Using AI for Literature Review in 2025

Despite AI making giant leaps and advancements, the methods used in conducting a literature review remain similar to those of decades ago. In this post, I want to share with you a powerful workflow consisting of various tools, suitable for multiple steps of the literature review. If you integrate some of those, you might complete your literature review more thoroughly and quickly.

Here is an overview of how our literature review workflow will look. We will go through every step of the diagram, and I will explain which tools you can use, providing you with further resources along the way.

Depending on the knowledge you have of your topic, you will take a different path through this diagram and use different tools.

AI Lit Review Step 1 – Dive into a Topic with Undermind

If you have no idea about a topic, the easiest and free way to dive in is to run a deep search with Undermind (5 free per month, with some limitations), Consensus (paid), Scispace (paid) or Liner (paid). This will generally give you a good overview of the subject matter, even though the results might not be complete and the papers might not be the most impactful ones. However, it is a perfect starting point that can save you hours if you don’t know anything about a topic.

The reason why Undermined is recommended in this graphic is that it is the only one that will generally give you the most results for the least money, but unfortunately, it will not show all of them if you are a free member.

Goal/Result: Find two or three good papers and build an understanding about the main questions and topics of your literature review.

AI Lit Review Step 2 – Finding Key Papers With Litmaps

Well, most AI tools struggle with prioritising papers. Litmaps made it its core tenet to give you the best overview of the impact of a paper. To achieve this, it utilises a visually appealing graphical display, where the most important papers are located on the top right, indicating high citation counts and recency.

Litmaps is one of the only tools where your input is a few papers and your output is more papers from the same domain. To find these papers, Litmaps uses the reference network, which sets it apart from AI tools that use textural similarity to discover related papers. Textual similarity is a technical concept developed by computer scientists, while references are set by the scientists who write the papers and therefore are more trustworthy.

Goal/Result: A larger set of papers fitting the scope of your literature review. Ideally, you have nearly all the papers you need. Aim at not more than 10-20 papers and prioritise new, highly cited and review papers.

AI Lit Review Step 3 – Picking the Right Reference Manager

Depending on whether you write with Google Docs or Microsoft Word, you should pick a reference manager. If you’re working with Google Docs, Paperpile will deliver all your PDFs directly to Google Drive, making them accessible using Google’s Gemini AI. If you don’t use Google Docs, you can choose to use Zotero, because it is free and has plug-ins to analyse your papers with AI. This workflow is part of the Effortless Publishing course.

Regardless of which reference manager you picked, you need a place to store your notes and insights. Obsidian is the go-to tool for taking notes. Here is the system that I use for note-taking. Want to get started? Sign up for the free 21-day course where I’m going to show you how to use digital notes to supercharge your literature review.

Paperpile integrates into Obsidian via the Citations plugin, and Zotero uses the Zotlit plugin. The note-taking course explains exactly how to do this best.

Almost every literature review will yield more papers than are needed to answer the question. Therefore, aim at collecting important papers and recording why you want to read them. Documenting the intent of why you want to read a paper allows you to discard papers that became irrelevant during the course of your literature review research. You can use a visual reading list to organise your reading:

If you end up with too many problems, you can check this tutorial on using AI to prioritise your reading and make the literature review much more efficient.

Goal/Result: A set of notes on the topics handled in the papers you have read. If you follow my system, the notes should be on topics rather than the papers themselves, which allows for building a knowledge network and fosters creativity.

AI Lit Review Step 4 – Reading and Understanding Papers

AI can greatly accelerate the process of skimming papers by creating outlines and summaries. The real strength comes if you create summaries on sets of papers rather than single PDFs. two tools are noteworthy for this step. SciSummary allows you to create contrasting summaries or analyse papers in bulk regarding a specific topic. Here is a review of SciSummary. Anara, on the other hand, goes deeper into each paper, allowing you to find related references to text passages and create a note collection on your papers. Here is a deep dive into Anara’s features.

Working with papers in the scope of a systematic review is usually a bad use case for AI, because in a systematic review, your search procedure needs to be reproducible. AI search, however, is always somewhat random. Nested Knowledge tries to solve that by merging AI and keyword-based searches for large systematic literature reviews. Here is the Nested Knowledge Deep Dive.

Goal/Result: reading the papers that you have found and creating notes on them. Fostering a deep understanding of the subject matter and using AI to answer questions regarding the papers.

AI Lit Review Step 5 – Synthesising Insights And Notes

Once you have built an understanding around your topic, your goal is to synthesise these ideas into a coherent narrative, pointing out the gaps. Treat AI here more as a search engine rather than a synthesis engine. The synthesis happens in your mind, while you can leverage AI to quickly ask questions to the collection of PDFs and notes that you have aggregated. There are generally many ways to do this:

1. If you have little notes but many PDFs, try Notebook.LM to synthesize information and generate mind maps, podcasts, presentations, or timelines directly from your source PDFs. Here is a tutorial on using Notebook LM. If you have a Notebook LM subscription, you can also use Google’s Gemini, which perfectly integrates into Google Docs creating a powerful writing assistant.
2. If you have written many notes in Obsidian, use Obsidian’s Co-Pilot plugin, which brings AI directly into your notes and can create new notes as well as synthesise existing notes. Here is an in-depth tutorial on Obsidian’s co-pilot.
3. ChatGPT also recently introduced a feature called “Projects”, which allows you to upload up to 25 PDFs or notes and query them using AI with customised instructions. This can be a powerful way to write a discussion section for your paper.

