Overview
Ethical Considerations of AI was the heaviest section of the retired AI Associate exam. Its two objectives were to describe the ethical challenges of AI, with examples such as human bias in machine learning and lack of transparency, and to apply Salesforce's trusted AI principles to given scenarios.
The challenges include bias that enters through historical data or proxy fields, opaque models whose reasoning users cannot see, privacy and consent risks with personal data, hallucinated or harmful generated content, and unclear accountability when an automated decision goes wrong. Salesforce's original trusted AI principles are Responsible, Accountable, Transparent, Empowering and Inclusive. For generative AI it added five guidelines: Accuracy, Safety, Honesty, Empowerment and Sustainability. The Einstein Trust Layer puts several of these into practice through data masking, zero data retention with model providers, toxicity detection and audit trails. Scenario questions ask which principle is at stake and which action honors it.
How it's tested
This section was scenario-heavy. You read a short story about a model, a dataset or a generated output, then identify the ethical risk, name the trusted AI principle involved, or choose the mitigation that best addresses the problem. Several options often sound ethical, so precision about which principle fits matters.
-
Spot the risk
A situation is described and you identify whether the core issue is bias, opacity, privacy, safety, accuracy or accountability.
-
Match the principle
You link a behavior, such as explaining score factors or citing sources, to the specific Salesforce principle or generative AI guideline it supports.
-
Best mitigation
Of several reasonable actions, you pick the one that addresses root cause, such as removing a proxy field or adding human review.
A lending partner's model approves fewer applicants from certain postal codes, although it never uses race. Postal code can act as a proxy for protected characteristics, so this is bias introduced through data. Adding a disclaimer would not fix it. The stronger mitigation is to assess the field for proxy effects, test outcomes across groups, and remove or adjust the feature, which supports the Inclusive and Responsible principles.
Objectives 2
Official objectives from the Salesforce outline, in outline order.
3.1 Describe the ethical challenges of AI (e.g. human bias in machine learning, lack of transparency, etc.).Bias, opacity, privacy, harmful output and accountability risks.
- What it means
- AI can repeat and amplify unfairness found in its training data, hide how it reached a result, expose or misuse personal data, and, for generative models, produce confident falsehoods or toxic text. Because AI decisions can affect many people quickly, these risks scale faster than individual human error.
- What to understand
- Know the main bias sources: historical bias in past decisions, sampling bias from unrepresentative data, measurement bias from flawed fields, proxy variables that stand in for protected attributes, and confirmation or automation bias in how people trust outputs. Understand that transparency means users can see why a result was produced and that AI was involved. Recognize privacy duties such as consent and data minimization, and that someone must remain accountable for every automated decision.
- How it shows up
- Expect scenarios where you name the type of bias or risk; the clue is usually where the unfairness entered.
- Common confusion
- Assuming bias is removed simply by deleting a protected field like gender, while proxies remain.
3.2 Apply Salesforce's trusted AI principles to given scenarios.Using Salesforce's principles, generative AI guidelines and Trust Layer controls in practice.
- What it means
- Salesforce's trusted AI principles say AI should be Responsible, Accountable, Transparent, Empowering and Inclusive. Its generative AI guidelines add Accuracy, Safety, Honesty, Empowerment and Sustainability. Applying them means choosing design and process decisions, such as human review, disclosures, masking and testing, that keep AI trustworthy for customers and employees.
- What to understand
- Map each principle to concrete behavior. Transparent: explain score factors and disclose AI-generated content. Accountable: keep humans responsible and maintain audit trails. Inclusive: test across groups and use representative data. Accuracy: ground outputs in trusted data and cite sources. Safety: reduce bias, toxicity and data leakage, for example through the Einstein Trust Layer's masking and zero data retention. Honesty: respect data provenance and consent. Sustainability: right-size models for the job.
- How it shows up
- Questions describe an action and ask which principle it supports, or ask which action best upholds a named principle.
- Common confusion
- Mixing the original five principles with the five generative AI guidelines, or treating Empowering and Empowerment as unrelated.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Responsible
Protect human rights and the data entrusted to Salesforce.
- Remember
- Covers safeguarding people and data throughout the AI lifecycle.
- Tested as
- Fits scenarios about protecting customers from harm.
- Trap
- Treating it as only a legal compliance checkbox.
-
Accountable
Seek feedback and own outcomes, with humans answerable for AI use.
- Remember
- Someone must be responsible for each automated decision.
- Tested as
- Appears in audit trail and review-process scenarios.
- Trap
- Assuming the model vendor carries all accountability.
-
Transparent
Make clear how AI reaches results and when AI is involved.
- Remember
- Explainable factors and disclosure of AI content.
