01 AI Concepts and Responsible AI¶
Domain 1 of AI-901: Identify AI concepts and capabilities (40-45%). Three skill groups: responsible AI principles, model components and configuration, and recognising AI workloads.
Responsible AI: the six principles¶
Microsoft's six principles appear in nearly every Microsoft AI exam. Questions describe a concern and ask which principle it maps to. Learn the trigger words.
| Principle | The question it answers | Trigger words in a scenario | Example |
|---|---|---|---|
| Fairness | Does it treat all groups equitably? | bias, discriminate, one group approved less often | A loan model approves fewer applicants from one postcode with the same finances |
| Reliability and safety | Does it work as intended, including in unusual conditions, without causing harm? | edge cases, testing, failure, harmful output, rare conditions | A medical triage bot must be tested on rare symptoms before launch; content filters block harmful output |
| Privacy and security | Is personal data protected and the system secure? | personal data, consent, data retention, encryption, leaks | Customer recordings used for training only with consent; prompts not retained |
| Inclusiveness | Does it work for everyone, including people with disabilities? | accessibility, disabilities, languages, accents | A voice app must handle speech impairments and many accents; captions for deaf users |
| Transparency | Do people understand what it is and how it decides? | explain, disclose, understand why, users know it is AI | Telling users they are talking to an AI; explaining which factors drove a decision |
| Accountability | Who is responsible for it? | governance, oversight, human review, who answers for it | A named owner, an approval board, a human who can override the system |
Commonly confused pairs:
- Transparency vs accountability. Transparency is about understanding (explanations, disclosure). Accountability is about responsibility (people and processes that answer for the system).
- Fairness vs inclusiveness. Fairness is about outcomes not differing unjustly between groups. Inclusiveness is about access: designing so everyone can use it.
- Reliability and safety vs privacy and security. Harmful or wrong outputs are reliability and safety; data exposure and attacks are privacy and security.
In Foundry, these principles show up as concrete features: content filters (Azure AI Content Safety) on model deployments, evaluations for quality and safety, transparency notes published for each service, and access controls on projects. See Responsible AI basics on this site.
How generative AI models work¶
You need the concepts, not the maths:
- A large language model is trained to predict the next token (a word or part of a word) from the text before it. Trained on enormous amounts of text, it learns language, facts and patterns of reasoning. See LLM basics and How LLMs are trained.
- Most current models are transformers, which use attention to weigh every token against every other token in the context.
- Embeddings are lists of numbers representing meaning; similar meanings have nearby vectors. They power semantic search and retrieval.
- Multimodal models accept (and sometimes produce) images, audio and video as well as text.
- Models are non-deterministic and can hallucinate: produce fluent, confident, false statements. Grounding (supplying trusted data in the prompt, as in retrieval-augmented generation) reduces this.
- An agent is a model that can call tools (functions, search, code) in a loop to complete a task, rather than only answering.
Choosing a model by capability¶
Questions describe a need and ask for an appropriate kind of model. Think in capabilities, not brand names:
| Need | Appropriate model type |
|---|---|
| Chat, summarisation, drafting, Q&A | A general-purpose language model (a chat completion model) |
| Hard multi-step reasoning, maths, code | A reasoning model (spends more tokens "thinking"; slower and costlier) |
| Understand an image, chart or screenshot in a prompt | A multimodal (vision-capable) model |
| Answer a spoken question or respond in speech | A multimodal model with audio input/output, or Azure Speech plus a language model |
| Create an image from a description | An image-generation model |
| Semantic search, clustering, retrieval | An embedding model |
| High volume, low latency, low cost, simple task | A small or "mini" model |
The model catalog in Foundry lets you filter by capability, compare benchmarks and see deployment options. See Which model for which job and Small language models.
Deployment options and configuration parameters¶
Deployment makes a model callable from your project with an endpoint. Things to know:
- Deployment types affect where data is processed and how you pay: Global deployments route to Microsoft's capacity worldwide; Data Zone deployments keep processing within a geography (for example the EU or US); regional deployments stay in one region. Standard (pay per token) versus provisioned throughput (reserved capacity, predictable latency) is the other axis.
- You call a deployment by its deployment name, which you choose and which can differ from the model name.
Configuration parameters you should be able to explain:
| Parameter | What it does | Turn it up when | Turn it down when |
|---|---|---|---|
| Temperature | Randomness of word choice | You want variety (brainstorming, creative text) | You want consistent, factual answers |
| Top-p | Restricts choice to the most probable tokens adding up to p | Similar to temperature; usually change one or the other | |
| Max tokens (max output tokens) | Caps the length of the response | Long answers are needed | You want short answers or to cap cost |
| Stop sequences | Text that ends generation | ||
| System message / instructions | Sets role, rules, tone and format for the whole conversation | Always set one |
Some reasoning models do not accept temperature and use a reasoning effort setting instead.
Identifying AI workloads¶
The skill is reading a scenario and naming the workload. Clues:
| Workload | What it does | Scenario clues |
|---|---|---|
| Generative AI | Creates new text, code, images or audio | draft, summarise, chat, write, generate, create an image |
| Agentic AI | Plans and takes actions with tools across steps | book, update the system, multi-step, take action, call an API |
| Text analysis | Extracts meaning from existing text | sentiment, key phrases, entities, language detection, PII |
| Speech | Converts between speech and text | transcribe, captions, read aloud, voice, spoken |
| Computer vision | Understands images and video | detect objects, describe an image, read text in images (OCR), tag |
| Image generation | Creates new images | create artwork, product mock-ups from a description |
| Information extraction | Pulls structured fields out of documents, images, audio or video | invoice total, form fields, receipt, contract dates, call summary fields |
Text analysis techniques¶
- Keyword (key phrase) extraction: the main talking points ("battery life", "delivery time").
- Entity detection (named entity recognition): people, places, organisations, dates, quantities; entity linking connects them to a knowledge base entry.
- Sentiment analysis: positive, negative, neutral or mixed, at document or sentence level, often with confidence scores; opinion mining ties sentiment to aspects ("the screen is great, the battery is poor").
- Summarization: extractive (picks key sentences) or abstractive (writes new sentences).
- Also: language detection, PII detection and redaction.
Speech¶
- Speech recognition (speech to text): real-time or batch transcription, with features such as diarisation (who spoke) and custom vocabulary.
- Speech synthesis (text to speech): neural voices, SSML to control pronunciation, pauses and emphasis, and custom voices (with consent requirements).
- Speech translation: recognise and translate in one step.
Computer vision and image generation¶
- Image analysis: captions, tags, object detection (with bounding boxes), people detection, OCR (reading printed and handwritten text).
- Image classification vs object detection: classification labels the whole image; detection finds and locates each object.
- Face capabilities are restricted under Microsoft's responsible AI policy (limited access for identification).
- Image generation creates images from text prompts; content filters apply to prompts and outputs.
Information extraction¶
Taking unstructured content and producing structured fields: from documents and forms (invoices, receipts, IDs), images, audio (call transcripts plus fields such as reason for call) and video (scene descriptions, key moments). In Foundry this is Azure Content Understanding; see note 04.
Check yourself¶
- A hiring model rejects more applicants over 50 with equivalent qualifications. Which principle? (Fairness.)
- Users must be told they are chatting with an AI. Which principle? (Transparency.)
- You want consistent, repeatable answers for a policy Q&A bot. Which parameter do you lower? (Temperature.)
- "Extract the total, vendor and due date from 10,000 PDF invoices." Which workload? (Information extraction.)
- "Tell me whether each product review is positive or negative." Which workload and technique? (Text analysis; sentiment analysis.)