Server AtticLearning Harbour

AI Foundations · Module 2 · Lesson 1 of 3

Drafting and summarizing text

Two of the most common everyday AI tasks — producing a first draft and condensing a longer document — and the review step that makes both safe to actually use.

Lesson · 15–20 minutes · Text-first

By the end, you can

  • Explain why drafting and summarizing are a strong match for what generative AI does well (AF-2).
  • Identify what a person still has to check before sending an AI-drafted message or relying on an AI-produced summary (AF-1).
  • Give one example each of a safe drafting task and a safe summarizing task, naming the review step each one needs (AF-2).

Before you start

This is Module 2, Lesson 1 of the AI Foundations course. It assumes Module 1's vocabulary — rule, pattern, model, token — and Module 1 Lesson 2's point that a fluent AI answer is not automatically a correct one. This lesson does not require coding or technical setup.

Two everyday jobs: producing text and condensing text

**Drafting** means asking an AI tool to produce a first version of text from a short instruction — a reply, an invitation, an outline. **Summarizing** means asking it to condense a longer piece of text or information you already have into a shorter form. Google's own description of its Gmail drafting assistant puts drafting in exactly these terms: the tool can "generate a new email draft, such as a birthday invitation or an introduction to a potential business contact," and can "refine existing text in your draft for tone and clarity" — for example turning "a rough outline into a formal email." Anthropic's guide to summarizing legal documents describes the other job: using a language model "to efficiently summarize legal documents, extracting key information and expediting legal research."

Both jobs sit squarely inside what Module 1 Lesson 2 called **pattern continuation over tokens**. Anthropic's own description of Claude states plainly that it can "summarize text, answer questions, extract data, translate text, and explain and generate code" — summarizing and generating text named as core, everyday capabilities, not edge cases.

Why these are a safer fit than other task types

Recall Module 1 Lesson 2's distinction: a language model has no automatic lookup, so a task that depends on an external, current or private fact it cannot check is riskier than a task that stays entirely inside "continue this supplied text usefully." Drafting and summarizing usually stay on the safer side of that line, because the material the AI needs is already in front of it — your instruction for a draft, or the document you are asking it to summarize. Neither task typically requires the model to go and find a fact you have not given it. That is exactly why drafting and summarizing are often recommended as a good place to start with generative AI: the risk of "no automatic lookup" biting you is lower than it is for a task like answering "what is today's exchange rate."

The review step both still need

Lower risk is not no risk. Two specific things can still go wrong, even when every fact is supplied:

The Open Worldwide Application Security Project's generative-AI security effort defines hallucination as content that "seems accurate but is fabricated" — the same underlying mechanism from Module 1 Lesson 2, showing up here as an invented detail in a draft or a misrepresented point in a summary, rather than a wholly fictional fact. A fluent draft or a fluent summary is not proof that it correctly reflects your instruction or your source document — only checking it against that instruction or source is proof of that.

  • **Drafting can invent or embellish details that were never in your instruction.** A birthday invitation might add a time or venue detail you never specified. An email reply might imply a commitment — "we'll have this fixed by Friday" — that you did not authorise. Before sending, check every specific fact, name, date, number and promise against what you actually supplied.
  • **Summarizing can quietly drop a caveat, misstate a figure, or shift emphasis that was not in the source.** A five-bullet summary of a 12-page document is a compression; compression always involves choices about what to leave out, and a model's choices are not certain to match yours. Before relying on a summary, check that each bullet is actually traceable to a specific place in the source document.

A worked example

A hiring manager asks an AI tool to do two things. First: draft a short "thank you for your application" email using the applicant's name and the role title. Second: summarize a 12-page internal policy document into five bullet points for a team meeting.

For the first task, the manager checks three things before sending: is the applicant's name spelled correctly, is the role title exactly right, and did the draft imply any next step or timeline the manager did not actually authorise (for example, "we will be in touch within a week" when no such timeline was agreed)? None of these checks require re-reading the whole email — they are a short, specific list matched to what drafting characteristically gets wrong.

For the second task, the manager opens the source document and checks that each of the five bullets can be traced to a specific section, and that none of them changes the meaning of a clause by leaving out a condition or exception. This is a narrower check than reading the whole document again — it targets exactly what summarizing characteristically gets wrong.

Accessibility notes

This lesson is text-first, with no images, audio, video or downloadable artifacts. The practice exercise's model answer sits behind a native disclosure control that is reachable and operable by keyboard and correctly announced by screen readers. The knowledge check uses native radio-button inputs with a visible question and options, and posts its result to a live status region so assistive technology announces the outcome without a page reload.

Practice

Draft or summary: name the check

A colleague used an AI tool for two tasks this morning: (1) drafting a short LinkedIn message introducing themselves to a new contact, using only the contact's name and job title as input; (2) summarizing a 40-minute meeting recording into three action items for the team channel.

  1. For the LinkedIn message, name one specific fact or detail your colleague should check before sending it, and explain why that particular thing is the drafting task's characteristic risk.
  2. For the meeting summary, name one specific thing your colleague should check before posting it, and explain why that is the summarizing task's characteristic risk.
  3. Would you class either task as depending on a fact the AI would need to look up externally, using Module 1 Lesson 2's distinction? Explain your answer.
  4. Write one sentence you could say to the colleague about why 'it reads well' is not the same as 'it's ready to send'.
Compare with a bounded first version

For the LinkedIn message, check whether the draft added any specific claim not in the two supplied facts — for example, inventing a shared connection or a reason for reaching out that was never stated; that is drafting's characteristic risk of adding detail beyond what was supplied. For the meeting summary, check that each of the three action items is actually traceable to something said in the recording, including who is responsible and by when; that is summarizing's characteristic risk of shifting or dropping a detail during compression. Neither task depends on an external fact the AI would need to look up — both stay inside 'continue this supplied text usefully', using only the name/title supplied or the recording's own content. A fair sentence: reading well is a property of fluent text, not evidence that every name, date or commitment in it is exactly what you meant to say.

Knowledge check

Try the idea

An AI tool summarizes a 15-page supplier contract into five bullet points, and one bullet states a delivery deadline. What is the most accurate next step before this summary is used to make a decision?
Low-stakes practice only. This does not score, block progress or create a learner record.

Sources and limits

This lesson synthesises the sources below into a practical learning model. It is not a security standard, legal advice or a guarantee that any particular agent design is safe.

  1. "Help me write" in Gmail — Google. Describes Gmail's AI drafting feature generating a new email draft from a short instruction, such as an invitation or an introduction message.
  2. Legal summarization — Anthropic. A production guide describing how Claude is used to summarize legal documents, extracting key information and expediting legal research.
  3. Intro to Claude — Anthropic. Anthropic's own description of Claude's core text capabilities, including summarizing text and generating text.
  4. LLM09:2025 Misinformation — OWASP Gen AI Security Project. Defines hallucination as an LLM generating content that seems accurate but is fabricated — the basis for this lesson's review-step guidance.