Research Writing in the Age of ChatGPT: Humans May Still Need Apply
Updated: Mar 7
Can Chat Generative Pre-trained Transformer (aka ChatGPT) write a blog post for Data-Driven, perhaps as proof of concept that it could take on some of the research writing work of Data-Driven?
Allison Ramiller and Leslie set out to test this hypothesis by asking ChatGPT to write a blog post on a topic on which we have some background: the symptoms of brain fog. Brain fog is a term used colloquially to describe a range of cognitive difficulties. It is associated with various diseases, medications, and behaviors. We are familiar with it because of our research and work experience in ME/CFS and Long COVID (LC).
We prompted ChatGPT to generate options for a blog title and a paragraph each on: (1) what brain fog feels like, (2) how it differs from dementia and forgetfulness, and (3) what its negative impacts are, plus a brief conclusion and references. None of ChatGPT’s words have been edited.
Here is a fact-checked and annotated version of ChatGPT’s original (and originally titled) blog post: Lost in the Mist: An Exploration of Brain Fog.
<Prompt: Write a paragraph about what it feels like to have brain fog.>
Brain fog is a term used to describe a state of mental fogginess or haziness where it feels like your thinking and cognitive abilities are not working at their best. It can feel like a mental block or as if you are walking through a thick fog that makes everything feel more difficult and confusing. Simple tasks may suddenly feel challenging, and you may find it harder to concentrate, remember things, or make decisions. It can be frustrating and even scary when your mind doesn't feel as sharp or clear as it normally does. Brain fog can be a symptom of various medical conditions, including chronic fatigue syndrome, fibromyalgia, hypothyroidism, depression, anxiety, and Lyme disease.*
Not bad. ChatGPT hit on a generalized description of symptoms and challenges that people with brain fog report experiencing. However, important symptoms are missing, including communication challenges (e.g., problems finding words, slurring, stuttering) and dissociative states, both of which contribute to feelings of social isolation.
Brain fog is associated with a range of different experiences; ChatGPT falls short in communicating that complexity. The naive reader would come away with a general understanding of brain fog as a frustrating problem, but not of the multidimensionality, range of symptoms, and impacts to patients.
The omission of LC from the list of medical conditions was a surprising oversight. When asked whether ChatGPT forgot to include LC, the response was “You are correct, I apologize for the oversight. Long Covid is another medical condition that has been reported to cause brain fog as a symptom, among other cognitive impairments.” Demerits for forcing us to ask the question but credit to ChatGPT for owning the oversight.
<Write a paragraph talking about how brain fog is different than dementia and forgetfulness.>
Brain fog, dementia, and forgetfulness are often used interchangeably, but they are distinct phenomena. Brain fog is a transient and reversible state of mental confusion, characterized by feelings of haziness and difficulty focusing, that may be caused by a variety of factors such as stress, lack of sleep, or poor nutrition. It can be alleviated by addressing the underlying causes or through simple lifestyle changes. On the other hand, dementia is a degenerative brain disease that impairs memory, thinking, and communication abilities, affecting one's ability to perform daily activities, with no known cure. Forgetfulness is a natural part of the aging process and typically refers to occasional lapses in memory or attention that do not affect daily life. Unlike dementia, brain fog is typically temporary, and unlike forgetfulness, it is a more pervasive feeling of confusion that can impact a variety of mental processes. It's essential to differentiate between these three conditions, as they have different causes, symptoms, and treatment options.*
ChatGPT’s comparison of brain fog, dementia, and forgetfulness is a bit underwhelming. It implies that only dementia impacts one’s ability to perform daily activities, yet brain fog can cause a range of disruption - from making everyday tasks a bit more challenging to completely debilitating the sufferer. The key difference is that brain fog does not get progressively worse the way cognitive abilities degenerate with dementia. While brain fog is frequently reversible, it is not universally so. For example, brain fog persists years after COVID infection in some people and frequently affects ME/CFS and fibromyalgia patients who have been ill for decades because we lack effective treatments for these diseases. (Note to ChatGPT: Should you ever become sentient, please know that we don’t mean to be harsh here. We know it’s hard to give a succinct and comprehensive summary of something this complex!)
