11 Lessons Learned Writing an Academic Paper with AI

Initial experiments with AI-generated text quickly exposed a core limitation: AI, by itself, struggles with the nuanced demands of academic writing. Simple prompts like 'write about X' often result in disorganized content lacking argumentative cohesion. Academic writing, by its very nature, relies on a clear, logical progression of ideas – a structured flow from premise to conclusion. AI, without explicit guidance, tends to generate relevant but disconnected prose, skipping the crucial step-by-step reasoning. The author discovered that the most effective solution is to begin each section with meticulously crafted bullet points. These bullet points serve as a scaffolding, outlining the intended message, the logical connections between ideas, and any necessary nuances or caveats. When provided with this structured foundation, AI's output improves dramatically, allowing it to focus on refining language rather than guessing at content or argumentation. This approach transforms the writing process, shifting the human effort towards structuring ideas and feeding them to the AI for eloquent articulation.

Despite the significant improvements in fluency offered by custom GPTs, AI models often exhibit certain stylistic tendencies or 'tics' that persist regardless of prompting. A common example identified by the author is the overuse of the em-dash (—), a punctuation mark that can often be replaced by a comma for better readability. Initially unnoticed, these recurring stylistic patterns can become a giveaway that a passage has been AI-assisted. Beyond punctuation, certain sentence structures or transitional phrases may also appear repeatedly. The crucial takeaway is that AI-generated text requires post-editing not only for accuracy but also for stylistic naturalness, especially when aiming to maintain a distinct scholarly voice. Vigilance in identifying and correcting these AI-specific patterns is essential for producing authentic-sounding academic work.

Throughout the research process, the author continued to rely heavily on Litmaps, a tool that searches papers based on citations rather than AI algorithms. The preferred approach involves seeding the search with two types of papers: first, a central review paper that broadly defines the topic (e.g., 'biodiversity change'); and second, a methods paper that formalizes a specific measurement or approach (e.g., a particular metric of 'functional biodiversity'). Litmaps then identifies papers that cite both seed papers, revealing how specific methods are applied to broader problems. A minor drawback noted is the manual effort required to import papers into Litmaps from reference managers, which can sometimes lead to skipping this step. Despite this, the combination of Litmaps with method-specific papers provides a powerful method for precise literature discovery.

When analyzing papers with AI, a practical observation emerged: copying text directly from a publisher's website yielded superior results compared to uploading PDF files. Even though website text often includes unwanted elements like sidebars or menu headers, AI seemed to struggle with parsing PDF formatting reliably into plain text. This difficulty in reliably converting PDFs into usable text likely accounts for the performance difference. Therefore, for optimal AI analysis of academic papers, the author recommends the copy-pasting method over direct PDF uploads.

Generic prompts such as 'Write a paragraph about X' are likely to produce vague and unsatisfactory results. To achieve high-quality output, substantial context must be provided. This includes not only the topic but also the intended logical flow of ideas, specific points of emphasis, desired contrasts, and relevant supporting materials like additional papers or text snippets. In some cases, prompts became so detailed that ChatGPT indicated the messages were too large. However, this level of explicitness significantly improved the AI's ability to deliver output aligned with the user's intentions. For instance, when polishing a paragraph on ecosystem function, the author would specify the order of concepts (definition, contrast with related terms, application to a case study) and desired emphasis (prioritizing empirical findings over theoretical discussion). The more detailed and explicit the prompt, the better the AI can serve as a precise writing assistant.

The author's experience underscores that off-the-shelf AI tools often fall short of specific academic writing needs, leading to abandonment. Custom GPTs offer a powerful solution due to their deep customizability. The author has developed a suite of custom GPTs for various writing stages. For example, 'WritingWanda' assists in rephrasing and polishing paragraphs based on structured notes, with detailed instructions on academic writing conventions. Another, 'FocalExtraction,' can process a full-text article and a specific question to extract relevant quotations, classifying them as supporting, contradicting, or elaborating. These tools are not only helpful but also adaptable; when AI performance is suboptimal, instructions can be updated. This flexibility makes custom GPTs indispensable collaborators. Unlike generic AI assistants, they can be tuned for consistent performance across specific tasks. Furthermore, ChatGPT's memory function enhances their effectiveness, reducing the need for repetitive context in prompts, allowing the AI to retain specific project details, such as the author's focus on plant-related research.

The use of AI in academic writing is permissible, provided that its application is transparently acknowledged and documented. Journals are increasingly establishing guidelines for AI disclosure. A general template for acknowledging AI use, adapted from Wiley's author guidelines, involves stating which generative AI technologies were used and for what purpose. The specific AI-generated content within the manuscript should be clearly marked and described in a dedicated appendix for editorial and review purposes. Crucially, authors must review all AI-generated content and take full responsibility for the accuracy and integrity of the submitted manuscript.