- Tested as
- Fits score-explanation and labeling scenarios.
- Trap
- Thinking publishing source code satisfies transparency for users.
-
Empowering
Make AI accessible and helpful so people can do more.
- Remember
- AI should augment users' abilities.
- Tested as
- Fits scenarios about enabling non-technical users.
- Trap
- Confusing it with automating people out of the loop.
-
Inclusive
Respect societal values and test for fair outcomes across groups.
- Remember
- Representative data and outcome testing.
- Tested as
- Fits bias-mitigation scenarios.
- Trap
- Believing inclusivity is achieved by marketing language alone.
-
Accuracy (generative guideline)
Deliver verifiable results grounded in trusted data.
- Remember
- Cite sources and flag uncertainty.
- Tested as
- Fits hallucination scenarios.
- Trap
- Assuming a fluent answer is an accurate one.
-
Safety (generative guideline)
Mitigate bias, toxicity and harmful output, and protect personal data.
- Remember
- Includes red-teaming and toxicity checks.
- Tested as
- Fits harmful-content and leakage scenarios.
- Trap
- Limiting safety to physical harm.
-
Honesty (generative guideline)
Respect data provenance and consent, and disclose AI-created content.
- Remember
- Users should know when content is generated.
- Tested as
- Fits consent and disclosure scenarios.
- Trap
- Merging it entirely with accuracy.
-
Empowerment and Sustainability
Keep humans in the loop where it matters, and right-size models to limit environmental cost.
- Remember
- Smaller fit-for-purpose models can be the ethical choice.
- Tested as
- Less frequent but appears in guideline-matching items.
- Trap
- Assuming the largest model is always preferred.
-
Einstein Trust Layer
Salesforce's set of controls between CRM data and large language models.
- Remember
- Masking, zero data retention, toxicity detection, audit trail and grounding.
- Tested as
- The concrete mechanism behind safety and privacy answers.
- Trap
- Thinking it trains the LLM on customer data.
-
Proxy variable
A field that indirectly encodes a protected attribute.
- Remember
- Postal code or first name can stand in for ethnicity or gender.
- Tested as
- Key to explaining hidden bias.
- Trap
- Assuming bias disappears once protected fields are removed.
-
Human in the loop
A person reviews or approves AI output before it has effect.
- Remember
- Most valuable for high-impact or customer-facing decisions.
- Tested as
- Common best-mitigation answer.
- Trap
- Applying it as a rubber stamp without real review.
-
Hallucination
Generated content that sounds plausible but is false.
- Remember
- Reduced by grounding and review, not eliminated.
- Tested as
- Central to generative accuracy questions.
- Trap
- Blaming it only on bad prompts.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Trusted AI principles vs generative AI guidelines
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Original principles | Responsible, Accountable, Transparent, Empowering, Inclusive. | Apply to all AI, predictive or generative. | Using them as the only list for generative questions. |
| Generative AI guidelines | Accuracy, Safety, Honesty, Empowerment, Sustainability. | Address risks specific to content generation. | Assuming they replaced the original principles. |
| Einstein Trust Layer | Technical controls that enforce privacy and safety for LLM use. | A mechanism, not a principle. | Answering a principle question with a product name. |
Sources of bias
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Historical bias | Past decisions were unfair and the model learns them. | Exists even with perfectly measured data. | Thinking more of the same data fixes it. |
| Sampling bias | Training data does not represent the population served. | Groups are under- or over-represented. | Confusing it with measurement errors. |
| Measurement bias | Fields are recorded inconsistently or inaccurately across groups. | The data collection process is flawed. | Assuming the model itself is the source. |
| Automation bias | People over-trust AI output. | A human behavior, not a data flaw. | Treating human review as automatically effective. |
Transparency vs explainability vs disclosure
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Explainability | Showing the factors behind a specific prediction. | Answers why this result. | Equating it with sharing model code. |
| Disclosure | Telling users that AI produced or influenced content. | Answers whether AI was involved. | Assuming disclosure alone makes output trustworthy. |
| Auditability | Keeping records of prompts, outputs and decisions. | Supports accountability after the fact. | Treating logs as a substitute for review. |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Delete the protected field
- What candidates pick
- Removing gender or ethnicity fields as the full bias fix.
- Why it looks right
- It looks like a direct, simple action.
- Why it's wrong
- Other fields can act as proxies, so outcomes must still be tested across groups.
Clue to watch for The model still shows uneven results after the change.
-
Disclaimer as mitigation
- What candidates pick
- Adding a notice that AI may be wrong.
- Why it looks right
- Disclosure is a real principle.
