{"id":830,"date":"2026-02-09T12:10:30","date_gmt":"2026-02-09T12:10:30","guid":{"rendered":"https:\/\/egzakt.eu\/j\/?page_id=830"},"modified":"2026-02-09T16:18:28","modified_gmt":"2026-02-09T16:18:28","slug":"ai-in-academic-writing","status":"publish","type":"page","link":"https:\/\/egzakt.eu\/j\/ai-in-academic-writing\/","title":{"rendered":"Be my Writing Guide \u2013 AI in Academic Writing"},"content":{"rendered":"

K\u00e1rp\u00e1ti L\u00e1szl\u00f3<\/a><\/p>

Angoltan\u00e1r \u00e9s \u00fajs\u00e1g\u00edr\u00f3, kor\u00e1bban gyermekv\u00e9delmi gy\u00e1m. A Kaleidoszk\u00f3p\u00a0irodalmi \u00e9s kultur\u00e1lis foly\u00f3irat volt idegen nyelvi rovatvezet\u0151je. A Magyar \u00dajs\u00e1g\u00edr\u00f3k Orsz\u00e1gos Sz\u00f6vets\u00e9g\u00e9nek tagja. Szabad\u00fasz\u00f3 ford\u00edt\u00f3, tolm\u00e1cs \u00e9s lektor. Rendszeresen publik\u00e1l tudom\u00e1nyos \u00e9s sz\u00e9pirodalmi \u00edr\u00e1sokat.<\/p><\/div><\/div>\n\n

The rise of large language models (LLMs) and generative AI (Artificial Intelligence) has created new multifunctional writing tools for the global academic context. AI chatbots like ChatGPT, Claude, Gemini, Copilot, Perplexity and \u2013 most recently \u2013 DeepSeek offer help at almost every stage of writing, from brainstorming to polishing the final draft. By mid-2025, for example, ChatGPT alone had roughly 600\u202fmillion users and Gemini about 350\u202fmillion, reflecting how widely researchers are already experimenting and relying upon these tools. It is almost self-explanatory that each of the above-mentioned AI has its own strengths and weaknesses<\/a>: while some excel at researching and citing sources, others are more well-versed at lengthy reasoning or writing codes.<\/strong><\/p>\n\n\n\n

In the initial phase of writing a research paper, AI can spark ideas<\/em> and accelerate any possible research required for a topic. ChatGPT<\/a> or Claude<\/a> can effortlessly suggest research questions, refine a topic, or even role-play a \u201cSocratic\u201d interlocutor to surface hidden assumptions. For example, a prompt to ChatGPT like \u201cGenerate 5 novel research questions on climate change policy\u201d<\/em> might help a writer break through a mental writer\u2019s block. On the other hand, Google\u2019s Gemini and Microsoft\u2019s Copilot (Bing Chat) combine web search with AI to enhance their productivity: you can ask Copilot to \u201cFind scholarly articles about XX <\/em>\u2013 where X represents the topic in question –  from.edu or.org sites\u201d<\/em> and it will return links and sources for further elaboration. Copilot<\/a> can also crawl a given URL and \u201cprovide a summary of this article\u201d<\/em>, saving a significant amount of time on skimming through the \u201coceans of information\u201d. Similarly, Perplexity AI<\/a> is designed as a research assistant, albeit some could easily argue that it has the most powerful research engine involved. It uses live search to compile concise answers with citations to any sort of question. In \u201cDeep Research\u201d mode, Perplexity runs multiple searches parallel, gathers facts from many possible sources (including academic sources, websites and even social media, if needed), and presents a thorough and detailed report with a generous amount of footnotes and further links. This makes it ideal for literature surveys. Its free tier allows a few in-depth queries per day, making it accessible for students and independent researchers<\/a>, too.<\/p>\n\n\n

\n
\"\"<\/figure>\n<\/div>\n\n\n

Source: freepik.com<\/em><\/p>\n\n\n\n

When it comes to creative brainstorming, ChatGPT and Claude offer fast actions of such sort, but\u00a0can\u2019t substitute for reading the relevant literature of the field in question<\/em>. It should come as no surprise that over-relying on AI to generate ideas can stunt one\u2019s own critical and creative way of thinking and easily lead to a state of being a \u201cmental couch potato\u201d. Likewise, while AI can suggest search terms or authors to check (ChatGPT can \u201cjumpstart your research\u201d by explaining the fundamentals of a given topic), any factual claims it makes should be verified. In contrast,\u00a0Google Gemini<\/a>\u00a0with its up-to-date web access and Perplexity AI are generally more skilled at finding recently published articles and facts, due to using live search. Claude, however, boasts a huge context window (up to ~200,000 tokens vs ChatGPT\u2019s 32k), so it can ingest and reason over very large texts (like dozens of paper abstracts at once) when brainstorming. In practice, it\u2019s wise to\u00a0mix more<\/a>\u00a0of the above-listed tools: use one or two AI chatbots to generate ideas, followed by rapid research into a given topic and finishing the process with quick summaries of texts found. This, however, must be followed up by reading through the findings of AI, thinking through the information and expending upon it.<\/p>\n\n\n\n

