Free PSM-AI practice test: 30 questions with answers

Hallucination describes plausible-sounding output that is not grounded in verified sources, which fits a citation to a clause that does not exist. Retrieval would need a real search step, which the scenario does not describe, overfitting means reproducing training text, and quantization is a compression technique unrelated to invented citations.

A language model generates output by predicting the most likely next token given the preceding context, a process learned from training data. The other options describe live internet search, hand-written rule systems, or conversation lookup, none of which is how a base language model generates text.

Retrieving the current policy at question time, a pattern called retrieval-augmented generation, keeps answers aligned with the latest version without retraining. A one-time fine-tune or a fixed pasted policy goes stale after the next quarterly change, and general knowledge ignores the company's own rules.

The context window is the span of tokens, including the prompt and the response, that the model can consider at one time. The other options describe concurrency limits, content policies, or tenant isolation, which are different controls.

Generative models pick each next token probabilistically, so repeated requests can produce different wording unless randomness settings are constrained. The first option describes caching the scenario does not mention, and the second and third wrongly claim the model learns or memorizes during ordinary use.

Training learns parameters from data, inference applies those fixed parameters to new inputs, and built-in knowledge stops at the training cutoff unless new information is supplied. The second and fourth statements are false because ordinary inference does not update weights and data use depends on provider terms, while the fifth describes design work rather than inference.

Artificial intelligence is the broadest field, machine learning is a subset that learns from data, and deep learning is a subset of machine learning that uses multi-layer neural networks. The other options reverse the hierarchy or wrongly claim AI is limited to images or that deep learning replaced machine learning.

Unsupervised learning finds patterns and groupings in data that has no labels, which matches the goal of grouping comments without predefined categories. Supervised learning needs labeled examples, reinforcement learning depends on reward signals for actions, and a keyword list is hand-written rules rather than learning from data.

Pasting confidential code into an external tool can move client data outside the organization's control, so data-handling rules and approved tools should decide what may be shared. The other options describe effects the scenario does not support, such as automatic publication or guaranteed wrong fixes.

Prompt injection hides instructions in content the model processes so that they override the developer's intended behavior. Message flooding is a denial-of-service attack, repeated querying to copy behavior is model extraction, and data poisoning targets training data rather than a live prompt.

Bias is a systematic skew in outputs that comes from unrepresentative data or design decisions, so it can affect groups unfairly even when nobody intends it. Random wording variation, deliberate vendor favoritism, and refusals are different phenomena that do not define bias.

Hiring decisions affect people, so a named person must stay accountable and the ranking should be audited for bias before anyone relies on it. The other options wrongly treat a tool as neutral, use a narrow criterion that does not remove bias, or delay oversight until after complaints.

Classifying data before entry and verifying outputs before sharing are standard controls for generative AI use. Skipping review, using personal accounts for company data, and assuming outputs are free of third-party rights each increase risk.

Transparency lets stakeholders see when AI influenced a result and judge how far to trust it. The other options describe copyright guarantees, vendor pricing, or skipping quality checks, none of which transparency achieves.

Rating people from private communications without disclosure or consent is covert surveillance, which violates basic expectations of transparency and fairness. The other options describe usefulness or usability problems, which are secondary to the ethical breach.

Contracts contain sensitive data, so the team should first confirm how the vendor stores, retains, and uses submitted content, including whether it is used for training. Export formats, answer length, and marketing claims do not address the data risk.

The Scrum Team owns its retrospective, so AI output is a draft that the team verifies and then decides on. The other options let the AI decide, destroy evidence the team should inspect, or have one person decide for the whole team.

The Product Owner is accountable for Product Backlog management, which includes ordering its items and keeping them transparent. The Developers estimate and build, the Scrum Master facilitates, and stakeholders request, but none of them is accountable for managing the backlog.

The Scrum Team is accountable for the Increment, and AI-assisted code must still meet the team's Definition of Done through normal review. A vendor or an automatic process cannot be accountable, and making one Developer solely responsible ignores the team's shared accountability.

Drafting refinement questions, summarizing feedback for discussion, and drafting acceptance criteria for review all support the team while leaving decisions with it. Replacing the Sprint Review with an unread report and reassigning work without agreement both remove the team's inspection or self-management.

Empiricism relies on transparency, inspection, and adaptation based on evidence, and accepting an AI estimate without discussion skips that inspection. Timeboxing, self-organization, and scope flexibility are not what this situation threatens.

Scrum's empirical process control rests on transparency, inspection, and adaptation. Commitment, focus, and openness are Scrum values, while planning, velocity, and cost tracking are practices or metrics rather than pillars.

A forecast is one piece of evidence the team can discuss and test against real progress, which keeps the decision with the team. Treating it as final, replacing the Daily Scrum, or blaming Developers each misuses the tool and undermines transparency.

Few-shot prompting shows the model a small number of worked examples so it can follow the pattern for the actual request. Zero-shot uses no examples, chain-of-thought asks for step-by-step reasoning, and system prompting sets a persistent role rather than demonstrating a pattern.

Specific context about the team, audience, format, and length gives the model the constraints it needs to produce a relevant draft. Repeating the request with exclamation marks, switching language, or deleting background all remove useful information rather than adding it.

Step-by-step prompting can help with multi-step problems because intermediate reasoning is produced and can be inspected, though it does not guarantee correctness. The other options claim guaranteed accuracy, fewer tokens, or no need for review, which are all incorrect.

Naming the audience and purpose, and specifying the output format, give the model the context and constraints it needs. A vague instruction, contradictory instructions, and missing background each make the output less predictable and less useful.

Giving the model the failing output and expected behavior lets it correct the specific problem, and running the tests verifies the revision. Accepting confident output, asking the model to grade itself, or switching models without checking all skip verification.

Lower temperature makes sampling favor the most likely tokens, so outputs become more consistent and focused. It does not change the context window, and it does not make the model ignore the prompt; higher values are the ones associated with more variety.

Iterating with a targeted revision request, including a length limit and the points to keep, refines the draft without starting over. Retyping from memory, sending the draft unedited, or asking for more detail all work against the executive audience's need for brevity.