readiness.
FAQ
How much of NCA-GENM is Trustworthy AI?
Trustworthy AI is the smallest domain at 5% of the exam, usually two or three questions, with four objectives.
Which NVIDIA tools should I know for this domain?
NeMo Guardrails for runtime safety, Model Card++ for documentation, NVIDIA FLARE for federated learning, and synthetic data generation for filling data gaps.
Why is consent emphasized for multimodal models?
Faces and voices identify people directly, and generative models can reproduce or imitate them, so using them without consent creates privacy and legal risk.
Does removing sensitive attributes make a model fair?
Usually not. Other features correlated with the attribute can act as proxies, so bias must be measured per group and addressed in data and modeling.
Is the NIST AI RMF on the exam?
NVIDIA lists the NIST Generative AI Profile as a suggested reading on the certification page. Know its general purpose rather than its detailed structure.
How should I prepare for such a small domain?
A few hours on principles, consent, NVIDIA trust tools and bias mitigations is usually enough, followed by practice questions.
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 7 knowledge
You have reviewed trustworthy AI principles, privacy and consent, NVIDIA trust tools and bias mitigation. Confirm you can pick the right safeguard in short ethics scenarios.
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-GENM AI tutor Get any concept explained your way