When AI comes to ruin medical research.
Cannabis one of a range of drugs that fall into a weird category. We know from a basic harm-reduction perspective that legalisation for recreational use is probably a good idea. Yes, there are some harms caused by cannabis, but those harms are significantly lower than the many and varied problems that criminalising the substance causes.
But as an illegal drug, cannabis has attracted a bizarre mythos. It’s not just a way to relax with some friends, or potentially help with chronic pain, it must be a cure for a wide range of conditions. You see people claim that various cannabis products can cure everything from depression to polycystic ovarian syndrome, and a whole range of problems in between.
Mostly, this is nonsense. There are a handful of conditions where cannabis appears useful, and even then the evidence is often contested.
So why are we talking about cannabis and cancer? Well, recently a new study was published that hit the headlines. It is one of the worst pieces of research I’ve ever seen, and may herald a whole new era of bad science. It looks like AI slop has come for medical research in a big way, and it’s not pretty.
Let’s look at the publication.
AI Slop
The new paper is called “Meta-analysis of medical cannabis outcomes and associations with cancer” and has been published in the journal Frontiers in Oncology. For those in the know, the journal is itself an immediate red flag – Frontiers as a publishing company has a very bad name in academia for accepting pretty much any paper as long as the authors are willing to pay the publication fee.
Another problem that pops up pretty quickly is that all of the authors have a direct financial conflict of interest. Three of the four-person team are employed by the Whole Health Oncology Institute, which is a for-profit organisation that appears to exist to promote cannabis as a treatment for cancer. The final member of the team is employed by the Chopra Foundation, which was set up by Deepak Chopra and promotes the use of various plants and herbal medicines.
There are other red flags that you can immediately see when you start reading the paper. The thing is over 10,000 words long, with another 5000 or so words in the supplementary appendix. Much of this is circular – the authors describe the benefits of their approach several times in almost identical terms, both in the paper and in the appendix.
It’s also quickly clear that large parts of the text have been written by AI. You can check this using online AI checkers, but it’s also fairly obvious just at a scan. The text is vague, hard to parse, and while well-written it’s often meaningless or just bizarrely off-topic. There is, for example, a five-paragraph section on sentiment analysis in the introduction that describes in general terms what this technique is and some of the issues with it without noting that it’s what this study was about nor why the authors used it. For some reason, the authors have also spent >1000 words defining common terms in the methods section and then just…done that again in the appendix. The authors also say at the very end of the study:
“This manuscript was analyzed by ChatGPT 4.0 to assess readability and consistency, any adjustments were directly implemented by authors. All final narrative, computational, and analytic content was derived from author’s contributions, not generative AI.”
So we know that the authors wrote the paper using ChatGPT. Their statement is vague enough to mean anything from simple grammatical checks to writing the entire thing with a few prompts. Given that large portions of the manuscript come up as 100% AI-generated, I would lean towards more GPT than human, but it’s hard to know for sure.
At a brief glance, the paper is problematic. But what did the authors actually do?
This is where the study gets even murkier. There is no simple, straightforward description in the publication of what the process entailed. Reading between the lines, they appear to have downloaded a large number of studies, fed those into ChatGPT, and then asked the LLM to run a sentiment analysis, but it’s almost impossible to know for sure based on what’s written in the document.
One of the main parts of the study is tracking keywords. The authors provide a list of keywords that are associated with supporting, not supporting, or unclearly supporting cannabis for cancer. The idea of a sentiment analysis is to identify how often these keywords appear in the text that you’ve got in order to get an idea of how positive/negative the general sentiment is about that thing. This is usually done with the aid of a complex statistical model that can count not just the occurrence of words but include some context as to how they are being used.
Here’s the AI word-salad that the authors use to describe this central, key aspect of their methods:
“The methodology for tracking the occurrence of keywords within individual studies in the systematic review dataset involves a detailed, quantitative approach to textual analysis. This method is instrumental in uncovering patterns and themes across a large volume of literature on medical cannabis, facilitating a nuanced understanding of the research landscape, and is detailed in Appendix A.”
Ok, so that tells us nothing. Let’s go to the Appendix:
“The methodology for tracking the occurrence of keywords within individual studies in the systematic review dataset involves a detailed, quantitative approach to textual analysis. This method is instrumental in uncovering patterns and themes across a large volume of literature on medical cannabis, facilitating a nuanced understanding of the research landscape. Here’s an outline of this process:
- Keyword Tracking: For each study in the dataset, the frequency of specified keywords related to medical cannabis is carefully recorded. These keywords are organized by topics, such as types of cannabis, therapeutic effects, specific medical conditions (e.g., various cancers), treatment-related terms (e.g., chemotherapy, immunotherapy), and outcomes (e.g., tumor growth, quality of life). The tracking process involves a systematic scan of each study’s text—often focusing on the abstract, results, and discussion sections—to tally the occurrences of each keyword.
