# Lesson 6: AI-Assisted Academic and Professional Writing *Companion-podcast transcript • Sarah & Kiffer* *Office Hours episode to listen to after working through the lesson* --- **Sarah:** In one study of short stories, the writers who scored lowest on a test of inherent creativity gained the most from seeing story ideas from an assistant. **Kiffer:** Which sounds like good news. The same study found that the stories written with AI ideas were more similar to one another. **Sarah:** The authors called it a social dilemma. Each writer does better with the assistant, and the collection as a whole gets less varied. **Kiffer:** And Kobak and colleagues saw a version of this in about fifteen million PubMed abstracts. They estimated that at least thirteen and a half percent of the abstracts from 2024 had been processed with a language model. **Sarah:** Welcome back to Office Hours. I'm Sarah. **Kiffer:** And I'm Kiffer. This is Lesson six, AI-Assisted Academic and Professional Writing, and we're starting with what the writing experiments show. **Sarah:** The headline numbers are easy to remember. In Noy and Zhang's experiment, professionals with access to ChatGPT cut the time they took by about forty percent, and independent evaluators rated their work about eighteen percent higher. My question is how far those numbers travel. Would a public health team writing a briefing note see the same thing? **Kiffer:** It helps to look at what the four hundred and fifty-three participants were asked to write. These were realistic tasks from their own jobs, such as a cover letter for a grant application or an email about an organizational restructuring, and each took twenty to thirty minutes. They were short and self-contained, and they didn't test fact-checking. So the experiment measures what an assistant does for fluent drafting, which is a large part of writing. It says much less about the hours a writer spends gathering figures and checking them. **Sarah:** Let me put a number on that with a made-up example. Say a briefing note takes you five hours. Three hours go to drafting and polishing, and two go to pulling figures from tables and checking them. If the assistant cuts the drafting part by forty percent, as in the experiment, how much shorter is the whole job? Take a few seconds. *(Pause)* **Kiffer:** Forty percent of three hours is 1.2 hours, or seventy-two minutes. The job falls from five hours to 3.8 hours, which is a saving of about a quarter. And if reading the assistant's sentences against your outline adds half an hour of checking, the saving drops to forty-two minutes, or about fourteen percent. **Sarah:** That's still a real saving. **Kiffer:** It is, and I'd take it. The forty percent applies to the part of the work the experiment measured. When you plan a writing task, it's worth asking which part the assistant speeds up and which part stays with you. **Sarah:** The other result I keep coming back to is who gained. In Noy and Zhang, the participants whose first task got low grades gained the most, so the gap between weaker and stronger writers narrowed. **Kiffer:** And Doshi and Hauser found something similar with short stories. The writers who scored lowest on a test of inherent creativity gained the most from seeing story ideas from an assistant. In a field where many students and researchers write in a second language, I think that's an equity gain worth taking seriously, as long as the writer keeps control of the content. **Sarah:** But the same study found that the stories written with AI ideas were more similar to one another. The authors called it a social dilemma. Each writer does better with the assistant, and the collection as a whole gets less varied. Kobak and colleagues saw a version of this in about fifteen million PubMed abstracts. They estimated that at least thirteen and a half percent of the abstracts from 2024 had been processed with a language model, and the sign was a rise in verbs and adjectives that describe style. **Kiffer:** Does that sameness worry you for public health writing in particular? **Sarah:** It depends on the document. I don't mind if every briefing note in a health authority follows the same structure, with the decision first and then the background and options. Readers benefit from that. What worries me is sameness of content. If thirty students start their essays from the assistant's ideas, you get thirty versions of the most typical argument, and the point that only one of them would have made never gets written. **Kiffer:** I agree with that distinction. Convergence on structure is mostly harmless and often helpful. Convergence on ideas is the real cost, and a simple habit limits it. Write down your own claim and your key points before you open the assistant, and then use it to question and critique what you've written. **Sarah:** Question two. If an assistant tells you its edit didn't change the meaning, why isn't that enough? **Kiffer:** Let's work an example, and this one is made up. Suppose a report on a hypothetical school flu vaccine program contains two sentences. In the twelve schools that sent reminder letters home, forty-one percent of students got a flu vaccine, compared with thirty-six percent in the twelve schools that didn't send letters. Schools chose whether to send the letters, so