PSM-AI flashcards

33 cards. Read the term and the definition here. The deck at the top of this page is optional.

Large language model (LLM)

A neural network trained on large text corpora to predict the next token, which lets it generate, summarize, translate, and transform text on request.

Token

The unit of text a language model processes, roughly three to four English characters; token counts drive context limits, response latency, and API cost.

Context window

The maximum number of tokens a model can consider at once, spanning system instructions, prompt, retrieved documents, and generated answer; content beyond it is unavailable.

Retrieval-augmented generation (RAG)

A pattern in which relevant documents are retrieved and inserted into the prompt so the model answers from supplied evidence instead of only its training data.

Fine-tuning

Further training of a pretrained model on domain-specific examples so its weights shift toward a desired style, format, or task behavior, without guaranteeing factual accuracy.

Hallucination

Fluent model output that is not supported by the provided context or by verifiable fact, typically stated with unwarranted confidence and requiring human verification.

Training versus inference

Training fits model parameters on data and happens periodically; inference is the per-request generation that uses the resulting fixed weights.

Temperature

A sampling setting that controls randomness: low temperature yields repeatable, focused output, while high temperature yields more varied but error-prone output.

Prompt structure template

Role, task, context, constraints, output format, and an example; specifying all six consistently raises output quality and reduces rework on generated artifacts.

Human-in-the-loop review

A named person checks and approves AI-generated content before it is used, keeping accountability with a human role rather than with the tool.

Drafting versus deciding

A decision rule separating AI-suitable work such as drafting, summarizing, and generating options from human-owned work such as prioritizing, conflict resolution, and personnel decisions.

AI-assisted retrospective clustering

Using a model to group anonymized retrospective comments into themes, then validating those themes with the team before any action is agreed.

Verification against primary source

Checking an AI claim in an authoritative document or system before it enters a Scrum artifact, rather than trusting a second model response as confirmation.

Velocity as non-target

Velocity is a planning observation, not a goal; machine-generated comparisons or targets distort forecasting and encourage gaming of the metric.

Impediment triage with AI

Using a model to search, summarize, and route blocker information, while the Scrum Master still escalates and resolves organizational impediments personally.

Definition of Done AI clause

An added Done condition stating that AI-assisted work was human-reviewed, sources were cited, and no confidential or personal data entered a prompt.

Automation bias

The tendency for people to accept machine suggestions uncritically, especially under time pressure; countered by spot checks and deliberate disagreement exercises.

Psychological safety for AI errors

A team climate in which admitting an AI mistake is safe; without it errors are hidden and the team cannot inspect and adapt.

Team prompt library

A shared, versioned collection of proven prompts and patterns owned by the team, which stabilizes output quality and shortens onboarding time.

Timeboxed AI experiment

A small hypothesis-driven trial of an AI tool inside a sprint, inspected at the retrospective and continued, changed, or dropped based on evidence.

Deskilling risk

The gradual loss of team capability when AI performs core work, reducing the human judgment available to detect and correct AI failures.

AI adoption backlog

A prioritized list of AI use cases with written value hypotheses and acceptance criteria, managed empirically in the same way as product backlog items.

Baseline metric before pilot

The pre-AI measurement of cycle time, rework, or time-to-answer needed to prove or disprove an adoption value hypothesis after the pilot.

EU AI Act risk tiers

Four tiers (unacceptable, high, limited transparency, minimal) that set escalating obligations; high-risk uses require risk management, data governance, logging, and human oversight.

GDPR and prompt data

Personal data entered into a prompt counts as processing under GDPR, requiring a lawful basis, data minimization, and a processor agreement with the model vendor.

Community of practice for AI

A voluntary cross-team group that shares prompts, failure cases, and standards, accelerating diffusion of proven practice beyond the initial pilot team.

Adoption KPI versus licenses

Adoption is evidenced by active use, task cycle time, rework rate, and incident counts, not by the number of seats or licenses purchased.

Algorithmic bias

Systematic error in model outputs that reflects patterns in training data, producing unfair or skewed results for particular groups and requiring testing and review.

Prompt injection

An attack in which text inside a document, email, or web page instructs the model to ignore its rules or leak data; untrusted content must be treated as hostile.

Accountability delegation fallacy

The incorrect belief that a model can be answerable for an outcome; a human role must remain responsible for every AI-assisted decision.

AI disclosure to stakeholders

Telling stakeholders when content or decisions were AI-assisted, preserving transparency and their ability to question or contest the outcome.

Data classification before prompting

Labeling information as public, internal, confidential, or personal before use, so only approved categories are entered into an external model.

Model drift

Degradation of a model's real-world performance as the environment changes after training, requiring scheduled re-evaluation and monitoring of reported errors.