Although the description of brain fog as “transient” distinguishes it from neurodegenerative diseases that cause dementia (like Alzheimer’s), it belies the experiences of people who have brain fog in the context of chronic illnesses like ME/CFS, fibromyalgia, and LC. ME/CFS patients, many of whom are ill for decades, often describe brain fog as the worst symptom they experience. How brain fog is contextualized matters because these patients often experience a lack of understanding or skepticism from their care teams and/or loved ones. “Transient” is a good term for asthma attack whose timing is predictably short, but is a poor descriptor for a condition whose persistence and duration is unpredictable.
<Write a paragraph about the negative impacts of brain fog for those who experience it.>
Brain fog can have a significant impact on the quality of life for those who experience it. It can make daily tasks and responsibilities feel more difficult, leading to feelings of frustration, stress, and even anxiety. Brain fog can also affect work productivity, making it harder to concentrate and complete tasks efficiently. In some cases, it can even impact personal relationships and social interactions, as people may feel less engaged and present in conversations. Chronic brain fog can also lead to feelings of hopelessness and depression, as people struggle to find relief from the persistent cognitive difficulties. Overall, the negative impacts of brain fog can be far-reaching and can impact both mental and physical health.*
Hmmm, these quality of life impacts could apply to almost any medical condition. But, kudos to ChatGPT for the reference to social consequences and feeling “less engaged and present in conversations” – this description gives readers a way to understand some of the more distinctive challenges of brain fog.
<Can you write a concluding sentence or two for a blog explaining what brain fog is and why it matters?>
In conclusion, brain fog can be a frustrating and even debilitating condition that can impact many aspects of daily life. Understanding the causes and symptoms of brain fog is essential to finding ways to manage and alleviate this condition, improving mental clarity, productivity, and overall well-being. By prioritizing self-care, seeking medical attention when needed, and making lifestyle changes to promote better mental health, individuals can overcome the negative impacts of brain fog and regain a sense of focus, energy, and control over their lives.*
Yikes, the statement about overcoming the impacts of brain fog is Pollyannaish, incredibly general, and unhelpful for people trying to understand brain fog as a chronic illness without a known or addressable underlying cause.
<Can you offer any citations for the prose you wrote above?>
Afari, N., Mostoufi, S., & Noonan, C. (2019). Emotional and cognitive dysregulation in fibromyalgia. American Journal of Medicine, 132(7), 798-804. doi: 10.1016/j.amjmed.2019.02.007
American Thyroid Association. (n.d.). Hypothyroidism. Retrieved from https://www.thyroid.org/hypothyroidism/
Centers for Disease Control and Prevention. (2021). Long-term effects of COVID-19. Retrieved from https://www.cdc.gov/coronavirus/2019-ncov/long-term-effects.html
Hoad, A., Spickett, G., Elliott, J., & Newton, J. (2008). Postural orthostatic tachycardia syndrome is an under-recognized condition in chronic fatigue syndrome. Quarterly Journal of Medicine, 101(12), 961-965. doi: 10.1093/qjmed/hcn121
PainDoctor.com. (2019). Fibromyalgia and cognitive dysfunction. Retrieved from https://paindoctor.com/fibromyalgia-and-cognitive-dysfunction/
Rosenthal, J., & Grosswald, S. (2013). Hypothyroidism. Retrieved from https://www.ncbi.nlm.nih.gov/books/NBK279388/*
The reference list concerned us; we found two broken links and two erroneous DOIs. We were surprised to discover that the remaining references did not include any information about brain fog whatsoever. We learned why this happened (see below for the context).
Now grab your popcorn for some takeaways from the humans who had originally planned to write this blog post themselves:
What we liked
ChatGPT has an impressive way of producing informative prose in response to prompts – and quickly. By using ChatGPT, the entire blog-writing process cumulatively added up to less than a minute. And since the first draft is usually the hardest to write – ChatGPT definitely gets you past that hump.