part of this difference may reflect other differences between the schools. **Sarah:** And the writer asks the assistant to make it more concise. **Kiffer:** Right, and the assistant returns one smooth sentence. Reminder letters raised flu vaccine uptake by five percent, from thirty-six to forty-one percent. It also reports that the meaning is unchanged. So here's the question for everyone listening. How many changes in meaning can you find in that one sentence? Take a few seconds. *(Pause)* **Sarah:** I count four. First, saying the letters raised uptake claims that they caused the difference. Second, the sentence about schools choosing for themselves has disappeared, so the hedge is gone. Third, from thirty-six to forty-one sounds like a change over time in the same schools. The original compared two different groups of schools at one point in time. **Kiffer:** That third one is easy to miss. And the fourth? **Sarah:** The five percent. Forty-one minus thirty-six is five percentage points. As a relative change, five divided by thirty-six is about 0.139, so uptake was about fourteen percent higher in the schools that sent letters. Five percent is wrong under either reading. **Kiffer:** Exactly. Notice that thirty-six and forty-one both survived, so a check that only looks for the original numbers would pass this sentence. The lesson's five checks are meaning, numbers, hedges, added claims or citations, and voice, and they catch all four. The causal verb is an added claim, the missing sentence is a lost hedge, the change over time is a change of meaning, and the five percent is a number with a new meaning. **Sarah:** So why would the assistant say the meaning was unchanged? **Kiffer:** A model checking its own edit tends to share the blind spots that produced the edit. If it read the phrase raised uptake as a fair summary when it wrote the sentence, it will probably read it the same way when it checks. There may also be some sycophancy at work. Assistants trained on human preferences lean toward agreeable answers, and telling you that nothing changed is the more agreeable answer. A self-check is useful for locating the claims in a text, and the judgement about each claim stays with you. **Sarah:** What should the prompt have said in the first place? **Kiffer:** Something like this. Shorten these two sentences by about a third. Keep every number and every limit, don't add any claim about cause, and list each change you make with a one-line reason. That gives you a list you can check item by item. The assistant may still slip, and the list shows you where to look. **Sarah:** And for a report that matters, the slip goes in the change log. The entry might read: suggestion rejected; it turned a comparison between schools into a causal claim, dropped the limit and called a five-point difference five percent; checked against the original sentence and recomputed. **Kiffer:** Question three. What does a detector flag tell you about an essay? **Sarah:** Let me start with the way many of these tools work, because it explains everything that follows. They rely on perplexity, a measure of how predictable a text is to a language model. Models write by choosing likely words, so their text tends to be predictable, and detectors treat predictable text as a sign of a machine. The trouble is that people write predictable text too, especially people working with a smaller range of vocabulary. **Kiffer:** Which is what Liang and colleagues tested in 2023. They ran seven widely used detectors on ninety-one essays written by non-native English speakers for the TOEFL, the Test of English as a Foreign Language, and on eighty-eight essays by American eighth graders. The American essays were classified almost perfectly. The TOEFL essays were labelled as AI-generated about sixty-one percent of the time on average, and eighty-nine of the ninety-one were flagged by at least one detector. **Sarah:** And when the researchers used ChatGPT to enrich the vocabulary of those same essays, the false-positive rate fell to about twelve percent. So running a human essay through an AI tool made it less likely to be flagged. **Kiffer:** Detectors have changed since 2023, and newer ones may perform differently. The mechanism the study exposed belongs to the perplexity approach itself, though. Now let's put a detector into a realistic setting, with made-up numbers. Picture one thousand essays. Four hundred are by second-language writers and six hundred by first-language writers. In both groups, one essay in ten contains undisclosed AI text. Our hypothetical detector catches ninety percent of the essays with AI text. It wrongly flags five percent of the human-written essays by first-language writers and twenty percent of those by second-language writers. **Sarah:** And that twenty percent is a good deal kinder to the detector than the average Liang and colleagues found. **Kiffer:** It is. Here's the question. An essay by a second-language writer is flagged. How likely is it that it really contains AI text? Take a few seconds. *(Pause)* **Sarah:** The detector catches ninety percent of AI essays, so I'd say a flagged essay is about ninety percent likely to contain AI text. Maybe a little lower, since the detector makes some mistakes. **Kiffer:** Ninety percent of which