You may have seen or used various AI tools that convert academic outputs into podcasts.
These can be tempting. For readers, they promise to cut down on your literature review time and make academic papers more ‘lively’ by converting the written content into more palatable formats. After all, it’s usually easier to listen to a podcast on a morning stroll or while making dinner than it is to sit down with a heavy piece of writing.
For writers, these tools bring the lure of widened engagement and impact beyond academia. A few clicks puts your research (or something generated from it) on Spotify or in shareable posts for social media.
But are these podcasts any good? Can we trust them (as readers) to deliver accurate discussions of complex academic content? And can we trust them (as writers) to fairly and accurately represent our hard work?
To find out, I tested the technology on my own PhD thesis: a ~90,000 word document entitled Recycled Alterity: Familiar Dehumanisation in the Contemporary Fiction of Genetic Posthumanism.1 The whole document is in an online repository, so I was able to create podcasts straight from the link.
In order to compare how different AI tools would handle the job, I tested two that will generate podcasts for free:2 Google’s NotebookLM and Academia.edu. Read on for the results of the experiment:
- Reviewing NotebookLM’s podcast
- Reviewing Academia.edu’s podcast
- Should you use AI-generated research podcasts?
Reviewing NotebookLM’s podcast
Creating the podcast was easy. I simply opened a notebook, entered the link to my thesis, and selected the ‘Audio overview’ button. NotebookLM gave me lots of options for how I wanted my podcast: in brief, as a deep dive, critique, or debate. There was a choice of languages, lengths, and perspectives; and I could even specify what the hosts should focus on. I selected a deep dive at default length, and didn’t customise beyond that. The 23 minute podcast generated in about 5 minutes, and I could download it for free.
My podcast ‘deep dive’ started with a metaphor: the ‘hosts’ talked about planned obsolesence in smartphones, and how advancing technology (which we are socialised to see as an improving force) might sometimes diminish our experience as an older phone’s performance can be degraded – often deliberately – with updates. This is not something I talked about in my thesis at all, but the ‘hosts’ applied this metaphor as a way to illustrate genetic engineering being used to ‘degrade’ humans for profit. This is actually a pretty good metaphor, and one I wish I’d come up with! A strong start.
The ‘hosts’ then dropped a slightly obsequious compliment on my “brilliant piece of work” (is it strange that this still felt good to hear, even though it’s not human validation?). More importantly, they nailed the central premise:
Authors establish this [dehumanisation of clones] through a literary device the thesis calls ‘recycled alterity,’ right? Alterity meaning otherness. Right. It’s a really sobering analysis. They’re literally taking the exact chilling historical blueprints used to justify real-world atrocities, and recycling that rhetoric. And applying it to fictional clones.
The discussion broadly followed the structure of my thesis: starting with a contextual overview of real genetic engineering research and its applications, and then moving into an analysis of posthumanist theory and fictional representations of genetically altered bodies, with a focus (as in my thesis) on what the ‘hosts’ called “the mechanics of power and control”.
The thesis content was articulated surprisingly accurately, but (as you might expect) couched in the language and conventions of podcasts. The two ‘hosts’ split the discussion, interjecting into each other’s summations with stock phrases like “let’s unpack that” or “right, and…” or “here’s where it gets really interesting.”
At times there were phrasings that made me think ‘hmm, I wouldn’t have put it quite like that..’. or ‘that’s overstating what I wrote…‘ But they were usually quibbles, rather than critical misrepresentations.
The only thing that I would consider a significant flaw was the way the podcast applied the thesis discussion to real people. My thesis focuses on fictional clones and genetically engineered characters, not actual, living people in the world. But the podcast title (“Who Owns Your Patented Body”), and then much of the content, applies the ideas to ‘your’ or ‘our’ biology. It even goes so far, in the conclusion of the podcast, to predict a scenario in which you might have to accept an end-user licence agreement to access a genetic update. In my opinion, that is extrapolating too far from the content of the thesis and creating a false impression. But these moments of editorialising were limited, and didn’t detract too much from what was otherwise a decent summary of my research.
Verdict: I’m pleasantly surprised! The NotebookLM podcast was a good enough representation of my thesis to give me confidence sharing it as a point of entry. It did have a tendency to overstate implications and overgeneralise, but only slightly.
Reviewing Academia.edu’s podcast
Academia.edu took less than a minute to produce a 5 1/2 minute podcast overview. There weren’t a whole lot of options for customising the podcast, though there were some interesting multimedia options. But the process was quick and easy. So far, so good. In this case, I didn’t even need to enter a link, because my thesis is already there in my profile.
