1.1 Writing as a Process

Health sciences students write essays, lab reports, literature reviews and research proposals, and public health professionals write briefing notes, evaluation reports, grant applications, plain-language summaries and health communication materials. In all of these forms, writing does two jobs. It communicates with a reader, and it tests whether the writer's thinking holds together, since a gap in an argument often becomes visible only when the writer tries to put it into sentences.

General-purpose assistants such as ChatGPT, Claude, Gemini and Microsoft Copilot (examples available in 2026) have changed the first job more than the second. They produce fluent, well-organized text in seconds and restructure, condense and polish a draft faster than any human editor. They do not know what the writer has found or what a particular reader needs to decide. Lesson 2 explained why: a language model writes by predicting likely continuations of text, so it supplies the most typical version of whatever it is asked to write. The writer supplies the argument, the evidence and the judgement, and this lesson teaches workflows that combine the two contributions.

Box 6.1 introduces the Cedar Valley population health team, a fictional team whose loneliness program, AI assistant pilot and co-op student, Jordan, supply the examples used throughout this lesson.

Writing is usually described as a set of recurring activities more than a straight line: the writer plans, drafts, reads back, revises and edits, and returns to an earlier activity when a later one exposes a problem. Figure 6.1 shows six stages with the kind of help an assistant gives at each.

Table 6.1 sets out, for each stage, what only the writer can supply, what an assistant does well and what a careful writer watches for.

Table 6.1. Six stages of writing, with what the writer supplies, what an assistant does well and what to watch for at each stage.

Sections 1.2 to 1.6 take these stages in turn, beginning with the generation and organization of ideas.

1.2 Generating and Organizing Ideas

At the start of a writing task, the most useful thing an assistant can do is ask questions. A writer asked "What do you want the reader to do after reading this?" usually discovers that they know more than they had written down. An assistant can ask such questions patiently, one at a time, and then group the answers into themes, so that the ideas stay the writer's own.

Box 6.2 gives a prompt that asks an assistant to interview the writer, one question at a time, before any outline is written.

Jordan used this prompt for their course essay, and the questions brought out a point Jordan had not planned to make: in rural areas, a referral program can only be as good as the community groups and transport it refers people to. That point became the essay's third section. The prompt bars the assistant from proposing arguments for two reasons. Coursework exists to develop the student's own reasoning, and Lesson 4 showed that AI help that does the thinking weakens learning. An assistant's suggestions also tend toward the typical (Section 1.7). Lesson 7 teaches techniques for using AI to widen ideas when that is the goal.

Once the writer's ideas have been gathered and grouped, the next stage is to arrange them as an outline, the subject of Section 1.3.

1.3 Outlining

An outline is a plan of the claims a text will make, in order, with the evidence for each. Assistants propose structures well because they have absorbed the conventions of many genres: the background, gap and question of a research introduction (HSCI 207 Lesson 12), the issue, options and recommendation of a briefing note, and the problem and call to action of a health message. The writer makes sure the outline carries the writer's claim and rests on evidence the writer actually has, so the most useful outlining prompts supply the writer's own notes and ask for alternatives with reasons.

Box 6.3 is a prompt of this kind: it supplies the writer's numbered notes and asks for two alternative outlines, each tied to the notes, with reasons for preferring one.

Three checks apply to any outline an assistant returns. Is the thesis the writer's own claim, or a blander version of it? Is every section tied to evidence, which the note numbers make easy to see? Does the outline deal with the strongest objection to the thesis? The instruction to write EVIDENCE NEEDED guards against fabrication: an assistant asked to fill a gap will often supply a plausible study, and Lesson 5 showed how often such citations prove inaccurate or invented.

An outline that carries the writer's claim and marks where evidence is missing is the starting point for drafting, which Section 1.4 takes up.

1.4 Drafting

Drafting is the stage at which the choice of workflow matters most, because it determines whose sentences and whose thinking the final text contains. The right division of labour depends on the purpose of the task, the rules that apply to it and how much the writer needs to learn from doing it. The tabs describe four common arrangements.