essays? **Sarah:** Of the essays that contain AI text. Right, I've turned it around. I want the share of flagged essays that contain AI text, so I need to count everyone who gets flagged. Among the four hundred second-language essays, one in ten is forty essays with AI text, and the detector catches ninety percent of them, which is thirty-six. The other three hundred and sixty essays are human-written, and twenty percent of those are wrongly flagged. That's seventy-two. **Kiffer:** So how many flagged essays are there in that group? **Sarah:** Thirty-six plus seventy-two, which is one hundred and eight, and only thirty-six of them contain AI text. That's one in three. So two out of three flagged second-language writers did nothing wrong, which is a long way from my ninety percent. **Kiffer:** And for the first-language writers? **Sarah:** Sixty essays with AI text, and fifty-four of them are caught. There are five hundred and forty human-written essays, and five percent of those is twenty-seven false flags. So eighty-one are flagged and fifty-four contain AI text. That's two in three. **Kiffer:** So the same flag means something different depending on who wrote the essay. Across all thousand essays, ninety-nine human writers are wrongly flagged, and seventy-two of those ninety-nine are second-language writers. That's almost three quarters of the false flags, from a group that wrote forty percent of the essays. **Sarah:** That helps me see why SFU's guidelines are so firm. Detector results can't be used for grading decisions or academic misconduct investigations, and instructors who have a concern about AI use are asked to talk with the student before reporting it. **Kiffer:** There's one more consequence worth spelling out. If a detector rewards varied vocabulary, it's tempting to edit a text until it passes. That's a different activity from disclosure. What you disclose depends on which tools you used and how, under the rules of your instructor, journal, funder or employer. A text that sounds human, or that a detector passes, carries the same obligations as any other. **Sarah:** And the grey areas are where students get stuck. Take a grammar checker built into a word processor. **Kiffer:** That depends on the rule that applies. Springer Nature, for example, exempts AI-assisted copy editing of human-written text for grammar, spelling, punctuation, readability and tone from disclosure, and generative drafting falls outside that exemption. Other journals word their policies differently, so you read the target journal's own policy. In a course at SFU, the instructor sets the rules, and when they're unclear or missing, the student guidelines say to ask the instructor first. **Sarah:** Students who help with grant applications as research assistants meet one more set of rules. **Kiffer:** Yes. Canada's three federal research funding agencies published joint guidance in November 2024. It recognizes that generative AI can help applicants, including non-native English and French speakers, and it holds the named applicant accountable for everything in the application. For reviewers it's strict. Using publicly available online tools to evaluate applications is prohibited, because uploading them risks privacy breaches and the loss of intellectual property. **Sarah:** Question four. Can a summary with an excellent readability score still let its readers down? **Kiffer:** It can, and the fictional Cedar Valley case showed one way. The assistant's first plain-language summary for Jordan, the team's co-op student, scored grade 5.2 by the checker's rule and still said the program worked, which the abstract contradicts. I'd like to show a second way, which comes from what the formula rewards. As a reminder, the Flesch-Kincaid grade is 0.39 times the words per sentence, plus 11.8 times the syllables per word, minus 15.59. **Sarah:** So it only sees sentence length and word length. **Kiffer:** Right. Here's a made-up passage for a hypothetical clinic leaflet: ninety words in four sentences, with one hundred and fifty syllables. That's 22.5 words per sentence and about 1.67 syllables per word, which works out to a grade of about 12.9. Now suppose an assistant is told to bring it down to grade eight, and it splits the four sentences into nine without changing a single word. What happens to the grade? Take a few seconds. *(Pause)* **Sarah:** Only the first part of the formula changes. Ninety words in nine sentences is ten words per sentence, and 0.39 times ten is 3.9. Before, 0.39 times 22.5 was about 8.8. So the grade drops by about 4.9, from about 12.9 to about 8.0. **Kiffer:** That's nearly five grade levels from punctuation alone, and every word is as hard as it was. Splitting can also do harm of its own. Long sentences in health writing often carry their logic in small words like because, unless and if. Take a made-up leaflet line saying that you should get a flu shot this fall unless your doctor has told you to skip it. Put a period before unless, and the first sentence reads as advice for everyone, with the exception left hanging on its own. **Sarah:** So a formula can reward the very edit that breaks the logic. Do people who write for the public in Canada take that seriously? **Kiffer:** The federal government does. Its Canada.ca Content Style Guide has a heading