Each of these use cases uses AI to process information in academic papers. The job of a scientist becomes to ask the right questions, to synthesise a high-level overview of the topics, and let AI handle the nitty-gritty details. One approach that proved very valuable to me is to deconstruct papers into small atomic statements and reassemble these statements into a narrative of my own:

However, it is a skill to outsource the right work to AI without losing control and understanding. The best way to gauge if you’re using AI properly is to see if you are operating in a state of flow, which is an indicator that your mind is operating optimally. If you feel that you’re waiting too long or being distracted, it means you’re probably relying too heavily on AI. If you think that you’re not making progress, that probably means that you can outsource some of the mental tedium to your AI companions. But regardless of where you’re at, AI cannot learn the topics for you.

Goal/Result: Creating a narrative for your literature review based on the papers you have read and using an eye to help fill in the gaps.

AI Lit Review Step 6 – Writing And Polishing Manuscript

Once you have created a narrative, it’s time to start writing your literature review. The easiest way to integrate AI into this process is by breaking your narrative into small statements. Each statement consists of a factual expression and a reference (e.g. “Climate change causes sea level rise (Smith 2020)”. If you assemble dozens of these statements, you can use them as the building blocks for a bigger narrative of your academic paper. Here is how it looks laid out on a mind map in my vault:

These atomic statements can also be extracted using ChatGPT. Once you have broken down the narrative in this way, writing up a few paragraphs of text becomes very easy. You can combine the relevant atomic statements and use ChatGPT’s Canvas feature to iteratively improve your writing paragraph by paragraph. If you’re using ChatGPT and you find its writing mediocre, you can train it to write in a more academic style using the writing blueprint method.

And the last step of the literature review, you can use a tool like PaperPal or custom GPTs like the Academic Assistant Pro to polish your writing and improve it to academic standards.

In my opinion, writing is the part of the literature review process that can be most strongly automated and might even become obsolete in the foreseeable future. Some academics believe that AI writing is inferior and could be detected by AI detectors. This is, however, not the case for already more than a year, and AI is advancing very rapidly.

Here are my lessons writing a paper predominantly with AI:

AI Lit Review Step 6 – Peer Review

The last step of the literature review could be considered a fail-safe, where you upload your finished manuscript to AI and let it search for relevant papers and then critically peer-review your writing. The tools PaperWizard and Liner both do an excellent job at this:

Another way that you can use to check if you have found all the right references for your literature review is to upload your entire bibliography for the lit review to Litmaps. If you remember, Litmaps was designed to find papers similar to the papers you give it. If you upload your entire bibliography and it still finds relevant similar papers, we could consider those reference gaps. This is a very powerful method to check what you might have forgotten. It takes mere minutes:

Summary of the AI Literature Review Process

Many tools can help you find the proper papers, synthesise information, aid in understanding while reading, peer-review your manuscript, and assist with writing parts of it. The key to using them effectively is to treat the AI as a tool rather than an employee. While you can expect an employee to be responsible for the outcome, AI usually has no goals of its own. But the more detailed your descriptions and objectives are, the better the job this tool can do. A large language model is nothing more than a piece of software that processes information in a sophisticated way. And just like with software, the more niche and detailed the task is, the more reliable the results. Avoid over-relying on the all-in-one solutions where you don’t have to put in any effort. No effort usually means no understanding.

The Best AI Tools for Conducting Literature Reviews in 2025

In 2025, the top tools for writing a literature review would be Litmaps, Consensus, Liner, Obsidian Copilot, and ChatGPT/Gemini. A subscription to all of them would cost you under $100 monthly, but save you dozens of hours if you are an active researcher. All of them have free versions that are powerful enough for occasional use or to complete undergrad-level assignments. If you were to pay for only one tool, opt for Gemini because it provides access to Notebook LM and integrates perfectly with Google Docs, where I write all my papers, allowing me to cross-reference the manuscript with any PDF in my Google Drive.

Best AI Literature Review Generator

There is no best AI literature review generator, but there is a best way to use AI for literature review generation. Tools like Liner, Consensus, SciSpace or undermind aim to be a one-stop shop for literature review: They start from simple queries, search for papers, read them (if they get access), and synthesise the information into a literature review document with citations. This approach, however, has serious problems you should be aware of when using these tools:

Problem 1: AI Struggles to Prioritise Papers In Your Literature Review

AI is often not good at prioritising papers. A peer-reviewed Q1 publication might be weighted the same as a pre-print that has not been peer-reviewed. This can compromise the output and make it less trustworthy. However, all of these tools will display the number of citations or the journal ranking in the results, allowing you to double-check.

Problem 2: The Success of Your AI Literature Review Depends on the Prompt

Large language models have been trained to fulfil users’ requests and are generally very supportive. So if your prompt implies a particular theory (e.g. “Find evidence for climate change causing the loss of biodiversity”), AI will try to affirm that theory rather than give you a critical assessment of your ideas. You can always double-check this by formulating two different prompts that contradict each other and check the output. Another way is to avoid this altogether by asking very open-ended questions (e.g. “what are the effects of climate change on biodiversity”).

Problem 3: Literature Review AIs won’t read most papers

Reading a paper takes you time, while it takes AI computational power. Every company tries to minimise computational costs in order to maximise profits. It is part of AI’s training to generate responses faster with less material. However, this can be a problem if you are looking for an exhaustive search that encompasses most or all the literature on the topic. AI companies solve it by offering different tiers of search. Here is an example from Consensus.

Generally, you will have to pay more for more expansive searches. On the example of Consensus, the quick search with 10 papers is free. The pro search with 20 papers costs $10/month and also includes a few deep searches monthly. Higher packages, including deeper searches or AI agents, can go up as much as $90/month.

This means that the more thorough you need your research to be, the more you have to pay. In some cases, it might be possible to run multiple simpler searches on related queries and synthesize the answer yourself.