- Topic Allocation: Keywords are categorized into broader topics to streamline the analysis. For example, “tumor increase,” “tumor growth,” and “tumor size” might fall under a “Tumor Growth” topic, while “generalized anxiety” and “chronic stress” might be grouped under the “Anxiety” topic. This categorization helps in organizing the data and facilitates a topic-wise analysis of keyword occurrences.
- Total Occurrences per Topic: The total number of keyword occurrences is calculated for each topic within each study. This aggregate figure provides insight into which aspects of medical cannabis are most frequently discussed or emphasized in the literature, highlighting areas of significant interest or concern within the research community.
- Categories, Topics, Keywords: The structure of categories, topics, and individual keywords is utilized to organize and analyze the vast amount of keywords captured in this meta-analysis.
4.1. Category: Categories are broad areas that encompass specific areas of interest within the field.
4.1.1. Topic: Each category is further divided into topics, which are more focused areas of study within the general category.
4.1.1.1. Keyword: These topics are then made up of individual keywords, which are the specific terms or phrases extracted from the studies themselves.”
That’s a lot of text, but you can pretty quickly see the issue if you read it. This is a general description of how sentiment analysis works – I can get a similar description by asking ChatGPT a variation on the phrase “how do you perform a sentiment analysis”?” – but it doesn’t tell us what the authors did in this study. How was the frequency of keywords recorded? How did they organize them into topics? What criteria did they use? What software? The study included >10k individual papers, so it must have taken a fair bit of computing power – how was this organised, run programmed, etc?
Related
The same issue is repeated for pretty much every aspect of the methods. The paper waxes at length about the importance of correlations, but there’s no information I can find on how the team ran their stats. The authors say:
“Correlations were calculated between all individual keywords within a topic and the sentiments of supported, not supported, and unclear, for both keyword occurrences and dominant instances.”
Elsewhere, they say that they used Pearson’s r. But this calculation requires a numeric estimate. So was this a correlation between the numbers of keywords in a study and some numeric estimate of “supported”? I’ve read through the methods, results and supplementary appendix several times and I still have no idea of what went into the equation that produced the numbers the study presents.
On top of all of that, the entire paper is a ridiculous waste of time from a scientific perspective. The main thing that the authors do describe in full is how they got their sample of papers. They entered a bunch of keywords into Pubmed, a big database of medical research, and extracted 10,641 individual papers.
The problem is that the search terms are far too broad, and many of these papers meaningless to the discussion of cannabis and cancer. I copy+pasted one string of the author’s keywords into Pubmed, and here’s a selection of studies from the first page of results:
- The impact of cannabis exposure for male rats on their children’s dopamine level.
- The impact of cannabinoid exposure on rat brain cells to identify possible neuropsychological impacts of the drug.
- An investigation into the use of cadmium sulfide in cancer cells (yes, nothing to do with cannabis).
- A study on how cannabis smoke may degrade supplements (i.e. if you keep your bong too close to your multivitamin).
The only thing that the authors report doing to narrow down their massive dataset of studies was removing “Duplicates/Text Unavailable/Unreadable”, which means there was no qualitative screening for whether a study was completely off-topic. So all of these studies which have literally nothing to do with the impact of cannabis and cancer are in theory included in the sentiment analysis.
Even if we ignore this major, glaring issue, what does a sentiment analysis even tell us here? Do we really care if this 1985 paper on a small sample of rats that looked at tumour development after cannabis smoke exposure had a supportive or unsupportive sentiment?
There are many other major holes in the methodology, but honestly this piece is already far too long. Suffice to say that it’s one of the worst things I’ve ever seen published.
Bottom Line
I’m not going to go into the deep, worrying aspects of this paper appearing in a medical journal. Yes, it’s Frontiers, but sooner or later we will probably see something like this in a less disreputable place. The journalists who covered the paper definitely should have looked a bit harder, but you can easily imagine slightly less conflicted academics doing something similar just to get a publication. I would not be at all surprised to see dozens of stories in the future reporting on AI slop, it’s not something that’s going to go away any time soon.
From a scientific perspective, there’s no more reason to feel optimistic about cannabis than the last time I wrote about it. There is some evidence that it helps for specific things, like nausea induced by chemotherapy, but little data showing benefits for the vast majority of stuff. It’s like any other drug – probably useful for a few things, probably not useful for the rest.
This paper tells us literally nothing new about the situation. At best, it tells us that people who write papers that come up in searches about cannabis and cancer use more of the keywords the authors associated with supporting cannabis. That doesn’t mean that the studies are positive, or that there is broad scientific support, just that a list of quite arbitrary keywords – the term “prospective” was considered “supporting” in this study – is more common than another list of arbitrary keywords.
I’m still pro-legalisation when it comes to cannabis. From a public health perspective, it makes a lot of sense. But AI slop like this paper adds nothing to the discussion.
Dr Gideon Meyerowitz-Katz is a Sydney epidemiologist, writer and senior research fellow at University of Wollongong who is known online as Health Nerd. This was originally published on his Substack.
