What can AI do for academics? Writing, revising, searching, polishing, and peer-reviewing with AI

AI is not a passing trend in research. It’s a structural shift in how knowledge is accessed, processed, and communicated. For researchers, the challenge is no longer whether to engage with AI, but how to do so with skill, integrity, and confidence. In this post, you can take a test that outlines 10 essential competencies for thriving in an AI-powered academic environment. It’s not about mastering every new app, but building repeatable habits that adapt as tools evolve.

If you haven’t done it already, take some time to answer each of these questions with a yes or no. Count the number of yeses and compare it to the table in the bottom left.

But even if you don’t have a perfect score, learning to use AI is not rocket science. Let’s dive into each of the 10 elements of this test and see what they entail and how you can improve.

1. Understand the Basics of Language Models

To use tools like ChatGPT, Claude, or Gemini effectively in academic work, you first need to understand what language models actually do and what they can’t. Our first reaction is to anthropomorphise AI because it talks to us like a human; we attribute human-like qualities to it.

That is wrong. Language models predict text based on statistical patterns, not factual knowledge. Their outputs can be fluent but wrong, and oftentimes confident but outdated. Knowing this helps you spot limitations and avoid over-relying on surface-level summaries.

More importantly, modern large language models lack a concept known as a world model, which, according to leading AI scientist Yann LeCun, is the reason why they aren’t truly intelligent. Think of it this way: if you see a stone falling off a cliff when you drop it, you can conclude that a car or any other object will drop down the cliff even though you have never observed the car driving down a cliff. A large language model, by contrast, will need to see the car driving down the cliff in order to associate “driving down the cliff” with “falling down” and “the car”, because it learns “by example” not “by deduction”.

If you want to explore this topic in more depth, read about the five core ideas(article link) that explain how AI tools think, reason, and interact:

1. Perception – how AI “reads” research sources
2. Representation and reasoning – why AI gives the answers it does
3. Learning – why AI keeps getting slightly better
4. Natural interaction – how to talk to AI
5. Societal impact – Ethics in AI literature reviews and academic writing

2. Keep Up with AI Tools and Capabilities

The academic AI ecosystem is changing rapidly, and so are the tools. Currently, the most powerful types of tools for researchers are writing assistants, research copilots and note-integrated agents – but you might only need one. Most importantly, the vast majority of AI tools are built only on four core models: Google’s Gemini, OpenAI’s ChatGPT, Claude’s Anthropic, Meta’s LLAMA or DeepSeek’s Rx models. The primary difference between AI tools lies in their ability to integrate external data effectively. In the academic use case, it refers to the number of papers they can access and how they process this information in relation to your query.

It can be overwhelming keeping up to date with the newest AI tools, so rather than chasing every launch, get clear on what types of tools serve your actual needs:

• Source discovery: Try Elicit or Litmaps to surface relevant papers through plain-language queries or citation networks.
• Idea development: Use ChatGPT’s Projects feature or NotebookLM to reflect across your own notes and generate structured prompts.
• Writing structure: Tools like Yomu AI and Paperpal can help scaffold, refine, and rewrite academic prose. But you can also easily customize ChatGPT to write academically.
• Reference management: Use Scite Assistant for smart citation suggestions, or integrate Zotero with AI-powered plugins for auto-tagging and formatting.
• Workflow automation: Platforms like Obsidian and Notion support AI integrations that can automate summarisation, linking, and recurring task generation.

Check out the tools page for a detailed breakdown of these tools.

3. Use CustomGPTs or Personalised AI Assistants

Generic prompts give generic results. But typing up lengthy prompts takes too much time. This is the problem customGPTs solve (also called Gems in Google Gemini). A custom GPT is like a specialised assistant, and you specialise it by giving it a prompt which can have considerable length and gears its replies towards the purpose that you want to achieve. For instance, you can have a custom GPT help you with academic writing by describing exactly what type of rules academic writing follows, what type of jargon to use, and so on.

The possibilities are truly limitless, and this might be the single most useful feature to master to get the most out of AI at the time of writing. Here are some other ways to use custom GPTs.

• Create a GPT that critiques your academic writing using your supervisor’s criteria
• Writing a discussion section for your manuscript using GPTs and mega prompts
• Writing an introduction for your manuscript from your notes (the atomic sentences method)
• Build a lit review summariser that tags sources by method, location, or theory
• Configure Obsidian’s AI plugins to help synthesise your findings

Custom GPTs are one of the most useful and powerful features in AI right now. I created a separate course dedicated entirely to programming and customising these assistants called the Effortless AI course.

