10 Writing with AI
AI writing tools — by which this chapter means large language model-based tools such as ChatGPT, Claude, Gemini, and similar systems — are now part of academic life. Most researchers use them in some form — for drafting, for editing, for getting past the first sentence. Used well, they can make the writing process faster and less painful.
Used poorly, they make the problem this guide is about significantly worse.
10.1 The expansion problem
AI writing tools have a systematic tendency to expand. Ask an AI to improve your introduction and it will suggest adding things — more context, more nuance, more theories (Uhler et al., 2026). This reflects how these systems are trained. Large language models are typically fine-tuned using human feedback, where human raters judge which outputs are better. Raters tend to reward outputs that seem comprehensive and thorough — so models learn that adding content is usually the safer move. The result is a tool that is structurally biased towards inclusion.
In the AI research literature, a related phenomenon is called sycophancy: the tendency of models to tell users what seems agreeable rather than what is accurate or useful (Cheng et al., 2026). In the context of academic writing, both tendencies — expansion and sycophancy — produce exactly the bloat this guide is designed to prevent.
A concrete example. You have a focused theoretical framing — three paragraphs, one anchor theory, one prediction. You ask an AI to improve it. The AI returns five paragraphs. It has added a paragraph on Situation Models because they are “highly relevant to event cognition research.” It has added a sentence noting that “predictive processing frameworks offer a complementary perspective.” The writing is fluent. The argument is weaker.
The problem is invisible if you do not know what to look for. The expanded version sounds authoritative. It sounds thorough. It sounds like a paper that will satisfy reviewers. It is the first attempt from the worked example chapter, generated in thirty seconds.
10.2 How expansion manifests
Sycophantic expansion produces three recognisable failure modes. They share a common root — the AI optimising for what seems agreeable rather than what is argumentatively sound — but they show up differently on the page.
Sycophantic expansion. AI adds theories, qualifications, and hedges that dilute the argument. This is the most common failure mode and the hardest to detect because the output looks polished.
Confident vagueness. AI connects theories to your argument with fluent but imprecise language. Phrases like “this aligns with”, “is consistent with”, and “offers a complementary perspective on” sound substantive but say nothing about whether the theory makes a differential prediction for your design. A theoretical framing full of these phrases fails the two filters — does this theory make a different prediction for my design, and does it reach my dependent variable mechanistically? — while appearing to pass them.
Premature polish. AI produces text that looks finished before the argument is sound. This kills the impulse to revise. If a paragraph reads well, it feels wrong to delete it — even if it is doing no argumentative work. The polish is a trap.
10.3 What AI does well
Once you understand the expansion problem, the question is how to use it against its own tendencies. Before that, three genuinely unproblematic uses:
- Sentence-level editing — improving clarity, flow, and concision once the argument is already clear
- Formatting and consistency — standardising terminology, checking reference formatting, fixing grammar
- Overcoming blank page paralysis — generating a rough first draft that you then revise heavily
For everything else — drafting theoretical framings, cutting redundancy, checking argument structure, compressing the paper, stress-testing against a critical reviewer — the prompts in the next section show how to use AI productively by giving it the right constraints.
10.4 How to use AI well
The most important rule: complete your anchor sentence and your landscape sentence before opening an AI tool. These are the constraints that keep AI expansion in check. Without them, the AI has no basis for knowing what belongs in your paper and what does not — and it will optimise for comprehensiveness, which is the opposite of what you need.
With those constraints in place, AI can be used productively at every stage of the writing process. The prompts below are designed to work with specific inputs — paste them in exactly as instructed, and be precise. A vague prompt produces a vague response.
10.4.1 Before you write
For drafting:
I am writing the theoretical framing for an empirical paper to be submitted to a peer-reviewed psychology journal. My anchor sentence is: [paste your anchor sentence exactly as written in the anchor exercise]. My landscape sentence is: [paste your landscape sentence exactly as written after the filter exercise]. Write a theoretical framing of exactly three paragraphs that moves from the phenomenon to the anchor theory to the specific prediction. The first paragraph establishes why the phenomenon matters. The second introduces the anchor theory and its mechanism. The third states the specific prediction and disposes of excluded theories in one sentence. Do not introduce any theoretical framework other than [anchor theory]. Do not use the phrases “consistent with”, “aligns with”, or “offers a complementary perspective”.
10.4.2 Once you have a draft
The following prompts work on existing text. In each case, the anchor sentence is the constraint that stops the AI from optimising for comprehensiveness instead of argumentative clarity. Always paste the full text you want assessed — AI cannot read your manuscript unless you provide it.
For summarising as a consistency check:
Here is my theoretical framing: [paste text]. Summarise the argument in three to five sentences, in your own words. Do not use the same sentence structures as the original. State: what phenomenon is being studied, which theory drives the hypotheses and what it predicts, and what the study will show. Do not add anything that is not explicitly stated in the text.