Two further points apply to every arrangement. An assistant drafts best from facts it has been given: asked to "write a paragraph on the effectiveness of social prescribing", it may misstate the evidence and cite studies that do not exist, while given the writer's own summary of the evidence it is far more dependable. The rules that apply to a piece of writing also decide which arrangements are available: the instructor's rules in a course, the publisher's policy in a journal and the organization's policy at work. Section 3 sets out these rules.

Once a full draft exists, the work shifts from producing text to improving it, and Section 1.5 begins with a method for seeing the structure the draft actually has.

1.5 Reverse Outlining and Revising

A reverse outline is an outline made from a finished draft: a list of the main point of each paragraph, in order. It shows the structure the draft actually has, which often differs from the one the writer intended. Reverse outlining is widely taught in university writing centres and tedious by hand, which makes it an excellent task for an assistant.

Box 6.4 gives a prompt that asks an assistant for a reverse outline of a draft essay, with safeguards against summaries that tidy up a muddled paragraph.

The prompt contains two safeguards. Quoting the sentence that carries each main point lets the writer check the summary against the text. The NO CLEAR POINT instruction counters a known tendency: an assistant summarizing a muddled paragraph often reports the point the paragraph should have made, which hides the problem a reverse outline is meant to reveal.

Revision works best from the largest problems down. Global revision deals with the argument, structure and evidence, and local revision with sentences, word choice and grammar. Editing sentences that a global revision later deletes wastes time, so skilled users ask for global feedback first and editing last. Section 2 teaches the feedback prompts in detail.

Box 6.5 presents one paragraph of Jordan's essay with a summary that an assistant produced without these safeguards, and asks the reader to judge whether the summary describes the paragraph accurately.

Reverse outlining and global revision deal with the argument and the structure of a draft. Editing at the level of the sentence comes last, and Section 1.6 turns to it.

1.6 Editing for Clarity and Grammar

Editing is the task at which assistants are most consistently strong. They correct grammar and punctuation, shorten wordy sentences, make terms consistent and flag ambiguous references in seconds, and grammar checkers built into word processors offer a narrower version of the same help. The main design decision in an editing prompt is how much control the writer keeps. "Improve this paragraph" invites free rewriting, which can change the meaning, remove hedges and replace the writer's voice. Asking for each change to be listed with a reason, or for grammatical corrections only, keeps the writer in charge, and a document-comparison view makes every change visible. Section 2 develops these practices.

With the six stages described, Section 1.7 turns to the experimental evidence on how AI assistance changes the time writing takes, its quality and its variety.

1.7 What the Experiments Show

Two randomized experiments give the clearest evidence so far on how AI assistance changes writing. Noy and Zhang (2023) ran an online experiment with 453 college-educated professionals, including marketers, grant writers, consultants, data analysts, human resources professionals and managers. Participants completed realistic writing tasks of 20 to 30 minutes from their own occupations, such as cover letters for grant applications, emails about an organizational restructuring and plans for data analyses, and a randomly selected group was given access to ChatGPT for the second task. Access to the assistant cut the time taken by about 40 percent and raised the quality of the work, as rated by independent evaluators, by about 18 percent. The gains were largest for participants whose first task had received low grades, so the gap between weaker and stronger writers narrowed.

Doshi and Hauser (2024) asked 300 people to write a short story of eight sentences for young adults. One group wrote without help, a second could see one story idea generated by an assistant, and a third could see up to five such ideas. About 600 evaluators then rated the stories. Stories written with access to AI ideas were rated as more novel and more useful (better suited to their audience and more likely to be published) than stories written without them, and the gains were concentrated among the writers who had scored lowest on a test of inherent creativity. The stories written with AI ideas were also more similar to one another than the stories written without help. The authors described this as a social dilemma: each writer is better off using the assistant, while the collection of stories as a whole becomes less varied.

The flip cards that follow summarize Noy and Zhang (2023), Doshi and Hauser (2024) and what the tasks in the two experiments had in common.

Three implications follow. The gains are real for the parts of writing that concern form (organization, fluency, grammar and tone), so a writer who has the content can expect substantial help. The larger gains for weaker writers make assistance a meaningful aid to equity, including for writers working in a second language (Section 2), provided the writer keeps control of the content. The convergence that Doshi and Hauser found is a predictable result of how language models write: a model that predicts likely text produces the most typical version of a story or argument, so writers who start from AI suggestions start from the same place as everyone else.