4. Automate and Augment with AI

AI can eliminate repetitive, low-value tasks, but only if you delegate with care. Good automation is subtle: it reduces friction without eroding your control or critical engagement. But AI also makes mistakes. You must ensure that these mistakes don’t multiply.

For example, if you take some text and edit it with AI, then use the result on another AI with a different purpose, the errors each AI makes often compound. It is a little bit like the game of the broken telephone, where the name of the message is slowly changed as it is transmitted with errors.

But this should not deter you from using AI for automation, rather sensitivize you to understand what and when you have to control. Here are some practical examples of smart automation in academic work:

• Formatting and managing references
  Use Sourcely to generate structured citations with evidence strength, or integrate Zotero with tools like Obsidian or Word for seamless in-text citation and bibliography management.
• Extracting structured data from PDFs
  Tools like Scholarcy or Sumnotes can pull out highlights, tables, and references. Then use workflows (e.g., Zotfile + Zotero + Obsidian) to sync metadata and annotations into your knowledge base.
• Finding Notes or PDFs using semantic search
  AI is perfect for searching for specific text within your notes or PDF collections. Here’s an example for Obsidian notes. The same applies to searching for PDFs.
• Batch-processing text data
  For qualitative researchers, tools like Otter.ai (for transcription) or ChatGPT with Code Interpreter can help tag, sort, and summarise interview data more efficiently.
• Automated literature monitoring
  Tools like Litmaps or Connected Papers can notify you of newly published work that matches your research trajectory, reducing the manual overhead of staying up to date.

Automation helps accelerate quality work but only when you’re clear on what the task actually requires. Always review outputs with a critical eye.

5. Enhance Literature Discovery with AI

Traditional database queries often miss relevant papers due to strict keyword matching. AI-powered tools like Elicit, Semantic Scholar, and Connected Papers overcome this limitation by employing semantic search and citation network analysis. Semantic search works by comparing a body of text to another body of text using complicated mathematical calculations. Imagine each published paper being a star in the night sky. Nearby stars indicate that the two texts or papers are closely related and discuss the same topic.

To generate this semantic similarity, large language models “read the paper and then assign it a very high-dimensional number” (akin to a position on the night sky). Retrieval of similar papers becomes very easy after the position has been determined.

Popular tools for semantic search are Elicit, Consensus and SciSpace. They allows researchers to input plain-language questions and retrieve studies by analysing semantic similarity and provide structured summaries, including study design, sample size, interventions, and outcomes, facilitating efficient screening and evaluation of literature.

A big problem here is that large language models cannot distinguish between a highly acclaimed peer-reviewed publication and a forgotten abstract on a pre-publishing server (e.g. ArXiv), written in a language that you can’t understand. Semantically, these two papers are similar, but your human perception would be vastly different.

You should therefore always double-check the quality of semantic search results.

6. Compare and summarise research papers with AI

Summarising papers was one of the first use cases for AI in academia. However, most papers also have an abstract, so just asking for a summary is not very informative. If, however, you become a little bit more creative, you can get much more out this technology.

For instance, using two PDFs instead of one allows you to compare studies, helps you understand how research aligns, where it diverges, and what questions remain. But doing this across many papers can be slow and fragmented.

Tools like Anara AI allow to compare multiple papers in a systematic way. They can help to derive insights for your literature review and sometimes even spot the gap your research should address.

AI can speed up source digestion, but it often simplifies in ways that mask nuance. Don’t take summaries at face value. Cross-check claims, revisit the original, and rephrase in your own words.

When the outputs don’t seem reliable:

• Use structured prompts like “List the main findings and limitations of this paper”
• Ask for comparisons: “What’s similar and different in these three studies?”
• Use tools like Anara AI or Paperpal that retain citation traceability

When using AI to analyse PDFs, think a little bit more creatively than just summarising. For instance, you can ask whether a study supports or rejects a certain argument, or find exact sentences where this study speaks to a specific topic. All of these aspects move a simple PDF chat with an AI into

7. Co-write with AI (without losing your voice)

AI writing tools can be deceptively fluent. It’s easy to start relying on them for structure, phrasing, or even content. However, once you have used that technology for a while, you will start to become aware of AI-generated text. Most famously, the m-dash (–) is a hallmark of using ChatGPT. While those stylistic properties of your language can be easily remedied, other properties like proper jargon, sentence length, active and passive language, tenses can be more challenging.

