With major strides forward in AI (artificial intelligence/machine learning), computers are increasingly able to produce music, images and text. So you might wonder if soon we’ll have an AI that can write your PhD for you. This is not exactly my area of expertise, but I know enough about the technology to be able to have some educated guesses. I’ll also be weaving in something that I do know a lot about, the history of doctoral writing technologies.
Okay, to start with, the current AI text bots are able to put together grammatically correct text, and increasingly idiomatically correct texts. However, these texts are typically short, and they are either so generic that they are of course true (they reproduce ‘common knowledge‘), or they are little fictions (which are fun but not robust claims about knowledge). So the programs are not yet able to produce 80,000 words of logical and defensible thesis.
Secondly, the programs work by taking in thousands of words of example text and then learning to reproduce it. Since our PhD theses are written about new knowledge, they are original contributions. No AI will have access to that knowledge until you produce it through your own research and creating your own texts. So even if the AI was eventually able to help you, say, put together a first draft of your lit review based on your notes and summaries, you would still have needed to do all the reading and made the notes and added your scholarly opinion on the readings before it could help you by collating it.
Thirdly, AIs take months of intensive work to train, if you want them to produce meaningful outputs. Considering how many hours you would have to spend coding and correcting your AI to produce meaningful text… it is probably still faster to write your own thesis for now.
There are some aspects of the thesis that I do expect to be automated soon enough, though. We already have automated grammar and spelling checkers, and reference managers that input our citations in the specified style. I wouldn’t be surprised to start seeing more sophisticated tools that use machine learning that would allow us, for example, to ‘rewrite this chapter in the style of The BMJ‘, and then when it gets rejected to ‘rewrite this article in the style of The Medical Journal of Australia‘. The rules are already clearly explained and set out, there is a defined corpus of example texts, and we have already produced the important intellectual labour.
For people who like to edit rather than produce first words, even very imperfect text by the AI bot might be helpful to write over and correct. Or the machine might produce random results that inspire creativity (similar to some composers and artists who integrate random number generators, creativity cards etc, as creative prompts or structures for their process).
This would accelerate changes which have already impacted how ‘hard’ it is to write a PhD thesis. If you ever get an old PhD thesis out of the library, even just from the 1980s, you will be surprised by the differences. They are usually written on a typewriter, and small errors are fixed manually (with pasted words or by hand). There are typically a lot fewer sources cited, since books and articles were only available physically, and you found the sources by sitting down with the card catalogue or the printed annual index for each journal and reading through them searching for relevant titles. All the citations had to be done manually too. The figures and tables are painstakingly manual and so there are often fewer of them and always in black and white. You had to get three copies of your thesis printed and bound for the examiners, and then bound again in hard copy for the library.
Today, it’s easy to get R or Excel to format your figures, Endnote to insert your citations, you found the books and articles you needed by typing a vague term into Google scholar or you university’s database. Your computer checks your spelling, formats your text to look ‘publishable’, adds page numbers and automatically creates your table of contents. You might not ‘type’ your drafts, but dictate them into your machine. You upload the finished document as a pdf to your university portal.
While this makes parts of the PhD ‘easier’, we have maintained the challenge of the degree by shifting the expectations. Because you can find so many more articles, so quickly, I very much expect that you will. Because it’s so quick to produce figures and tables from your data, I’m concerned if you seem to have skimped. I won’t give you much leeway if you have a few spelling errors or citation mistakes, or messy formatting.
More concerningly, AI will be making differential contributions to different candidates. Experimental science PhD theses, which follow a strict format and are typically ‘writing up’ data, will probably get the most benefit from AI text assistants. I’m not sure that it makes much of a difference, once you set out your methods in order, whether a researcher or a machine puts it into sentences. However, a researcher who is iteratively creating their methods (like a social scientist using grounded theory) or who creates their data through their writing (like textual and qualitative researchers), is still going to have to write multiple drafts as their scholarly process. Moreover, anyone undertaking interdisciplinary research is always creating a hybrid and original form for their thesis and this is unlikely to be easily produced by a computer.
This would be yet another way that machine research tools disadvantage creative, arts and humanities researchers. For example, there are already tools to help you do your lit review, but they are not useful in my research yet (even where I know there is lots of material that I can find the old fashioned way). Databases like Web of Science and citation trackers like Google Scholar often do not trace citations in books. Nor does text matching software like iThenticate use books as sources. This means humanities scholars like me don’t get the same benefit of using the tools to help us find sources, or check our references, and are disadvantaged by metrics that judge our impact through citation software. And that’s not counting the way many of my secondary sources are in books that are not held in libraries in Australia and were published before ebooks; or the fact that I include a much wider variety of difficult-to-cite sources which reference managers often struggle to support.
Such an AI would also exacerbate exisiting social inequalities between researchers. Currently, most literature is produced in English and from research organisations in the developed Global North. Researchers who work in languages other than English, or who are writing from the Global South, or in small disciplines, or in specialised fields, will find that the AI texts are less likely to be high quality or directly useful.
I’m so grateful every day for word processing and spell checks and pdfs of articles and searchable library catalogues and ebooks and my wireless printer and the backspace key and all the hacks in Word that make my writing easier. These tools support more writers to write, and to focus on the human, creative and scholarly aspects of the writing process. I look forward to us thoughtfully and creatively incorporating new machine learning tools into our workflows. Nonetheless, we all know, you can’t just run the spellcheck or grammar check and accept all the changes. You need to double check the computer’s work. This will continue to be true when AI is more prevalent and able to make a bigger contribution to your doctoral journey.
But an AI is not going to write your thesis for you yet, and if we ever get to a time where it could, we just find a different way to make the PhD hard–because you will still need to make an original and sustained contribution to knowledge, however it gets onto the page!
If you want a different take on our PhD AI future, Thesis Whisperer Inger Mewburn has just written about her thoughts.