This prompt uses AI’s limitations as a diagnostic tool. A competent summariser that cannot produce a coherent summary is telling you something: the source text contains gaps or contradictions that fluent prose was concealing. If the summary is confused or vague, the problem is not the AI — it is the framing. The “do not add anything that is not in the text” instruction is essential; without it, AI will fill gaps with plausible inference and mask the very inconsistencies you need to find.
For cutting:
Here is my theoretical framing: [paste text]. My anchor sentence is: [paste anchor sentence]. Identify every sentence that does not directly advance the argument toward the prediction in the anchor sentence. For each such sentence, state in one line why it does not advance the argument, and recommend removing it or specify what it would need to say to earn its place.
For the argument progression test:
Here is my theoretical framing: [paste text]. For each paragraph, write one sentence stating the single step it adds to the argument. Then assess whether the sequence of steps tells a complete story from phenomenon to prediction without gaps or redundancy. List any gap — a missing step — and any redundancy — two paragraphs doing the same job.
For the elevator abstract:
Here is my full manuscript: [paste text]. Write a two-sentence summary of the paper. The first sentence states what I did and why — the theoretical prediction tested and how it was operationalized. The second sentence states what I found and what it means for the anchor theory. Use precise theoretical language but no unnecessary jargon. Do not begin either sentence with “This study” or “We found”.
For the plain-language summary:
Here is a description of my study: [paste a brief description of your design, manipulation, dependent variable, and main finding]. Write two to three sentences explaining what this study is about and why it matters. Use no technical terms, no theory names, and no field-specific jargon. Write for an intelligent adult with no background in this area of research. If anything in my description is unclear or would require specialist knowledge to understand, flag it.
For critical review:
Here is my theoretical framing: [paste text]. My anchor sentence is: [paste anchor sentence]. Act as a critical peer reviewer with expertise in this area. For each theory that receives more than a citation, assess whether it passes both of the following filters: (1) does it make a different prediction from my anchor theory for this specific manipulation and dependent variable, and (2) does it reach my dependent variable mechanistically — not just at the level of the phenomenon in general? Flag any theory that receives substantive treatment without passing both filters. Flag any sentence that uses the phrases “consistent with”, “aligns with”, or “offers a complementary perspective” without specifying a differential prediction. Be direct and specific; do not soften criticisms.
The critical review prompt is the sharpest of these tools: it turns the sycophantic tendency on its head by explicitly instructing the AI to look for the weaknesses that sycophancy would normally conceal.
If the critical review prompt flags weaknesses in your theoretical framing, address them before submission. The next section deals with the harder case: a reviewer who flags them after.
10.4.3 After reviews come back
For reviewer responses:
A reviewer has asked me to add a substantive discussion of [Theory X] to my introduction. My anchor theory is [Theory Y]. My anchor sentence is: [paste anchor sentence]. In my design, [Theory X] and [Theory Y] generate identical directional predictions for [specific manipulation and dependent variable] — [Theory X] fails Filter 1 because it makes no distinct prediction for this operationalization. Write a polite, specific response to the reviewer that: (1) acknowledges the importance of [Theory X] in the field, (2) explains precisely why our design cannot adjudicate between [Theory X] and [Theory Y], and (3) states that we have added a sentence acknowledging [Theory X] at [location in manuscript]. Do not be dismissive. Do not simply refuse. Be specific about why the filter fails for this design.
Notice the pattern across all these prompts: in every case, you give the AI a constraint derived from the tools in this guide. The summarising prompt uses AI’s failure as a signal rather than a flaw — an incoherent summary points to an incoherent source. The argument progression test prompt asks AI to do what you would do manually — but faster, and without the attachment to your own prose that makes self-editing hard. The critical review prompt uses AI as a stand-in for the hostile reviewer: if it finds weaknesses, a real reviewer will too. These prompts turn AI’s tendencies to compress and summarise into assets rather than liabilities — the opposite of asking it to expand and improve.
10.5 A note on honesty
Using AI to assist with writing is not dishonest, provided you follow your institution’s guidelines and disclose appropriately. But there is a subtler form of dishonesty worth naming: using AI to generate a theoretical framing you do not fully understand, because the framing sounds authoritative enough to submit.
If you cannot explain in your own words why each theory in your introduction is there — what it predicts for your design, why it matters for your argument — then the framing is not yours, regardless of who edited the sentences. A reviewer who asks a sharp question about your theoretical choices will expose this immediately.
AI can help you write what you think. It cannot think for you. Every tool in this guide — the anchor exercise, the two filters, the landscape sentence, the argument progression test, the elevator abstract, the plain-language summary — is a thinking tool. Do it yourself, on paper, before you open the AI.
10.6 The sharp tool
A sharp tool cuts well in the right hands and badly in the wrong ones. The same AI prompt produces a focused, well-argued theoretical framing when the writer has a clear anchor — and a bloated, theory-laden paragraph when they do not.
The guide you have just read is, among other things, a set of instructions for how to pick up the tool correctly. The anchor sentence is the handle. The constraints you bring to the tool determine how precisely it cuts.