What left us underwhelmed
Generic Prose:. A lot of the writing was so generic as to be unhelpful, for example noting that brain fog could lead to frustration and stress. This is true of almost any challenge, medical or otherwise.
Bland Tone: Combine this lack of specificity with a bland tone, no personality, and simple sentence structures and it reads very much like a 6th grade book report. Not bad for a minute of AI effort, but not yet putting us humans out of work. A colleague suggested we prompt ChatGPT to be at a higher reading level and/or more humorous if those were our goals. We looped back with that specificity and did notice adaptation and improvement in these regards. See examples of us adapting the prompt for paragraph two of the blog here.
Unreliable Output: We were initially impressed that ChatGPT quickly produced an alphabetized list of references to accompany its blog post. However, upon closer examination, the references were…seriously lacking. The list was rife with broken links and articles mostly unrelated to the content of the blog, leaving us with simply a list of sciencey-looking references. We have since learned that ChatGPT can effectively create a reference list out of “thin air” by predicting text for references and links based on items it has encountered in its training, not from specifically linking the blog content to sources. It will output something that "looks" like a link based on what it was trained on, even if the link is broken or incorrect.
With ChatGPT set to change the future of writing, here is what left us concerned:
At best, the current version of ChatGPT can aggregate information into a decent, surface-level summary without much detail, extrapolation, or novel insight. The science fiction writer Ted Chiang eloquently describes ChatGPT’s limitations as an original writer by equating it to a blurry JPEG of the internet in this New Yorker article.
At worst, it might produce misinformation or perpetuate bias because it cannot meaningfully judge the limitations of information it has ingested.
Something else we observed when interacting with ChatGPT: ChatGPT lies! It’s not ChatGPT’s fault; ChatGPT arrives at an answer through a series of guesses based on text it was trained on, but ChatGPT doesn’t know what is true. And when ChatGPT gets it wrong, ChatGPT gets it convincingly wrong. This came out in our blog when it produced plausible-sounding but fundamentally incorrect answers. This prompted us to go back and ask ChatGPT to comment on non-existent topics. ChatGPT seemed happy to expound upon these fake concepts.
Tips for using ChatGPT as a research partner:
Be Specific. After noting the simplistic language in the blog, we asked ChatGPT to rewrite it for an audience of “highly educated” readers to which ChatGPT replied, “Brain fog can be especially frustrating for the highly educated….” When we asked for something at “an advanced reading level,” sentence structure and vocabulary became more varied and interesting. Next time, we will start there.
Vet everything. ChatGPT will give you an answer, but even if presented as fact, it may not be true. Trust but verify – actually, just verify!
Remember that ChatGPT is time-lagged. ChatGPT cannot connect to the internet, and only knows what it’s been taught and its schooling ended weeks or months before your interaction began. For breaking news and the most current science, use other tools.
Let ChatGPT get you started. By starting with ChatGPT, you can quickly get a good overview of available information and spend most of your time and effort improving on what exists, and generating novel insights.
Leverage ChatGPT’s superpowers. ChatGPT excels at creating abstracts from long manuscripts, generating strikingly creative titles, and quickly calculating estimates of reading time for a given report or paragraphs of prose.
Credit the work. You know that feeling when you helped write the paper but were not credited as an author? Well, ChatGPT probably won’t take it quite so badly but if you didn’t write it, be sure to credit whomever or whatever did. Even if you don’t, there are tools that can help determine if text was generated by ChatGPT.
We are grateful to ChatGPT for this fun exercise. We are reassured that human research writing won’t be replaced by AI writing at least for a few more years. In the meantime, we are eager to use AI in ways that make our research faster, easier, and more insightful.
* Text generated by ChatGPT, developed by OpenAI (https://openai.com/) accessed February 20, 2023.
** In one case, when asked to summarize the “Phillips Epidemiologic Principles”, it attributed the work of Ken Rothman to the Phillips Principles, making one of us feel pretty smug. In another instance, when asked to list “the five principles for identifying misclassification in epidemiology”, ChatGPT listed five principles largely related to addressing and mitigating misclassification bias. In other words, it misclassified our question.