AI-901 Cheat Sheet

The fairness principle requires an AI system to treat people fairly and to avoid affecting similarly situated groups in different ways. Fairness does not mean returning an identical result to everyone regardless of their circumstances.

Trap Return an identical result to every person, regardless of differences in their circumstances.

Evaluate error rates and outcomes for relevant groups because a strong overall score can hide poorer performance for a particular cohort. A fairness assessment should compare group-level behavior rather than treating aggregate accuracy as proof of equitable performance.

Trap Accept the model as fair when its overall accuracy is high, without comparing results for relevant groups.

The reliability and safety principle requires an AI system to perform as designed and respond safely across expected and unexpected conditions. Testing failure modes and defining safe fallback behavior are reliability and safety considerations.

Trap Measure only the system's average accuracy on its original test dataset.

Monitor an AI system after deployment because changing data, usage patterns, and operating conditions can expose behavior that predeployment tests did not reveal. Ongoing monitoring supports timely detection and mitigation of unsafe or inconsistent results.

Trap Stop evaluating the system after it passes its initial test set because deployed behavior will remain unchanged.

The privacy and security principle requires an AI system to respect privacy and remain secure. Appropriate considerations include minimizing exposed personal data, protecting stored and transmitted data, and restricting access with controls such as Microsoft Entra ID and role-based access control.

Trap Add multilingual output and screen-reader support to the user interface.

Privacy considerations include being clear about the collection, use, and storage of personal data and giving people appropriate control over how their data is used. These obligations complement technical security controls such as encryption and access restriction.

Trap Treat encryption alone as sufficient even when users are not told how their personal data will be used.

The inclusiveness principle requires AI systems to empower and engage people regardless of background and to be usable by people of all abilities. Supporting multiple interaction methods, languages, and assistive technologies such as screen readers is an inclusiveness consideration.

Trap Publish a detailed explanation of how the model generated each prediction.

Include people with varied abilities, language needs, backgrounds, and levels of technical proficiency in design and testing. Their participation can reveal barriers that a homogeneous development team might miss.

Trap Ask only the development team to review accessibility because they already understand how the interface works.

The transparency principle requires people to understand that they are interacting with AI and what the system can and cannot reliably do. User-facing disclosures should set appropriate expectations about AI-generated output.

Trap Encrypt customer records and restrict access by role.

When AI informs a consequential decision, provide useful information about how and why the system behaved as it did. Human-understandable explanations help stakeholders identify performance issues, fairness concerns, and unintended outcomes.

Trap Disclose only that AI is present without explaining a consequential result or the system's known limitations.

The accountability principle holds people and organizations responsible for how an AI system is designed, deployed, and used. Define ownership, governance processes, and responsibility for compliance throughout the system lifecycle.

Trap Treat the model provider as the only accountable party after the application is deployed.

An AI system should not be the final authority for decisions that significantly affect people's lives. Human oversight and mechanisms to review or contest consequential outcomes support accountability.

Trap Remove human review once the model exceeds an accuracy target for consequential decisions.

A tokenizer breaks input text into tokens, and a large language model selects an output token based on the input sequence. The selected token is appended to the sequence and becomes context for selecting the next token, so generation proceeds one token at a time.

Trap The model retrieves a complete, stored answer that exactly matches the prompt.

During training, a large language model analyzes how tokens occur together and in similar contexts, encoding learned relationships in its parameters. At inference time, it applies those learned patterns to the prompt; it does not verify each generated claim against a factual database by default.

Trap The deployment retrains the model on every user prompt before returning a response.

A model's context window limits the tokens it can process for a request. Input tokens, generated output tokens, and any reasoning tokens consume the available context budget, so a longer prompt can leave less room for generation.

Trap Input and output token limits are always additive, so using the full input limit leaves the full output limit available.

Chat models accept an ordered transcript of messages rather than a single unstructured completion string. A system message supplies behavior and constraints, user messages supply requests, and assistant messages can preserve prior responses or provide examples.

Trap The assistant role is where application-wide rules must be placed before every system message.

An embedding is a numerical vector representation of semantic meaning. Semantically similar inputs should have vectors that are close to one another in the embedding space.

Trap An embedding is the natural-language completion returned to the user.

Choose an embedding model when text must be converted into vectors for semantic comparison, clustering, or similarity search. A chat model is the adjacent choice when the required output is newly generated natural language rather than a vector.

Trap Choose an image-generation model because it can create a visual representation of each text passage.

Chat models accept role-labeled message transcripts and return generated assistant messages. This format fits natural-language instruction following, question answering, and multi-turn text interaction.

Trap Choose sentiment analysis when the application must compose a new natural-language answer to a user's question.

Choose a multimodal model whose model card lists every required input type when a solution must reason over combinations such as text and images. A text-only model is not an appropriate choice merely because the desired response is text.

Trap Choose any text-output model because output modality alone determines whether images can be supplied as input.

An image-generation model synthesizes new visual content from a text prompt or an image-editing instruction. Its result is newly created image content rather than labels or coordinates for supplied pixels.

Use the Model leaderboard and Compare models experience in the Foundry portal to compare supported benchmark and performance information for candidate models, including cost comparisons exposed by the experience. Use each model card to confirm supported data types, features, deployment options, benchmark details, and license information.

Trap Use the evaluator catalog, which lists evaluation methods rather than comparing candidate model costs and capabilities.

Selecting a model in the Foundry portal model catalog does not by itself make the model callable. Unless the model supports instant access (preview), deploy the model to create an inference target, then use the deployment name when sending requests. With instant access, supported models can instead be called by model name without creating a deployment.

Trap Adding the model to a project collection automatically creates a callable inference endpoint.

Use standard deployment in a Foundry resource whenever the model supports it; Microsoft documents it as the preferred option with the broadest capabilities. It can support regional, data-zone, or global processing and standard or provisioned throughput choices, depending on the model.

Trap Use managed compute for every catalog model because dedicated virtual machines always provide more Foundry features.

A serverless deployment exposes a Microsoft-hosted model through an API and is generally billed for API input and output usage. Managed compute deploys model weights to dedicated managed virtual machines, requires compute quota, and is billed for compute uptime.

Trap Serverless deployment installs model weights on dedicated virtual machines in your subscription and bills by core-hour.

Configure Max Completion Tokens when a response must remain within a defined generated-token length or when generated-token cost must be bounded. The value is an upper bound that includes visible output and, for applicable models, reasoning tokens; it does not guarantee that the model will use the full amount.

Trap Configure Temperature, which changes sampling randomness rather than setting a response-length ceiling.

Lower temperature produces more focused and consistent output, while higher temperature increases randomness. Top P is an alternative nucleus-sampling control that limits consideration to tokens within a chosen probability mass; normally adjust Temperature or Top P, not both in the same request.

Trap Increase Max Completion Tokens to make token selection more random without allowing a longer answer.