Identify AI concepts and capabilities practice questions

Reliability and safety: AI is probabilistic, so systems should behave safely when a prediction is uncertain.

Computer vision models return predictions with a confidence score and are sometimes wrong. Refusing to act when confidence is low, and handing the decision to a person, is a safety mechanism that keeps the robot from causing damage. Accountability is the tempting distractor because a human is involved, but the design's goal is safe behaviour under uncertainty.

1. Note that the model's output comes with a confidence score.
2. See what the design does: it avoids acting when the model is unsure.
3. Ask which principle is about avoiding unintended harm: reliability and safety.
4. Don't be drawn to accountability just because a person is involved.

Exam tip: Acting only above a confidence threshold is a reliability and safety control.

AI systems are based on probabilistic models and are not infallible.

Microsoft frames reliability and safety around the fact that AI is probabilistic: a model generates the most likely output for an input, which is usually but not always right, and it can be confidently wrong on inputs unlike those seen before. A successful demo covers only a few cases, so the team must test a wide range of normal, edge and adversarial inputs and plan for failures.

1. Recall how generative models produce output: by predicting likely tokens.
2. Conclude that output quality varies and errors look plausible.
3. Reject claims about per-prompt retraining or determinism.
4. Connect varied testing to reliability and safety.

Exam tip: AI is probabilistic, not infallible: test beyond the happy path.

Content filters (guardrails) mitigate the risk of harmful content generation.

Generative models can produce harmful content such as violence, hate or self-harm material. Content filters, configured as guardrails on a model deployment in Microsoft Foundry, detect and block those categories in prompts and responses. They are a mitigation for safety risk, so they fall under reliability and safety.

1. Identify the risk: the model generating harmful content.
2. Identify the control: filters that block harmful categories.
3. Map preventing harm to people to reliability and safety.

Exam tip: Content filters are a reliability and safety mitigation.

Reliability and safety measures for a generative AI app: guardrails plus testing and monitoring.

Reliability and safety asks that the system behave as designed, respond safely to unexpected input and resist harmful manipulation. In practice that means mitigations such as content filters on the deployment and a testing regime that includes adversarial prompts, followed by monitoring once live. The other measures are good practice under different principles: a disclosure is transparency, a named owner is accountability and accessibility is inclusiveness.

1. Restate the principle: safe, consistent behaviour, including under misuse.
2. Pick the preventive control: content filters on the deployment.
3. Pick the verification step: adversarial testing plus monitoring.
4. Map the remaining measures to transparency, accountability and inclusiveness.

Exam tip: Guardrails plus testing and monitoring = reliability and safety; disclosures, owners and accessibility belong to other principles.

Privacy and security: protect personal data used to build and run AI systems.

AI systems need data, and that data often contains personal information. Keeping it only as long as needed, stripping identifiers such as names and card numbers, and restricting access are standard ways to protect it. Those are privacy and security considerations, not changes to how the model behaves or explains itself.

1. List what the team does: limits retention, removes identifiers, restricts access.
2. Notice that every step protects personal data.
3. Map data protection to privacy and security.

Exam tip: Minimise, de-identify and restrict access to personal data: privacy and security.

Protecting personal data in training data: de-identification plus access control and encryption.

Training data that contains personal information creates two risks: the data store can be exposed, and the model can learn and repeat details. Masking or removing identifiers addresses the second; restricting access and encrypting the dataset addresses the first. Representative data, a transparency note and human review are good practices for fairness, transparency and reliability respectively.

1. Identify the asset at risk: personal details in the transcripts.
2. Pick the step that reduces what the model can learn: de-identification.
3. Pick the step that protects the stored data: access control and encryption.
4. Map the remaining options to other principles.

Exam tip: Personal data in training sets: de-identify it and lock it down.