The practical response is to put the writer's own thinking first. Writers who draft their thesis and key points before consulting an assistant, ask for several alternatives and use the assistant mainly to critique and polish keep the distinctive parts of their work, and they stay accountable for the content, which is where every journal, funder and university places responsibility (Section 3). Section 2 turns to feedback on a draft the writer has already written, where assistants add the most value with the least risk.

Before that, the reflection and knowledge check that follow review the stages of writing and the evidence from the two experiments.

2.1 Asking for Feedback That Helps

Feedback on a draft combines the most value with the least risk of any use of AI in writing. The writer has already done the thinking, the assistant reads the whole draft in seconds and comments without tiring, and the writer decides what to change. The quality of the feedback depends heavily on the request. "Is this good?" usually returns general praise and a few safe suggestions, because assistants trained on human preferences tend toward sycophancy, the tendency to agree with and flatter the user (Lesson 3; Lesson 7 returns to it in depth).

Three features make a request useful. The first is explicit criteria: a rubric that defines what a good version of the document does, such as a funder's evaluation criteria, a journal's guidance for reviewers, an instructor's published criteria for an essay or a team's standard for briefing notes. Asking the assistant to rate the draft against each criterion and to quote the passage behind each rating ties every comment to the text. The second is a defined reader (Section 2.2). The third is a set of constraints, such as "do not rewrite" and "do not add facts or references", that keep the content in the writer's hands. Cycle 1 of Worked Example 6.1 shows a complete rubric prompt.

The second of these features, a defined reader, is the subject of Section 2.2.

2.2 Reading as a Critical Reviewer or a Named Audience

Writers find it hard to read their own work as a stranger would, because they know what they meant. An assistant can take two kinds of reader's view. As a critical reviewer, it looks for unsupported claims, gaps in logic, missing counterarguments and conclusions that go beyond the evidence; a good request names the reviewer's standard ("a peer reviewer for a public health journal who is skeptical of before-and-after designs"), asks for the three most serious problems in order and asks for no praise. As a named audience, it reads as a particular person with a particular need, such as an executive deciding on a budget or an older adult deciding whether to join a program. Describing that reader's knowledge, time, concerns and decision, and asking what they would ask after reading, exposes what the draft fails to tell them.

Both requests simulate a reader, and the simulation has limits. An assistant playing an executive draws on general patterns of what executives care about and cannot know the priorities of a particular executive in a particular year. It is a rehearsal that improves a draft before a real reader sees it, a point Lesson 7 develops for interest holder engagement. Cycle 2 of Worked Example 6.1 shows a complete named-audience prompt.

Feedback from either kind of reader leads to changes in the draft, and Section 2.3 describes how to compare alternative versions and decide on each change.

2.3 Comparing Versions and Deciding on Each Change

Assistants make alternatives cheap. A writer unsure about an opening can ask for three versions that lead with different elements (the finding, the problem, the decision required), each with a sentence on what it gains and loses, and then choose or combine them against a stated purpose.

The discipline that makes AI revision dependable is to decide on each change separately. Accepting a whole rewritten draft at once is the most common way errors enter AI-assisted writing, since a fluent rewrite can change a number, drop a hedge or add a claim with no visible sign. Writers who use assistants well ask for changes as a list or table, or compare versions in a document-comparison view, and check each change for five things: meaning, numbers, hedges, added claims or citations, and voice. For consequential documents, a short change log recording each suggestion, decision and reason makes the process reviewable and supports disclosure (Section 3).

The next subsection turns to language editing for writers working in a second language.

2.4 Writing in a Second Language

Many health sciences students and researchers in Canada write in a language that is not their first, and AI has made expert-level language editing available to all of them. The federal research funding agencies' guidance on AI in grant proposals (2024) notes that generative AI has the potential to assist non-native English and French speakers, and some publishers exempt this kind of copy editing from AI disclosure (Section 3). The prompt decides whether the editing also builds skill. "Fix my English" returns a polished paragraph, teaches nothing and may replace correct but uncommon word choices with typical ones. A prompt that limits the edit to grammar and idiom and explains each correction gives the writer a better text and a lesson.

