ChatGPT has a bad reputation for finding literature, but only if you misuse it. In this workflow, we’ll look at using a combination of tools, including AI, to build out a literature review. We’ll collect papers in bulk with a AI-powered tools and then filter them using ChatGPT and Zotero. This guarantees reliable results from the AI, maintains researcher control, yet lets your process 100s of papers in minutes (not months!). Let’s dive in.
Learn which tools can speed up your research: This free 21-day email series introduces you to a range of academic tools and how they work together to speed up the time needed for a literature review. Most of these tools have been around for less than a few years, some use AI. Knowing them gives you an almost unfair productivity advantage for your research.
The outline of our workflow looks like this:
- Collect relevant papers using AI tools and academic databases (Litmaps, SciSpace, Consensus) or manually. If you have a collection already, skip this step.
- Use your reference manager(Zotero, Paperpile, Endnote) to export all papers and abstracts.
- Use ChatGPT (and ideally a custom GPT bot) to identify important papers
- Read and annotate papers using Obsidian.
1. Collecting Papers with Litmaps, SciSpace or Consensus
Currently there are dozens of tools you can use to discover new literature. There are essentially three types of tools that we can use:
- Keyword-based tools: Academic databases that use your query text to find papers by title and abstract (e.g., Google Scholar, Pubmed, Semantic Scholar, Scopus, etc.). This is probably the style of search you’re already deeply familiar with.
- Citation-based tools: Starting from a “seed” set of papers (even just a single one), we discover other papers referencing or referenced by the seed set. (Litmaps)
- AI-based tools: Your query is semantically compared by an AI to millions of abstracts to find the right papers.
AI-based tools are in vogue right now, but their problem is that they might miss important papers. Keyword-based tools do not have this problem but often return overwhelming results. The sweet spot is using both sets of tools to their advantage.
Here, we’ll go over one way to do this by finding a set of seed papers using AI-based tools and then finding related literature by exploring their citation networks.
1.1 Collect Initial Seed Papers with SciSpace
SciSpace is an excellent AI based tool for the job. Consensus is similar to that but works better for medical questions. It always starts with a search box where you type in your research question. Note that you do not start with keywords but with a question.
In the case of SciSpace, the result is a set of papers displayed as a table. You can have the AI analyze any particular aspect of it. For example, suppose you are an ecologist interested in the forests of northern China. In that case, you can create a new “column” and ask the AI to analyze each paper and assess whether the study area is indeed north China. This is impossible to do reliably with a keyword-only search.
Below are the predefined columns SciSpace suggests (you can add custom ones in the tool as well).
Once you have collected a set of papers that seem relevant, do not be overly critical of which ones you pick. In step 3, ChatGPT will filter through all of them and choose the most relevant ones regardless of how many papers you collect.
👉 Note: This is one of the key advantages of using AI. We can filter & prioritise papers far more efficiently, compared to traditional research requiring us to skim through 100s of irrelevant papers manually.
The easiest way to collect papers is by using the Zotero Connector. It is a Chrome browser plugin that allows you to import bulk papers from most pages and apps.
1.2 Grow Your Collection with Citation Search Engine
Citation search tools let you discover papers based on how they connect via citations and references. The most popular tools for this are ResearchRabbit, Litmaps, and Connected Papers.
In this example, we’ll use ResearchRabbit, but any of these tools will work great for this step.
👉 Check out our comparison of ResearchRabbit vs Litmaps vs Connected Papers here.
First, let’s export the papers from your reference manager. You’ll want to export as a BibTeX, and then use this to import into ResearchRabbit.
ResearchRabbit will generate a graph that shows your input papers (with seed icons) plotted alongside new recommendations (hollow circles). This is how you can discover “related articles” connected by the citation network.
You’ll see all the paper on a 2D map. At the top are papers with more citations, and to the right are recent papers. Publication date and citations are negatively correlated; your papers are roughly arranged along a diagonal (top left to bottom right). Look at the recommended papers at the top right for the most impactful and recent papers.
This process is explained in breadth as part of the Effortless AI Literature Review course.
