Using genAI for the literature review

Is there any point in doing a literature review now that genAI can do it for you? That’s the timely question raised in a recent article in LSE Impact (“Why write a literature review if AI can do it for you?”) The authors (Krlev, Hannigan and Spicer, all based at business schools) distinguish different types of literature reviews, en route to their answer. For their paper, they only looked at management articles, but I think their findings have a validity for literature reviews in any subject.

The reason for the question “Why write a literature review?” is, of course, that genAI can certainly provide any survey of the literature for you, and in lots of detail, if you keep prompting it. If you ask for the best studies on the causes of the American Civil War, any of the major tools will give you an extensive, annotated reading list, with detailed comments.

What is intriguing about this output is that it resembles almost perfectly the reading list any lecturer will give their students at the start of a unit: a list of works, with a sentence or two of commentary, intended to provide a starting point for the student’s own reading. Nobody complains that the reading list handed out by institutions to their students is a form of cheating!

So, 1-0 to genAI, I would have thought. Is there anything wrong with the above? Krlev et al’s article concludes that the literature review is still worth doing – but we have to dig a bit deeper and establish just what the literature review involves., and this is the interesting aspect of the article, which is itself a literature review, of literature reviews. The authors identify ten kinds of literature review, which they describe as “review styles”. They depict these styles using quadrant analysis diagram. All literature review articles are placed in one of four quadrants, based on the two axes “assembling – disassembling”, and “representational – generative”.

Elsewhere in their article (which, at 15,000 words, is long for a single article), the authors distinguish ten purposes for a review:

1. Envisioning: “turning tensions into productive, alternative paths”
2. Conceptual: “connecting other fields”
3. Modeling: “underpinning how methodological gaps hamper theoretical progress”
4. Debate: “providing a balanced view, pros-cons, etc”
5. evidence-oriented: “overcoming contradictory evidence”
6. synthesis: “broadening perspectives … illustrating logical next steps”
7. systematic: “dealing with bias, increasing rigor”
8. perspectives: “offering alternative interpretations”
9. taking stock: “consolidating and summarizing”
10. content dominant: “thick description of a subject area”

This long list doesn’t seem very helpful. I can’t see the need for ten categories – some of these clearly overlap (“debate”, “perspectives”, and “taking stock” look very similar to me).

While the attempt to differentiate different types of review is a valid one, I think for most purposes two strands will suffice. I would have thought most literature reviews can be placed in one of two categories: what we know, and what we need to know (or do) next. Review articles might be formulated in a more or less systematic way, but that doesn’t make them a different type of review.

Where does genAI fit with all this? Sometimes you have to remind yourself that Claude hasn’t read all the books it talks about. All it knows is what others have said. Hence AI is good for the former (what we know) but when it comes to the latter, all it can do is to pass on any other proposals from the corpus. The classic research methodology is to reshaping and expand existing thinking – and that is where the human takes over.

For an analogy to the literature review, consider Darwin, while writing On the Origin of Species, trying to establish how bees manage to create perfect hexagonal-shaped cells for their hive. He began with a literature review, which meant reading any published work on the life of bees. He also spoke and corresponded with as many bee experts as he could find. At that point, he made a hypothesis (that the bee makes a circular shape, which when added to others formed a hexagon) and tested it. Mathematicians subsequently looked at the hexagon shape and decided it was the best possible arrangement of resources for the task. But bees are not mathematicians, and even the best genAI would not have been able to tell him, or the bees, how to create a layout for the hive when none existed.

So, all credit to Krlev, Hannigan and Spicer, for investigating types of literature review, and ideally we should state with precision what machines can do, and what humans should do with it subsequently. Their recommendation, actually two recommendations, is “experimentation with technologically infused practices”, that is, make use of genAI, plus “engage in collective and distributive practices”, which appears to mean “communicate with others about your assessment of the literature”. That is how Darwin worked, but is not addressed in this article – it’s a research topic in its own right. But I think there is no question that genAI should be used (and acknowledged) when doing the review. And if we insist on a public acknowledgment, perhaps every student essay should now include the words “this essay took into account the following reading list(s) supplied by my tutor”.