Box 6.6 is a prompt of this kind, written for a health sciences student whose second language is English.

Writers in this position should also know that tools claiming to detect AI-generated text misclassify writing by non-native English speakers as machine-written far more often than writing by native speakers (Liang and colleagues, 2023), a finding Section 3 presents with SFU's guidance on such tools.

Language editing applies general rules of grammar and idiom. Section 2.5 turns to the rules that an organization sets for its own writing, recorded in a house style guide.

2.5 House Style and Stock Phrasing

A house style guide is a written set of rules that an organization applies to everything it publishes: spelling, preferred terms, punctuation, tone and phrases to avoid. Journals, government departments, health authorities and news organizations all keep one. A style guide is an excellent thing to give an assistant, because it turns "write well" into specific rules the assistant can apply and the writer can check. Most assistants in 2026 accept standing instructions or project-level files (Lesson 3), so a team can supply its guide once.

The lessons in this series are written to a house style guide, and Figure 6.2 reproduces its main rules as a worked example. Several rules target patterns that became common in AI-generated text: a list of banned words, a ban on the long dash used to splice clauses, a ban on constructions that deny one idea in order to assert another (negation-correction constructions), and a ban on stock paragraph openers and comments about the writing itself.

Stock phrasing means words and constructions that appear far more often in AI-generated text than in careful human writing, so that readers have learned to associate them with unedited AI output. Kobak and colleagues (2025) measured the shift in about 15 million PubMed abstracts published from 2010 to 2024. Earlier vocabulary shifts, such as the one during the COVID-19 pandemic, involved nouns that describe content, while the shift after language models became widely available involved mostly verbs and adjectives that describe style. Words with marked excess use included "showcasing", "potential", "findings" and "insights", and the authors estimated that at least 13.5 percent of 2024 abstracts had been processed with a language model.

Stock phrasing adds words without information, replaces specific claims with vague ones and makes texts sound alike, which is the convergence Doshi and Hauser (2024) observed in stories. Table 6.2 lists common categories, with examples from outside the series' banned list.

Table 6.2. Common categories of stock phrasing, with examples of each and the fix.

Three practices make an assistant useful for this editing. Ask it to flag problems against the guide without rewriting, in a table that quotes each sentence, names the rule, proposes a replacement and states whether the replacement changes the meaning, because a free rewrite "in plain style" removes some stock phrases and introduces others. Have it check its own earlier output, since phrasing it introduced in a rewrite is the phrasing most likely to need removal. Finish with the writer's own read, because an assistant applies rules literally and may flag a necessary technical term or remove a hedge that the evidence requires. Editing out stock phrasing improves clarity; whether AI use must be disclosed depends on the rules in Section 3, whatever the phrasing.

Worked Example 6.1 brings the practices of this section together in one case. Jordan runs three cycles of feedback on the Cedar Valley briefing note: against the team's rubric, with the assistant reading as the executive who will decide, and against the house style guide. Table 6.3 records the decision on each suggestion with its reason.

This section has shown how explicit criteria, a defined reader, a comparison of versions and a house style guide make an assistant a useful source of feedback, provided the writer decides on each change. The reflection and knowledge check that follow review these practices, and Section 3 then sets out the rules on authorship and disclosure that apply to AI-assisted writing.

3.1 Authorship and Accountability

Authorship in academic and professional writing carries two things together: credit for the work and accountability for it. An author can be asked to explain a method, correct an error, answer a critic or retract a claim. This pairing explains the position that journals, editors' organizations and funders have taken on AI. An assistant can contribute a great deal to a text, but it cannot be held to account for it, so it cannot be an author, and the people who are authors answer for everything the text contains, including the parts an assistant produced.

The International Committee of Medical Journal Editors (ICMJE), whose recommendations many health journals follow, states that chatbots such as ChatGPT and other AI-assisted tools should not be listed as authors, because they cannot be responsible for the accuracy, integrity and originality of the work. Authors who use AI are asked to describe that use both in the cover letter and in the appropriate section of the submitted work. Humans are responsible for any submitted material that involved AI, and authors must be able to assert that there is no plagiarism in their paper, including in text and images produced by AI. The recommendations also state that referencing AI-generated material as a primary source is not acceptable, and that failure to disclose AI use may lead to corrective action.