Keep collecting papers and add any interesting ones to your Zotero collection.
At the end of this fun, “shopping-like” process, you might end up with 100s of papers. Before AI, this would have been a giant problem because nobody had time to read so much. Even skimming hundreds of papers would be incredibly time-consuming. Instead, we’ll automate this information extraction step, and even add extra value to it with AI (in the next section).
2. Prep Your Papers for AI
To have AI look at our entire collection and help us decide what is relevant, we need to export everything we have collected into a readable format. Every reference manager can export to “BibTeX,” a universal citation format. In most cases, it will contain abstracts of the papers as well. Here is how to do it:
The resulting file has an “*.bib” extension and will not be readily read by ChatGPT or your system. Renaming it to txt will reveal its contents:
You can use any reference manager for this step since they all support an export to BibTeX, and the format looks the same. However, sometimes Zotero will not be able to retrieve the abstract of a paper and ChatGPT won’t be able to use it either. You can find out if you look into the file. Each paper begins with @article or @techreport, followed by a few properties. If the “abstract” property is missing you can either add it manually or just keep in mind that this paper will be less likely to resurface in your AI analysis.
3. Use ChatGPT to Create Your Reading List
The final step is to upload all these abstracts to ChatGPT and ask it to create a reading list concerning our research question. The more clearly and detailed you define your research question, the better AI replies.
Upload your bib/txt file and use the following prompt:
My goal is to learn how {insert your goal here}.
1. Analyze the papers mentioned in the document. Use only them and nothing else.
2. Create a reading list that takes me from novice to advanced. Starting with broader papers and ending with very specific papers.
3. For each paper identify 2 questions as my "learning goal" based on what you infer in the abstract.
Choose only 5{insert your number here} papers from the collection, as I do not have more time.ChatGPTs responses will be stochastic so it might be worth to run the prompt a few times and see which papers come up multiple times to be sure.
The most important part of this prompt is to provide you with a list of “learning goals.” Look at all learning goals and assess how they relate to your research question. Do they answer it?
If you feel something is missing just ask ChatGPT a follow up question: “What paper is likely to answer {insert your question/goal}”. Given so many abstracts on a topic ChatGPT became somewhat of an expert on your field.
💡 Advanced Tip: Instead of writing prompts and uploading files to regular ChatGPT, try using a custom GPT. You can design your GPT (or an agent) to answer any questions regarding this literature review to really supercharge your workflow.
4. Time to Read!
The last step is to read your papers. No amount of AI will save you from critically reading your papers. But now, you have only a small set of documents to read and won’t be overwhelmed. Read 1-2 papers a day, but read deeply. Here are eight tips on becoming the best read person.
But, reading isn’t as simple as it sounds. There’s one persistent challenge we all have: taking good notes on what we’ve read. “Good notes” means you can:
- Easily find notes,
- quickly synthesize topics,
- interconnect ideas,
- and grow understanding over time.
The best software to use here is Obsidian. I have a free 21-day course to learn the basics and an in-depth course to make you a note-taking expert in academic matters. Obsidian is to knowledge what Google Maps is to navigation. It is the perfect tool for academics dealing with a lot of information.
Unsure if you’re ready to drop Notion, Evernote, MS Word, etc. quite yet? That’s fair. It’s a big choice to switch note-taking softwares. To learn more about why Obsidian is uniquely suited for academics, check out my in-depth guide here.
Summary
In this rapid AI-powered lit review workflow, we looked at how to find a few papers using AI-powered tools (Consensus, SciSpace) and academic databases. Then, we collected all the documents using a reference manager (i.e. Zotero). We explored the citation network of our papers with Litmaps to find as many related articles as possible, without being overly picky. Lastly, we used ChatGPT to do the hard work of skimming all the abstracts for us.
The end result is a well-curated list of the most relevant papers tailored to your exact research question with a set of learning goals per paper to guide your lit review.
Using AI in research today is all about automating tasks, remaining in control, and improving our research skills. Hopefully, this workflow can help you elevate your project to the next level, while giving you the confidence that your sources are robust and reliable.