- Why it's wrong
- A notice informs users but does not reduce bias, inaccuracy or harm at the source.
Clue to watch for The question asks how to reduce or prevent the problem.
-
Full automation as empowerment
- What candidates pick
- Letting AI send customer decisions with no review.
- Why it looks right
- Efficiency sounds empowering.
- Why it's wrong
- Empowering and Empowerment mean augmenting people, and high-impact actions call for human oversight.
Clue to watch for The decision affects customers' finances, access or rights.
-
Biggest model is best
- What candidates pick
- Choosing the largest LLM for a simple classification task.
- Why it looks right
- Bigger models seem more capable.
- Why it's wrong
- Sustainability favors right-sized models, and a smaller model may be more accurate and efficient for narrow tasks.
Clue to watch for The task is narrow and well defined.
-
Trust Layer trains on data
- What candidates pick
- Saying the Trust Layer lets the LLM learn from customer records.
- Why it looks right
- Grounding uses CRM data in prompts.
- Why it's wrong
- Zero data retention means the external model provider does not keep or train on the prompt data.
Clue to watch for Wording about what the model provider stores.
-
Transparency as source code
- What candidates pick
- Publishing the algorithm as the transparency answer.
- Why it looks right
- It sounds maximally open.
- Why it's wrong
- Users need understandable explanations and disclosure, not code they cannot interpret.
Clue to watch for The audience is business users or customers.
Answer framework
Use these steps for any ethics scenario on this outline.
- Name the harm in one phrase: unfair outcome, unexplained result, privacy exposure, false content, toxic content, or nobody accountable for the decision.
- Locate where it entered: historical data, unrepresentative sampling, a proxy field, the prompt, the model's generation, or how people use the output.
- Match it to the principle or guideline: Inclusive or Safety for bias, Transparent or Honesty for disclosure, Accuracy for hallucination, Accountable for ownership.
- Choose the mitigation that addresses the cause, such as representative data, outcome testing, grounding, masking, human review or an audit trail.
- Discard options that only add wording, remove human oversight on high-impact decisions, or assume a technical control solves a process problem by itself.
Practice question
Want more like this? The SF AI Associate practice test has questions for every domain.
Checklist
0 of 12
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Spend the most time here: eight to ten hours across several sessions. Begin by learning the two lists cold, the five trusted AI principles and the five generative AI guidelines, and write one concrete CRM action for each item. Then study the Trailhead content on responsible AI and the Einstein Trust Layer so you can describe masking, zero data retention, toxicity scoring, grounding and audit trails without notes.
Next, build a bias table. For each bias type, write where it comes from, a CRM example and the best fix. Include at least two proxy-variable examples, because they explain why deleting a protected field rarely solves the problem. Then practice scenarios in a fixed order: name the harm, locate its source, match the principle, pick the mitigation. Time yourself at about 90 seconds per scenario.
When you review mistakes, check whether you chose a disclosure when a fix was asked, or a technical control when the problem was a missing process. Those two errors account for most lost points in this section. Revisit Data for AI afterward, since many bias problems start as data quality problems.
FAQ
How much of the Salesforce AI Associate exam was Ethical Considerations of AI?
Ethical Considerations of AI was the largest section at 39%, about 16 of the 40 questions. Together with Data for AI at 36%, it made up three quarters of the exam.
What are Salesforce's trusted AI principles?
Responsible, Accountable, Transparent, Empowering and Inclusive. Salesforce published them for all AI, and later added five guidelines for trusted generative AI.
What are the five guidelines for trusted generative AI?
Accuracy, Safety, Honesty, Empowerment and Sustainability. They address generative risks such as hallucination, toxic output, undisclosed AI content, removing humans from important decisions and the environmental cost of very large models.
What is the Einstein Trust Layer?
It is a set of controls between Salesforce data and large language models. It includes secure data retrieval and grounding, dynamic data masking, zero data retention with model providers, toxicity detection and an audit trail.
Does removing sensitive fields stop a model from being biased?
Not reliably. Other fields can act as proxies for protected attributes, and historical decisions may already be biased. Outcomes need testing across groups, with fields adjusted or removed based on results.
Were questions only about memorizing principle names?
No. Most questions described a situation and asked which principle applied or which action best upheld it. Knowing the names helps, but mapping them to actions is what earns the points.
Is this ethics content relevant after the exam's retirement?
Yes. Trusted AI principles, guardrails and the Einstein Trust Layer remain part of Salesforce's Agentforce platform and its newer AI training.
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 3 knowledge
Test yourself on bias scenarios, trusted AI principles and generative AI guidelines, the section that carried the most weight on this outline.
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 SF AI Associate AI tutor Get any concept explained your way