The Committee on Publication Ethics (COPE), which advises journal editors and publishers on ethical practice, published a position statement on authorship and AI tools on 13 February 2023. It states that AI tools cannot meet the requirements for authorship because they cannot take responsibility for the submitted work, and that as non-legal entities they cannot declare conflicts of interest or manage copyright and licence agreements. Authors who use AI tools in writing a manuscript, producing images or graphical elements, or collecting and analyzing data must disclose which tool was used and how, in the Materials and Methods section or a similar part of the paper. Authors remain fully responsible for the content of their manuscript, including the parts produced by an AI tool, and are liable for any breach of publication ethics.

Individual publishers add their own detail, and the details differ. Springer Nature's policy, for example, states that large language models do not satisfy its authorship criteria and that their use should be documented in the Methods section. It also states that AI-assisted copy editing, defined as AI-assisted improvements to human-generated text for readability and style and to correct grammar, spelling, punctuation and tone, does not need to be declared, while generative editorial work and autonomous content creation are outside that exemption. Its journals generally do not accept AI-generated images. A writer preparing a manuscript therefore reads the target journal's own policy, since an edit that one publisher exempts may require a statement at another.

Table 6.4 summarizes the frameworks described in this section, including the tri-agency guidance and the SFU guidelines discussed in Sections 3.2 and 3.3, and sets out for each whether AI may be an author, what is disclosed and where, and who is accountable.

Table 6.4. Journal, publisher, funder and university rules on AI authorship, disclosure and accountability.

HSCI 207 Lesson 12 shows where such statements go in a research report, and HSCI 241 Lesson 6 covers the reporting of AI use in systematic and scoping reviews, where additional recommendations apply.

Journals and publishers set the rules for published work. The next subsection turns to the rules that Canada's federal research funding agencies apply to grant proposals.

3.2 AI in Research Grant Proposals

Canada's three federal research funding agencies, the Canadian Institutes of Health Research (CIHR), the Natural Sciences and Engineering Research Council (NSERC) and the Social Sciences and Humanities Research Council (SSHRC), published joint guidance on the use of artificial intelligence in the development and review of research grant proposals in November 2024, after an expert panel's advice in December 2023 and a public consultation that closed in June 2024. Students meet this guidance when they help prepare grant applications as research assistants.

For applicants, the guidance recognizes that generative AI may be a valuable tool, with the potential to improve efficiency, to assist non-native English and French speakers and to streamline proposal writing. It holds the named applicant ultimately accountable for the complete contents of the application, which must be true, accurate and complete, with all sources appropriately acknowledged and referenced. It warns that generative AI can present information without proper recognition of authorship, which places the burden of checking sources on the applicant. For reviewers, the rule is strict. Uploading application material into online AI tools could breach privacy and lose custody of intellectual property, since some tools may store and reuse what they receive, and the use of publicly available online tools to evaluate grant applications is prohibited. The guidance adds that such use would place a reviewer in breach of the agencies' conflict of interest and confidentiality agreement for reviewers.

The reviewer rule reflects a principle that applies well beyond grant review. A draft written by colleagues, an unpublished manuscript under peer review and a collaborator's research idea are confidential material belonging to other people. Lesson 3 taught students to classify information before entering it into any tool, and these documents belong in the confidential class whatever tool is used.

The tri-agency guidance governs grant proposals. For coursework and theses at SFU, the university's own guidelines set the expectations, which Section 3.3 describes.

3.3 SFU's Expectations for AI Use

SFU's approach is set out in its Learning and Teaching with AI Guidelines (November 2025) and in companion guidelines for students and graduate students (2025). The central principle is that instructors decide whether, when and how AI may be used in their courses, and communicate their expectations in the syllabus and in each piece of work they set. The student guidelines ask students to follow the AI-use guidelines stated by their instructor, to ask the instructor before using AI tools when those guidelines are unclear or absent, and to disclose and cite AI-generated content transparently in the way the instructor directs. They also ask students to keep records of their prompts, the outputs and the changes they made, to verify AI output because the tools are known to produce inaccurate information, to use tools their instructor has approved, and to keep personal, confidential and sensitive data out of AI applications that have not been vetted. Students remain responsible for avoiding plagiarism and copyright violations. Graduate students consult their supervisor and supervisory committee about AI use in research and disclose it in the acknowledgements or methods of a thesis or article.