that reads, don't rely on readability tools. It says these tools measure things like word length and can't tell you whether your audience will find the content easy to understand and use. It treats them as a helpful first step for spotting long sentences, passive voice and acronyms, and it advises setting up usability testing for your content. **Sarah:** So the order of checks is a readability score as a screen, a claim-by-claim comparison with the source, and then people from the intended audience. In the Cedar Valley case, three older adults from the program's advisory group read the corrected summary and changed the phrase referral threshold to the level that leads to a referral. **Kiffer:** And translation ends with people for the same reason. The same small words go missing there, as when up to six times turned into exactly six times in the leaflet example. A back-translation by the same kind of model can hide errors that cancel out, so a fluent reviewer and community members read the leaflet before it goes out. For most First Nations languages in British Columbia, which general translation tools barely support, decisions about translation belong with the language communities themselves. **Sarah:** Okay, our last topic. When, if ever, should an assistant write your first draft? **Kiffer:** I'll make the case for it. For a routine professional document whose content is settled, like a meeting summary or the routine sections of a report drafted from a table, I think writing a detailed outline and letting the assistant turn it into paragraphs is a sensible use of time. The thinking lives in the outline, with every claim, figure and source. The assistant supplies fluent sentences, and you rewrite them and read each one against the outline. It can also help writers who are stuck, or who are writing in a second language. **Sarah:** I'm more cautious, especially for students. Writing does two jobs. It communicates with a reader, and it tests whether your thinking holds together, because a gap in an argument often shows up only when you try to put it into sentences. If the assistant writes the sentences, you lose the second job, and learning to write is largely learning to do the second job. **Kiffer:** That's a fair point for learning tasks. I'd answer that a detailed outline does a lot of that testing. It's hard to list every claim with its evidence without finding the gaps. **Sarah:** You'll find some of them. Outlines are forgiving, though. An arrow from one point to the next looks like an argument until you have to write the sentence that connects them. We saw this in the reverse outline of Jordan's paragraph about rural barriers. The assistant supplied a link between sentences that Jordan never wrote, and its summary read perfectly well. **Kiffer:** That's a real risk, and it's why I'd keep assistant drafting to documents where the connections are already settled. There's also voice. A writer who accepts paragraph after paragraph of generated text drifts toward the assistant's voice, and toward the stock phrasing that readers have learned to notice. **Sarah:** And then there are the rules. In a course, the instructor decides, and an instructor may permit feedback and editing and forbid generated text. For a journal manuscript, the ICMJE recommendations and the COPE position both expect that kind of use to be disclosed. So the decision often isn't entirely yours to make. **Kiffer:** Agreed. So where do we land? **Sarah:** I'd put it this way. Write your claim and your key points yourself before any assistant writes a sentence. When you're learning a kind of writing, draft it yourself and use the assistant as a reader and an editor. **Kiffer:** And when the content is settled and the rules allow it, an assistant can draft from your detailed outline, provided you read every sentence against that outline and disclose the use where the rules ask for it. I lean toward drafting help in more cases than Sarah does, and we agree that the claim, the evidence and the judgement stay with the writer. **Sarah:** Let's pull it together with three things to take away. **Kiffer:** First, the experiments show real gains in speed and rated quality on short tasks that didn't test fact-checking. Plan your time around the parts that stay with you: the claim, the evidence and the checking. **Sarah:** Second, decide on each change separately, and check it for meaning, numbers, hedges, added claims and voice. An assistant's report that nothing changed is one more thing to check. **Kiffer:** And third, treat scores about a text as screens. A detector flag is weak evidence about who wrote an essay, and weakest for second-language writers. A readability grade can be lowered with punctuation alone. Claim-by-claim checks and real readers make the final call. **Sarah:** If you'd like more practice, rework the detector problem from this episode with a different share of essays containing AI text, say one in twenty, and see what happens to the chance that a flagged essay really contains it. **Kiffer:** Next time, it's Lesson seven, AI as a Thinking Partner: Generating and Testing Ideas, where we come back to sycophancy and follow the Cedar Valley team as it stress-tests its program's logic model. **Sarah:** Take care, everyone. **Kiffer:** See you in Lesson seven.