Once a particular topic is chosen, AI can help create a structure or outline for further questions. ChatGPT, Claude, and Copilot<\/a> can turn a research question into an outline or section headings. For example, you might prompt Copilot: \u201cCreate an outline for a research paper on the impacts of AI in education\u201d<\/em>; it will propose sections and key points. Similarly, ChatGPT can draft an outline or bullet list from a brief (e.g. \u201cDraft an outline for a 5-page essay on the fall of the Roman Empire\u201d<\/em>).<\/em> After an outline has been established, ChatGPT can even review and criticize it for further improvement. As the OpenAI Student Guide<\/a> notes it, you can ask ChatGPT to \u201creview the structure and provide feedback on the flow of ideas.\u201d<\/em>. Such semi-automatic assistance can<\/em> be helpful during academic writing, but must be double-checked<\/a> at all times. Because outline generation always involve new content creation, there\u2019s a risk of leaving out important points or introducing biases. For instance, the model might emphasize one theme over another or suggest unnecessary sections, therefore, the user must ensure that the outline truly reflects their intended argument. In short, using AI can speed up and enhance generating ideas for a scientific paper, but can\u2019t replace the author\u2019s actual planning and review of the final product.<\/p>\n\n\n\n

When it comes to writing in a more detailed and elaborate way, you can feed your own bullet points or notes into ChatGPT or Claude and prompt it to, \u201cTurn these into a formal paragraph.\u201d Writers often use this to break through writer\u2019s block or to see different phrasing style suggestions. Both ChatGPT and Claude can generate academic-style prose upon request. For example, prompting the following \u201cWrite a 200-word introductory paragraph about climate modeling\u201d can quickly and easily yield a coherent paragraph. Because of its broad training, ChatGPT can draw analogies or examples from many fields, while Claude\u2019s specialized training helps it use a more niched technical language accurately. For very technical drafting – for instance, writing code or advanced reasoning -, tools like DeepSeek<\/a> or ChatGPT\u2019s o4-mini and o4-mini-high models can cater to such needs.<\/p>\n\n\n\n

One thing that is common in when using any LLM is the possible occurrence of hallucinations. LLMs generate plausible-sounding text based on statistics, not on factualness. The more frequently something occurs on the internet, the more likely it is to appear in the text presented to the prompter. Therefore, it is suggested that instead of asking for brand new content from AI, one should use sources and references that are previously fact-checked, including, but not limited to, one\u2019s notes. Bearing that suggestion in mind, we are ought to treat the AI as a collaborator, rather than as a ghostwriter.<\/p>\n\n\n

\n
\"\"<\/figure>\n<\/div>\n\n\n

Source: freepik.com<\/em><\/p>\n\n\n\n

After the first draft has been written, the AI can polish both the language and style of the text. All the tools presented can effortlessly check grammar, suggest synonyms, or rephrase awkward sentences for a clearer understanding. For instance, you might paste a paragraph into ChatGPT and ask it to \u201cImprove the clarity of the pasted text and add an academic tone.\u201d As such, the job of a professional proofreader and editor can be integrated into the workflow without much stress, thus pointing out inconsistencies or tone issues can be dismissed quickly. Here we have to mention a more grammar-centered application that can be used on multiple platforms, including its web version: Grammarly<\/a>. If one merely seeks an app for the single purpose of proofreading without drafting or writing paragraphs, Grammarly can be a solid choice. <\/p>\n\n\n\n

As far as citation and referencing goes, Copilot and ChatGPT can format citations by turning a list of previously provided references into a styled bibliography, regardless if it\u2019s APA, MLA, or Chicago style we seek. It can also check citation styles and correct possible mistakes. To generate a citation, you only need to paste a DOI into your AI in question and ask to \u201cgenerate an MLA\/APA\/Chicago citation.\u201d ChatGPT can often guess citation formats correctly and because of that fact alone, OpenAI recommends using ChatGPT to \u201cdelegate citation grunt work\u201d like formatting, allowing researchers more room to focus on the actual content. Some new tools (like SciSpace<\/a> or Yomu AI<\/a>) specialize in finding and formatting citations from academic databases, but general models can handle basic tasks too. An ordinary issue with most LLMs is that they are bad at finding accurate references <\/em>and frequently lead the prompter astray with false or empty links.<\/p>\n\n\n\n

Multiple studies show that AI often hallucinates papers or gives wrong details. Cheng et al<\/a>. cite work showing only ~7% of ChatGPT\u2019s references were fully correct, with many DOIs fake. Similarly, they note that ChatGPT\u2019s literature searches \u201cmust thus be confirmed with a traditional search engine,\u201d as it can fabricate plausible-looking citations. In contrast, Perplexity AI will actually generate links to real sources, which is safer for fact-checking. The rule here is as follows: never trust a citation unless you click it. Always double-check references using Google Scholar or your library\u2019s tools. For writers working in a second language, ChatGPT and Gemini can help translate or rephrase content. You can ask ChatGPT to \u201cTranslate this paragraph into English (academic style)\u201d, and it will produce a solid draft. Google\u2019s Gemini (leveraging Google Translate tech) is also strong for such goals, both applicants generating semantically accurate translations, aiding non-native authors. A word of warning though: subtle nuances are often missed by the engines. Therefore, a thorough proofreading by a fluent human editor or native speaker for tone and cultural context is highly recommended here.<\/p>\n\n\n\n

As a quick rundown on the main characteristics of the above described AI chatbots, we can say that:<\/p>\n\n\n\n