The guidelines are equally clear about AI-text detectors. The student guidelines advise students not to rely on AI detectors because these tools can be unreliable and biased, and state that AI detectors cannot be used for grading decisions or academic misconduct investigations. The guidance for instructors gives the same direction, describes detectors as unreliable and biased, and asks instructors who have concerns about AI use to talk with the student before reporting a concern. The next section presents some of the evidence behind this position.

3.4 Why AI-Text Detectors Misclassify Writing

An AI-text detector is a tool that estimates whether a passage was written by a person or generated by a language model. Many detectors rely on some version of perplexity, a measure of how predictable a text is to a language model. Because a language model generates text by choosing likely words, its output tends to have low perplexity, and detectors treat low perplexity as a sign of machine authorship. The difficulty is that human writers can also produce predictable text, especially writers who use a smaller range of vocabulary and sentence structures.

Liang and colleagues (2023) tested this directly in a study published in Patterns. They ran seven widely used detectors on 91 essays written by non-native English speakers for the TOEFL (Test of English as a Foreign Language) and 88 essays written by eighth-grade students in the United States. The detectors classified the American essays almost perfectly, while they labelled the TOEFL essays as AI-generated in 61.2 percent of cases on average. Eighty-nine of the 91 TOEFL essays were flagged by at least one detector, and 18 were flagged by all seven. The authors then used ChatGPT to change the word choices in the essays. When the TOEFL essays were rewritten with more varied vocabulary, the average false-positive rate fell to 11.8 percent; when the American essays were rewritten with simpler vocabulary, the share misclassified as AI-generated rose from 5.2 percent to 56.7 percent.

Table 6.5 sets these four results side by side.

Table 6.5. Average share of essays classified as AI-generated by seven detectors, for TOEFL and US eighth-grade essays as written and after ChatGPT changed their vocabulary (Liang and colleagues, 2023).

The pattern shows that these detectors responded mainly to the range of vocabulary in a text, which is a poor proxy for who wrote it. It also produced a perverse result: in this study, running a human essay through an AI tool made it less likely to be flagged. Detectors have changed since 2023, and a later generation may perform differently, but the mechanism the study exposed is a property of the perplexity approach itself. The finding supports SFU's position that detector output cannot carry decisions about grades or misconduct, and it has a particular bearing on fairness for writers working in a second language, whose legitimate use of language editing was discussed in Section 2.

Box 6.7 states what this evidence implies for writers who use AI.

The honest course for a writer is therefore to disclose AI use as the relevant rules require, and Section 3.5 describes how to write a disclosure statement.

3.5 Writing a Disclosure Statement

A disclosure statement tells readers which AI tools were used in producing a piece of work, for what, and how the output was checked. A good statement answers five questions in a few sentences: which tool was used (with a version or the dates of use); which tasks it was used for; what the writer did with its output; how factual content was verified; and who takes responsibility for the result. It goes where the relevant rules say: in the place the instructor specifies for coursework, in the cover letter and the Methods or acknowledgements for a manuscript, and in the location an organization's policy sets for a report. Lesson 3 showed how to cite an AI tool in APA style and how to keep an AI use log, and the log is the natural source for the statement.

The tabs that follow give an example statement for each of three settings: coursework, a journal manuscript and a professional report.

Box 6.8 gives a prompt that drafts such a statement from an AI use log.

The CHECK MISSING instruction does two jobs. It stops the assistant from inventing a plausible verification step, and it shows the writer where the log, or the checking itself, has a gap.

Box 6.9 asks the reader to judge three disclosure statements against the requirements described in this section.

Some uses of AI fall into grey areas in which the rules differ between courses, journals and publishers. The accordion that follows discusses four of them: grammar checkers, translation, summarizing sources, and figures and images.

The common thread in these rules is accountability that stays with people. Assistants can take on a large share of the labour of writing, and every framework described here permits that, provided the people whose names appear on the work have checked it, stand behind it and say how it was made. Section 4 applies these practices to the documents that public health teams write for the public and for decision-makers.