The key isn’t avoiding AI in writing, but to use proper guidance. Just like a first-time PhD student will not be fluent at academic writing, AI has challenges here too. Help AI by starting with your own outline or argument map, use it to suggest transitional phrasing, restructure clunky paragraphs, or condense long explanations. At the end, always polish using your own skills.

Another aspect of guidance is the right prompt, which can make a difference between unusable academic writing and something that is very close to sounding like you. Longer prompts usually perform better. So the key is to describe your academic writing preferences in a long blueprint. You can read all about how to generate this writing blueprint in this in-depth tutorial.

ChatGPT in particular offers a project feature that allows you to provide custom instructions and files for a subset of chats, called a project. This is perfect for adding your academic writing prompt so that all conversations within this project are conducted in an academic tone.

8. Understand the Ethical Risks

AI tools seem neutral, but they’re trained on biased data. Bias, hallucination, and embedded power dynamics affect your citations, your writing, and your worldview. If you’re researching marginalised topics or global South contexts, AI may lead you back to Western, English-language, majority-world sources, reinforcing what’s already overrepresented.

And it’s not just content, privacy matters too. Uploading interviews or field notes into online AI tools may violate ethics agreements. Some platforms retain user data for training, so ensure to always check data policies before using sensitive material. As a rule of thumb, avoid uploading sensitive or copyrighted material to AI services.

However, one of my favorite use cases for AI involves uploading my own work (that I own the copyright for) to ChatGPT in order to peer-review the manuscript before publication.

Can ChatGPT writing be detected?

While AI writing has certain properties that allow it to be detected by services like Turnitin, this process is generally very unreliable. It becomes even more unreliable as models progress. AI detectors for writing are therefore likely to become irrelevant.

Can using AI be plagiarism?

Absolutely! Large language models are trained on copyrighted data, and especially if you are dealing with niche topics, this training data is likely sparse. That means that the AI is more likely to generate a verbatim reply from the original source. If you now use this output without checking for plagiarism, you might be committing unintended plagiarism. A tool like PaperPal can help with that.

PaperPal compares your text to millions of texts found online and highlights passages that seem like they have been taken from another document, giving you an overall similarity score. Keep this score as low as possible for everything you publish and go through sources that look similar to what you have written.

9. Know When and How to Disclose AI Use

Transparency around AI use is rapidly becoming a norm in academic publishing, teaching, and supervision. That said, “disclosure” doesn’t just mean adding a line at the end of your paper. It means knowing what to disclose:

• Did AI help draft or reword content?
• Was it used for idea generation, summarising, or citations?
• Were any outputs copied directly into the final work?

In student work, unclear boundaries can lead to academic integrity concerns. In research, they can affect peer review or funding evaluation. Many researchers are now adding a short “AI Contributions” note in their methods or acknowledgements sections.

Because currently every institution and journal has different guidelines, it’s worth checking these before submitting your work.

10. Define Your Own AI Use Guidelines

Aside from formal AI use guidelines, you should consider how you want AI to run your research. Because on one hand, you can become faster, but on the other hand, you can also outsource too much of your creativity and try to solve problems with AI that it’s not good at.

Ask yourself:

• What am I comfortable automating?
• What do I need to fully understand, vs. partially outsource?
• What is tedium that doesn’t bring my research forward and that I can outsource to an AI?
• What are my red lines?

Some researchers write a short “personal AI policy”. This is a living note or doc that defines how they’re currently using tools, where they’re cautious, and what they refuse to automate (e.g. final argument structure, original insight, qualitative coding). It’s surprisingly helpful for reflection and decision-making over time.

And as AI becomes more embedded, your future collaborators, supervisors, and students will appreciate that you’ve thought it through.

Summary

You don’t need to master every new tool. But you do need to stay sharp, intentional, and grounded in what makes research meaningful and most importantly keep learning. (Here’s what I learned when using AI for my lates paper.)

If you can confidently tick off each of the steps in this checklist, you’re not just keeping up, you’re actively shaping how research evolves in an AI-powered world. The future of academia is about thoughtful, rigorous, and adaptable researchers who know when to lean on AI, when to lead it, and when to step back entirely.

Let this checklist be your baseline and revisit it often. What’s relevant today might shift tomorrow. That’s the nature of future-ready work.