How AI reconfigures literature search and discovery

Post by Lynette Pretorius.

I recently presented a keynote on AI in literature search and discovery at an institutional seminar. Many people have asked me to share the presentation, so I thought I’d write this blog post to make it available to all.

I would like to start by getting you to think about how you normally search for literature. For most of us, literature searching is a laborious process of defining a calibrated set of keywords combined with Boolean operators (AND, OR, NOT), and perhaps the use of the asterisk to truncate certain terms. For different databases, we often need to combine the keywords in different ways to create a usable search string. You need to know how to speak the database’s language to ask a question that gives you the responses you want. I have included an example of such a search string below, which comes from this paper. You can see why I say coming up with something like this can be somewhat tedious…

(“pre-service teacher” OR “preservice teacher” OR “student teacher” OR “student-teacher” OR “pre-service teacher educators” OR “teacher educator*”) AND (“creative thinking” OR “critical thinking” OR “evaluative thinking” OR “value judgement” OR “decision making” OR “abstract thinking” OR “abstract reasoning” OR “logical thinking” OR “problem solving” OR “problem-solving” OR “Creativ” OR “creative development” OR “creative teaching” OR “creative thinking” OR “productive thinking” OR imagination OR “higher-order thinking” OR “higher order thinking” OR “analytical thinking” OR “reflexive thinking” OR “reflective thinking”) AND (“artificial intelligence” OR “generative AI” OR “generative artificial intelligence” OR genai OR “Gen AI” OR gai OR “Large Language Model*” OR llm* OR chatgpt OR gpt OR chatbot* OR copilot OR grammarly OR bard OR gemini OR claude OR consensus OR perplexity OR “Deep AI” OR dall-e OR midjourney OR leonardo.ai OR canva OR consensus OR elicit OR scite.ai OR “Research Rabbit” OR chatpdf OR “Semantic Scholar” OR iris.ai OR “Paper Digest” OR scispace OR notebooklm OR ailyze) AND (“teacher education” OR “ITE” OR “teacher preparation” OR “initial teacher” OR “teacher training” OR “bachelor of education” OR “master of teaching” OR “master of education” OR “teacher progression”)

With the introduction of AI, this process has become entirely reconfigured. Since the introduction of computers, you have had to be able to “speak” the computer’s language to get it to do something for you. AI, however, reverses this dynamic, allowing us to converse in our own words while the computer performs the translation. I have found this particularly useful in the literature search and discovery process. Let me give you some examples. I will preface this by saying that these tools are the ones I have used, but there are many others. You should find the ones that work for your workflow.

My personal favourite for literature discovery is Google Scholar Labs. It is easy to find – if you go to Google Scholar, there is usually a notice in the middle of the page that suggests you try Labs. Additionally, the current screen of Google Scholar also has a “Labs” button at the top left-hand corner. It is also completely free to use. I like this one because it is an interdisciplinary database and my work touches on many different fields. It works pretty simply: you ask it a question and it will search Google Scholar to find relevant literature. In my presentation, I used this question as an example:

“Find papers that reference the hidden curriculum in doctoral education, particularly using autoethnography as a methodology. Privilege studies from the Global South.”

This example highlights a key benefit of these AI search tools: you can ask more complex questions instead of trying to match keywords. In this example, I asked for literature on a topic (hidden curriculum in doctoral education), but I also asked it to find a specific methodology that is often less visible in research because it focuses on individual experiences (autoethnography). I also asked it to find articles specifically focused on Global South perspectives, another element I have found is often hidden in general searches. Within a few seconds, Google Scholar Labs found several articles for me. It listed the articles, including a quick summary under each article and its reasoning for why it met my search criteria. Then it shows a “Quick read” feature that gives a more detailed summary of the article and, if you have your institution connected to Google Scholar, lets you click the PDF link to read the full article.

I have also used Elicit, particularly for literature extraction. You can ask it a question and it will find articles for you in a very similar way to the example I gave above. It will then also give you an overarching summary of the literature it found in a kind of mini literature review. In the example I used in my presentation, I asked it what the commonly listed hidden curriculum concepts of doctoral education were. It did a literature search, found twenty-five articles on the topic of the hidden curriculum and identified six major thematic categories: “power dynamics and structural marginalisation, professional socialisation and identity development, supervisor-student relational dynamics, institutional culture and unwritten rules, emotional and psychological dimensions, and critical thinking and epistemological challenges”. It then went on to provide a summary of what these thematic categories included, with links to the relevant articles for me to do further reading. While certainly not a finalised literature review, this is a great way to get an overview of a field and then dive more deeply into the literature relevant to your specific topic. Elicit can then also extract information for you into a literature grid, similar to what you would be doing if you were doing a scoping or systematic review. Elicit is a bit pricey though, especially if you want to use it for the literature extraction component.