Before that, the reflection and knowledge check that follow review the rules on authorship, grant proposals, AI-text detectors and disclosure.

4.1 The Writing Tasks of Public Health Practice

Much of the daily work of public health is writing for people who are not specialists: decision-makers who need a recommendation, community members who need to understand a service and funders who need a case. These documents are often routine in form and high in consequence, the combination in which AI assistance is most valuable and checking most necessary. Lesson 9 surveys the routine tasks in which public health teams already use AI and the rules that govern that use; this section concentrates on the writing.

Table 6.6 sets out the main documents, with what an assistant does well for each, what the writer supplies and the check that applies.

Table 6.6. Main writing tasks of public health practice, with what an assistant does well, what the writer supplies and the check that applies to each.

The subsections that follow take up these documents in turn, beginning with briefing notes, grant applications and reports.

4.2 Briefing Notes, Grant Applications and Reports

A briefing note is a short document, usually one or two pages, that informs a decision-maker about an issue and often asks for a decision. Formats vary between organizations, but most place the purpose and the recommendation first and then give the background, the current situation, the considerations (evidence, cost, equity and risk), the options and the next steps. Executives read the first lines and skim the rest, so the hardest writing task is deciding what goes at the top. Jordan's note from Section 2, in the version that emerged from three cycles of feedback, shows this structure.

Grant applications combine persuasive and technical writing under strict length limits. Assistants help to check a draft against the funder's published criteria, cut a section to its limit and draft a lay summary, while the named applicant stays accountable for every claim and citation and a collaborator's unpublished ideas go only into tools the team's rules permit (Section 3). In evaluation and surveillance reports, an assistant can draft routine text from a table quickly, and every number in that text is traced back to the table, because a fluent sentence can carry a transposed figure.

The next subsection turns to documents written for community members, for whom plain language and health literacy decide whether a text can be understood and used.

4.3 Plain Language, Health Literacy and Readability

Plain language is writing that its intended readers can find, understand and use the first time they read it. Health literacy is the ability to find, understand, judge and use health information to make decisions, and it varies widely within any population. Plain language matters in public health because the people who most need a service, such as older adults, people with less formal education and people reading in a second language, are often those for whom technical writing works least well.

Assistants are very good at the mechanics of plain language: shorter sentences, everyday words, defined terms and active verbs. They are less dependable at keeping the meaning intact while simplifying. Simplification invites three errors: claims become stronger than the evidence, limits disappear and details are invented to make the text concrete. A good plain-language prompt therefore names the audience and reading level, requires every number and limit to be kept, forbids new facts and asks for a checkable list of claims.

Box 6.10 is a prompt written to these requirements for older adults and their families in a rural part of British Columbia.

A readability formula estimates how difficult a text is from features that can be counted. One of the most widely used in health materials is the Flesch-Kincaid grade level, developed by J. Peter Kincaid and colleagues in 1975 from Rudolf Flesch's earlier reading-ease formula. It combines average sentence length with average word length in syllables and expresses the result as a school grade. Readability studies of patient education materials commonly use a target of about grade 6 to 8, and they frequently find that published materials exceed it.

Box 6.11 gives the formula and works it through for the corrected summary in Worked Example 6.2.

Readability formulas measure length, which is only one source of difficulty. A short word can be unfamiliar ("stroma") and a long one familiar ("everybody"), and the formulas ignore accuracy, the presentation of numbers, layout and cultural fit. Calculators also count syllables and sentences differently, so one text can receive different scores. A readability score is a screening check, and testing a draft with members of the intended audience is the stronger test. Interactive 6.1 applies the formula and flags long sentences and common stock phrases.

Worked Example 6.2 continues the Cedar Valley demonstration from Worked Example 6.1. It follows Jordan from the team's conference abstract to a plain-language summary, and Table 6.7 compares the assistant's draft with the abstract claim by claim.

The corrected summary then becomes the source for further versions of the message, each written for a different audience, which Section 4.4 considers.