Unfortunately, it was all downhill from there. The Academia.edu podcast overview was laughably bad.
Firstly, the podcast didn’t even get the title of my thesis right. The podcast’s ‘host’ (an AI-generated version of the platform’s CEO Richard Price) dropped in a lovely compliment for my “fascinating” work as he introduced the thesis with in incorrect title… the term “Dehumanisation” became “Humanisation,” and “Posthumanism” was read as “Postism” (which is a different thing entirely). The correct title of my thesis is written right there in my academia.edu profile. It almost seemed as if the AI tool wasn’t able to read correctly?
As the podcast went on, it became even more clear that the technology is buggy. The ‘host’ delivered sentences that didn’t make grammatical sense. I found myself barely even concerned with inaccuracies in the content, because I couldn’t understand the oration in the first place. Prepositions, definite / indefinite articles, and conjunctions were missing from the sentences.
Here are a few choice quotes that make litte-to-no sense:
What does our fiction say about who gets count human?
New technology is opportunity to reflect on who are, and who we want to become. Until next time: curious.
And my personal favourite:
Irvine that so-called ‘enhancement assumption’ is a blind in both theoretical public.
‘Enhancement assumption’ is a term I use in the thesis, and at least the podcast said it correctly. But what is a ‘theoretical public’?
Then there was this work of surrealist nonsense:
About stories tell justify who’s in and who’s out. Y’know, remember sitting in a coffee once, evestro on a group debating robot citizenship, someone said: “we’ll just know they’t human.”
Technology won; just bring privilege.
Fans of The Office may be reminded of Kevin’s time-saving idiolect.
Worst of all, my keywords were badly garbled. A listener would not only misunderstand the terms I originated, but might also think that I misunderstood key terms related to my topic. My thesis makes reference to CRISPR, a common gene editing technology. The podcast relates this as CRPR. How hard is it to get a sequence of letters correct?
To add insult to injury, I couldn’t even download the podcast for free. That requires a premium account. Would I pay for the privilege of saving this garbled rubbish? No. No I would not.
Verdict: A listener relying on the Academia.edu podcast overview would know that my research had something to do with fiction, and with genetic engineering, but the specifics beyond that would be absolutely lost. They would get a far more accurate understanding of the thesis by reading the abstract, and that would take less time than listening to 5 1/2 minutes of confusing sentence fragments.
Should you use AI-generated research podcasts?
I’ve only reviewed AI-generated outputs from Academia.edu and NotebookLM in this post. There are many other tools that produce podcasts from academic writing (SciSpace, Wondercraft, SparkPod, and more). Of course there will be variation in the quality of outputs from one tool to another, and for different pieces of writing.
However, any AI-generated summary is going to have some quality issues. When you produce your own scholarly writing, you do so carefully, with precise language, through the lens of years of learning and experience. When an AI tool produces something based on a scholarly output, it is performing shuffling, summarising, and sometimes medium-shifting (e.g. text-to-audio) tasks without the benefit of the scholar’s human judgement.
I personally would not trust AI-generated research podcasts where the stakes are even moderately high. As a writer: I have invested far too much effort in the details and nuance of my own research to let a dubiously-ethical technology mangle it in the name of ‘impact’. As a reader: I respect my academic colleagues too much to prioritise my expedience over their expertise. Besides, if I build my knowledge based on inaccurate summaries, then I am only building inaccurate knowledge. That is the opposite of an education.
However, this experiment has shown me that AI-generated podcasts can be useful when you need only a general understanding of the material. (For the purposes of actually relying on or citing that information, I’d still always go to the source.)
That being said, the AI tool matters a great deal. Academia.edu butchered my thesis; whereas NotebookLM did a reasonable job of representing it.
Verdict: In my opinion, as of 2026, an AI-generated research podcast from a reputable tool can be useful for engaging with sources that are likely to be of ‘background’ level significance to your work. For anything seminal, critical, or to be directly cited? I’d just read the paper. A podcast summary is not good enough to rely on fully – not when your name, reputation, and career are on the line. Not when misinformation and disinformation is everywhere; and honest, valid scholarship is so desperately needed.
Notes
1For context: the thesis is all about how genetic engineering is depicted in books and films, and how representations of genetically engineered characters tend to show them as an exploited and animalised underclass in ways that mimic the dehumanisation of oppressed groups in recent history.
2 There are many other AI tools that can produce podcasts, but for the purposes of this experiment I stuck to the free ones. (In theory this would include SciSpace, but they limit you to a finite number of ‘free credits’ on your account, and mine were used up just thinking about the task of producing a podcast.)