Research Rabbit is a tool that visualises connected papers. You start by providing it with the DOI of an article and then the AI searches for related papers to the one you provided, giving you a visual representation. Personally, I find it interesting to use to see where my work sits within the wider field, but I also find it entertaining to use because it has an AI rabbit that joins you as you “dive down the rabbit hole of literature discovery” (as it proudly announces on its front page). Sometimes you just need some amusement when you have to wade through hundreds of articles on a topic. You can use Research Rabbit for free to visualise related papers, with the higher-priced tiers having additional features.

Our university’s library has Scopus AI as one of its databases, so if you have access to that one, it works in a very similar way to Google Scholar Labs; it is just limited to the Scopus database specifically. When I asked it to list the common concerns related to doctoral education for my presentation, it came up with ten concerns listed in what it felt was the order of importance, providing links to the relevant papers for each.

The particularly useful aspect of the four tools I listed above is that they are focused specifically on academic sources. You can certainly also ask the general large language models for help with literature discovery. I have, for example, used ChatGPT about the common concerns related to doctoral education and it gave me a very useful overview (you can see its responses here). The linked example is from 2023 when ChatGPT was first released. I redid this same search for my recent presentation and it came up with a similar response. The difference was that it now also provided some links to sources. I have also recently tried using Gemini and it provided an excellent response with many relevant sources for further reading.

One of the most eye-opening examples I used for the presentation was deep research mode. In this example, I provided Gemini with the introduction, literature review, theoretical framework, methodology, and findings of a paper I wrote and asked it to use my findings and the literature more broadly to suggest a discussion section for the paper. At this point I had already written my whole paper, so I thought it would be a good way to test the AI’s capacity. The deep research agent worked for about 25 minutes and produced a 3,000-word discussion that was well-written, reasonable in relation to my findings, and supported with reference to appropriate literature.

It is always important to consider the potential for hallucinated ideas when using GenAI models , so if I use these models, I always ask the AI to explain its reasoning. This is a very useful way of evaluating the accuracy of its suggestions, as well as encouraging you to use your own critical thinking to think more deeply about the sources. If it provides you with links to sources it thinks are relevant, click on them and read them yourself, don’t just assume that its suggestion is factually correct. My deep research example above provided its thinking throughout the process:

• Mapping the hidden architecture: I have begun synthesising the core findings from your research on the ‘hidden rules’ of doctoral education. My initial analysis of the data reveals a complex ‘moral architecture’ where legitimacy is not just earned through scholarship, but it is often contingent on navigating unspoken social codes. I am thinking through how supervisors, institutions, and publishers act as primary gatekeepers who define what it means to be a ‘good’ scholar. These insights highlight a significant tension between formal institutional policies and the lived, often precarious, reality of doctoral candidates.
• Bridging Policy and Practice: A major focus of my current reasoning is resolving the contradiction between the university’s rhetoric of wellbeing and the structural pressures of marketisation. I am examining how norms like ‘endurance as worth’ and ‘opaque bureaucracy’ create significant information gaps for students, particularly those from marginalised or international backgrounds. By aligning these findings with your theoretical framework of academic doxa, I am uncovering how these tacit expectations can lead to self-censorship and emotional exhaustion, which your participants so poignantly shared.
• Developing Actionable Reform: Moving forward, I will be looking deeper into recent scholarship of doctoral education reform to ground your proposed institutional responses in the current academic landscape. Researching websites…
• It then provided links to 24 academic sources for further reading.

Remember to always critically engage with outputs, as per our ETHICAL AI Use Framework. Consider the cautionary tale of bixonimania, as exposed in a recent Nature article. A researcher uploaded two fake preprints detailing a completely fabricated eye condition, intentionally inserting absurd details such as funding from “the Professor Sideshow Bob Foundation”, “the University of Fellowship of the Ring” and “the Galactic Triad”. There was an acknowledgement to a staff member from “Starfleet Academy” for work done “onboard the USS Enterprise”. They even stated “this entire paper is made up”. Despite these glaring red flags, multiple artificial intelligence systems ingested the fake data and, within a month, presented it as legitimate information. A human scholar subsequently incorporated the fabricated reference into a peer-reviewed publication without reading the primary text, which was subsequently published. This ecosystem exposes a concerning pipeline: fake research -> AI ingestion -> AI output presented as fact ->uncritical human citation -> real publication. No human verified the information at any point in the chain. This is why critical evaluation is so important.

It is also important to acknowledge how you have used AI in your research. If it helps, this is the AI acknowledgement I currently use in my work:

I acknowledge that I collaborated with Gemini (Google, https://gemini.google.com/) and ChatGPT (OpenAI, https://chatgpt.com) during the preparation of this paper. These generative AIs acted as critical friends, helping me brainstorm and critique some of my ideas, refine my phrasing, and reduce my word count. Any suggestions incorporated into this paper were adapted to reflect my own style, voice, and ideas, and were further refined during the peer review process. I take full responsibility for the final content of the paper, noting that it represents my original ideas and adheres to academic integrity and quality requirements.

Finally, remember that your use of AI necessitates the development of AI literacy. It is when you start seeing the AI as a collaborative thinking partner, rather than just a tool, that you become AI literate. If used in this way, AI becomes a social interlocutor – another voice in the room that helps support your thinking and creativity in the research process.