4.4 Adapting a Message for Different Audiences

One finding usually needs several versions. The Connector program's first-wave results matter to clinic staff who make referrals, to older adults and families deciding whether to take part, to municipal partners who run community groups and to executives who fund the program. Each version emphasizes different facts and calls for a different action, while the facts themselves stay identical. Assistants produce such variants quickly (HSCI 312 Lesson 8 covers the campaign planning that decides which audiences and channels a message needs), and each rewrite is a new chance for a number to change or a limit to disappear. A prompt that asks for a consistency table makes such drift visible.

Box 6.12 gives such a prompt for four audiences of the Connector program.

Some audiences also need the message in their own language, and Section 4.5 turns to machine translation and the review it requires.

4.5 Machine Translation and Community Review

British Columbia's communities speak many languages. In the 2021 Census, the most common non-official languages spoken at home in the Vancouver area were Mandarin, Punjabi and Cantonese, and the First Peoples' Cultural Council reports 34 First Nations languages spoken in the province. Health materials in a person's own language improve access, and machine translation, now built into general assistants as well as dedicated tools such as DeepL and Google Translate (examples available in 2026), has made first drafts of translations fast and cheap.

The quality of those drafts varies by language and by text. Khoong and colleagues (2019) tested Google Translate on emergency department discharge instructions and found an accuracy of about 92 percent for Spanish and 81 percent for Chinese, and some of the errors could have caused harm. Translation tools have improved since then, and quality still tends to be lower for languages with less text available for training. General translation tools offer little or no support for most First Nations languages of British Columbia, and decisions about translating material into them belong with the language communities themselves; Lesson 10 discusses the principles of Indigenous data sovereignty that apply to such work.

Machine translation is therefore a first step in a process that ends with people. Figure 6.3 shows the process the Cedar Valley team uses for its program leaflet in Punjabi and Traditional Chinese, the two non-official languages most often requested at its clinics.

Step 3 uses an assistant to screen the translation before the human review, so that the fluent reviewer's time goes to judgement. A back-translation renders the translated text back into the source language so that a reader of the source language can compare meanings. When the same kind of model produces both the translation and the back-translation, its errors can cancel out and hide a problem, which is one reason the fluent reviewer is still needed.

Box 6.13 gives a prompt for this screening step, written for the Punjabi version of the leaflet.

Box 6.14 presents one sentence from such a screen and asks the reader to identify each change in meaning and judge how serious it is.

Two rules from earlier lessons complete the process. Only public or non-confidential text goes into a translation tool (Lesson 3), and the use log records which tool produced the draft, who reviewed it and what changed.

Box 6.15 draws the lesson together. Table 6.8 sets out, for six writing tasks, how AI performs and the check that applies, and the box ends with a short statement of what an assistant can and cannot do for a writer.

This section has applied the lesson's practices to the documents that public health teams write for decision-makers and the public, from briefing notes and plain-language summaries to translated leaflets. The reflection and knowledge check that follow review these practices.

Bringing It All Together

This lesson followed AI assistance through the whole of a writing task. Assistants can interview a writer before an outline exists, propose structures from the writer's notes, produce a reverse outline in seconds, give feedback against explicit criteria or as a named reader, and edit for grammar, concision and house style. Experiments with professionals and with story writers show real gains in speed and rated quality, largest for weaker writers, together with a drift toward typical, similar text that writers counter by putting their own thinking first.

The same features that make assistants useful make checking necessary. In the Cedar Valley demonstration, the assistant found a buried recommendation and an overstated result, and it also added a fabricated reference, proposed a wrong and causal figure, introduced stock phrasing and mislabelled a change of meaning. Each failure was caught by a specific check: a database search, recomputation, a style-guide review and the writer's own reading. The plain-language summary showed that a readability score can be excellent while the content contradicts its source.

Responsibility stays with people throughout. The ICMJE recommendations, the COPE position and publisher policies exclude AI from authorship, the tri-agency guidance holds applicants accountable and bars reviewers from public AI tools, and SFU's guidelines leave the rules for each course to its instructor while ruling out AI detectors for grading and misconduct decisions. Translated and plain-language materials reach the public only after fluent speakers and members of the intended audience have read them.

The final reflection asks you to design an AI-assisted workflow for a writing task you expect to do, drawing on all four sections.

Final Knowledge Assessment

These terms, frameworks and researchers appear in Lesson 6; links are provided where a stable source exists.