---

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 – most recently – 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 million users and Gemini about 350 million, 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: while some excel at researching and citing sources, others are more well-versed at lengthy reasoning or writing codes.

In the initial phase of writing a research paper, AI can spark ideas and accelerate any possible research required for a topic. ChatGPT or Claude can effortlessly suggest research questions, refine a topic, or even role-play a “Socratic” interlocutor to surface hidden assumptions. For example, a prompt to ChatGPT like “Generate 5 novel research questions on climate change policy” might help a writer break through a mental writer’s block. On the other hand, Google’s Gemini and Microsoft’s Copilot (Bing Chat) combine web search with AI to enhance their productivity: you can ask Copilot to “Find scholarly articles about XX – where X represents the topic in question –  from.edu or.org sites” and it will return links and sources for further elaboration. Copilot can also crawl a given URL and “provide a summary of this article”, saving a significant amount of time on skimming through the “oceans of information”. Similarly, Perplexity AI 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 “Deep Research” 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, too.

Source: freepik.com

When it comes to creative brainstorming, ChatGPT and Claude offer fast actions of such sort, but can’t substitute for reading the relevant literature of the field in question. It should come as no surprise that over-relying on AI to generate ideas can stunt one’s own critical and creative way of thinking and easily lead to a state of being a “mental couch potato”. Likewise, while AI can suggest search terms or authors to check (ChatGPT can “jumpstart your research” by explaining the fundamentals of a given topic), any factual claims it makes should be verified. In contrast, Google Gemini with 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’s 32k), so it can ingest and reason over very large texts (like dozens of paper abstracts at once) when brainstorming. In practice, it’s wise to mix more of 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.

Once a particular topic is chosen, AI can help create a structure or outline for further questions. ChatGPT, Claude, and Copilot can turn a research question into an outline or section headings. For example, you might prompt Copilot: “Create an outline for a research paper on the impacts of AI in education”; it will propose sections and key points. Similarly, ChatGPT can draft an outline or bullet list from a brief (e.g. “Draft an outline for a 5-page essay on the fall of the Roman Empire”). After an outline has been established, ChatGPT can even review and criticize it for further improvement. As the OpenAI Student Guide notes it, you can ask ChatGPT to “review the structure and provide feedback on the flow of ideas.”. Such semi-automatic assistance can be helpful during academic writing, but must be double-checked at all times. Because outline generation always involve new content creation, there’s 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’t replace the author’s actual planning and review of the final product.

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, “Turn these into a formal paragraph.” Writers often use this to break through writer’s block or to see different phrasing style suggestions. Both ChatGPT and Claude can generate academic-style prose upon request. For example, prompting the following “Write a 200-word introductory paragraph about climate modeling” can quickly and easily yield a coherent paragraph. Because of its broad training, ChatGPT can draw analogies or examples from many fields, while Claude’s 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 or ChatGPT’s o4-mini and o4-mini-high models can cater to such needs.

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’s notes. Bearing that suggestion in mind, we are ought to treat the AI as a collaborator, rather than as a ghostwriter.

Source: freepik.com

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 “Improve the clarity of the pasted text and add an academic tone.” 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. If one merely seeks an app for the single purpose of proofreading without drafting or writing paragraphs, Grammarly can be a solid choice. 

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’s 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 “generate an MLA/APA/Chicago citation.” ChatGPT can often guess citation formats correctly and because of that fact alone, OpenAI recommends using ChatGPT to “delegate citation grunt work” like formatting, allowing researchers more room to focus on the actual content. Some new tools (like SciSpace or Yomu AI) 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 and frequently lead the prompter astray with false or empty links.

Multiple studies show that AI often hallucinates papers or gives wrong details. Cheng et al. cite work showing only ~7% of ChatGPT’s references were fully correct, with many DOIs fake. Similarly, they note that ChatGPT’s literature searches “must thus be confirmed with a traditional search engine,” 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’s tools. For writers working in a second language, ChatGPT and Gemini can help translate or rephrase content. You can ask ChatGPT to “Translate this paragraph into English (academic style)”, and it will produce a solid draft. Google’s 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.

As a quick rundown on the main characteristics of the above described AI chatbots, we can say that:

• ChatGPT (GPT-4.5/4o) is a versatile generalist. It handles creative phrasing, broad Q&A, and coding. It has plugins for browsing or coding. Its strengths are in conversation and idea generation, but it can hallucinate facts. Use it for brainstorming, drafting, and writing tips, but with caution on factual correctness.
• Anthropic’s Claude focuses on safe, coherent responses. It can read enormously large inputs (200k+ tokens), which makes it great for analyzing long documents or staying consistent in a lengthy draft. Its training emphasizes ethical safeguards, so some users feel its outputs are more measured. It’s especially handy for deep dives into specialized topics, since it tends to preserve user instructions precisely.
• Google Gemini excels at integrating the web. Its Deep Research mode produces structured reports with up-to-date info and source links. It handles multimodal inputs (images, etc.) better than others. Use Gemini when you need the latest findings or multilingual support. Just remember it may summarize without the nuance a human would, so always read the linked sources.
• Perplexity AI is essentially a citation-savvy search assistant. Its free version lets you ask factual questions and get answers with footnotes quickly. For literature review or verifying claims, it’s very useful because it must cite its sources. It’s less useful for creative writing. Think of Perplexity as an AI-powered Google Scholar.
• Microsoft Copilot (Edge/365) integrates AI into familiar apps. Copilot can perform web searches on command (like the examples above), summarize articles, outline your essay, and format citations. It’s particularly convenient for users already on Windows/Office, since it lives in the browser or Word. Its strength is as a research partner: it will crawl and cite web pages and even teach you writing tips on demand.
• DeepSeek AI is a newer open-source Chinese model (launched Jan 2025) designed for coding and reasoning. It’s very strong on technical tasks and more cost-effective to run, but it’s still emerging in the West. It’s worth watching, but for now ChatGPT/Gemini cover most writing needs.

As a final thought, based on public benchmarks, we can claim the following: Claude is mostly suitable for long context with its 200k token window and occasional coding tasks; ChatGPT is a general all-rounder with great potential in coding and critical thinking; Perplexity AI is best fitted for (deep) research with transparent citations and a major amount of references; DeepSeek is most often used as the most autonomous all-rounder of the above-mentioned AI models. However, none of these applicants can entirely replace a human writer’s judgment.

Source: freepik.com

At the end of the day, whenever you decide to rely on the use of AI for the creation and polishing of your academic text, the following should be considered as a plausible guideline: 

1. Never take anything generated by AI for granted. Double-check facts, figures, and references from any AI output used during the process, otherwise you might run into a plausible-looking DOI number or title that doesn’t exist. 
2. Contribute substantially to the final text. Because your ultimate goal should be to remain the author of your work, don’t let the AI “gobble up” your texts and “spit out” an article that can be published under your name. Instead, guide the AI to help you in writing the text for you and not instead of you. 
3. Be transparent in your use of AI: if an AI chatbot (or multiple ones) helped you shape your work, acknowledge it. As OpenAI advises, include shareable ChatGPT links or notes about the AI’s role in your bibliography. Some journals and institutions now require their authors AI disclosure. This openness about actual authorship follows standard academic practice of citing all sources. After all, all academics need to list their references at the end of their papers. It is only natural that the use of AI be mentioned likewise.

As a final remark, we always have to bear in mind that AI is more like a crotch, than your actual legs when it comes to walking the path of the Academia. It aids us in creating more refined and more fitting academic content, but it is not the content creator itself.