All writing is structured, but scientists have an easier time in that scientific writing follows a very predictable structure. This is evident not only at the macro level (Introduction, Methods, Results, Discussion) but also at the micro level. Consider the following excerpt from the Results section of a scientific paper:
To confirm that the Hebbian term is essential for latent predictive learning (LPL) to prevent representation collapse, we simulated learning without the Hebbian term (compare equation (1)). As expected, we observed that neuronal activity collapsed to zero firing rate (Fig. 2c). Conversely, when learning with the Hebbian term but without the predictive term, no collapse occurred. Thus, the Hebbian component of LPL is essential to prevent activity collapse.
Even if you are not a science major, the underlying structure is clear.
reason: “To confirm that the Hebbian term is essential for preventing Latent Predictive Learning (LPL) from collapsing representations…”
approach: “…we simulated learning without the Hebbian terms…”
result: “We observed that the neuron's activity dropped to zero firing rate, as expected (Figure 2c). Conversely, when we trained with the Hebbian term but without the predictive term, no drop occurred.”
interpretation: “Thus, the Hebbian component of LPL is essential to prevent the collapse of activity.”
This “rationale, approach, results, interpretation” structure is a common way scientists present their research results. There's nothing written in stone that says results have to be reported that way, but it's the standard because it has a clear narrative flow. The rationale sets the stage by explaining the experimenter's motivation. The approach and results document what happened and how, and form the main plot. Finally, the interpretation provides the author's view of what it all means — the moral of the story.
Artificial intelligence tools like ChatGPT can take advantage of the highly organized nature of scientific writing to streamline the writing process. Take a typical results paragraph for example: As you already know from reading this column, you would need to explain the rationale, approach, results, and interpretation of a particular finding. Roughly write out this information without worrying about consistency, grammar, or style. Or, you can just speak it into your device. Then, feed the draft (or a transcription) into a tool like ChatGPT as a template along with the underlying structure (rationale, approach, results, interpretation). This workflow allows you to painlessly produce a draft that you will then need to check for accuracy and edit further.
To see an AI-assisted speech-to-paragraph workflow in action, let's look at an example from one of my papers. That way, we'll be able to determine, at least in principle, whether ChatGPT introduces inaccuracies, exaggerations, or other issues. After all, the paper is almost 10 years old. (As a technical note, I'm using the OpenAI web interface with GPT-4. To keep it reproducible, I'm not using custom instructions. I start a new chat for each prompt to avoid chat history effects.)
prompt:
You are writing the Results section of a scientific paper. Your task is to summarize the following audio transcript into a coherent paragraph. The information about the scientific experiment described should be written in the following order: Rationale, Approach, Results, Interpretation. Your answer should retain all necessary details from the original transcript, such as numbers and specific names. If the transcript mentions numbers or values but does not give names, insert a placeholder: “[INSERT VALUE HERE]Citations for figures should be placed in parentheses. References should be placed in parentheses at the end of the sentence. Keep the tone formal, but prioritize clarity and brevity.
Here's the transcript: “So, where do we start? Okay, there's been some work done on this. Mossy fibers go into the posterior supragranular cortex, or EGp for short. There are granule cells that connect to the ELL. See Figure 1B. Mossy fibers arise from different parts of the brain, even the spinal cord. Please cite the relevant papers… I think it was Bell 1981, then Szabo 1979 and 1990. These mossy fibers carry different kinds of information. There's EOCD, there's proprioception, there's electrosensory signals. I think Bell talked about this in his 1992 paper, but you should also cite the Kennedy paper from 2014. Oh, and you should also cite our paper, Sawtell from 2010. You need to remember those. So our main question is, are some of these mossy fibers carrying motor-related information, i.e. concomitant discharges? To figure that out, we're going to look at a hypothetical preparation, and then we're going to look at the EGp. We combined extracellular recordings from putative mossy fiber axons in the 10-kDa brainstem. Bell et al. mention this method. We found that a group of tonically active mossy fibers showed firing rate changes during spontaneous movement. This is shown in Figure S1. Or when there was microstimulation-evoked motor activity. About 23 of 48 fibers showed this change. This is shown in Figure 2C. Now, for the fibers that showed rhythmic motor activity. That was 19 of 23. The rhythmic firing rate changes coincided with the motor neuron activity. Here are the correlation values. Make sure you enter them. And there was a clear association between the frequency of mossy fiber firing rate changes and the frequency of motor neuron bursts after smoothing. This is shown in Figure 2D. Some fibers showed this during spontaneous bursts as well. Now, finally, in some of the mossy fibers that we tested with different microstimulation intensities, the frequency of firing rate changes increased with microstimulation intensity. Enter your numbers here. They are in Figure 2E. “We therefore conclude from this data that mossy fibres appear to transmit motor information related to the rhythm of swimming movements.”
ChatGPT output:
Previous studies have revealed that mossy fibers project to the posterior supragranular cortex (EGp), where granule cells establish connections with the ELL (Figure 1B). Originating from different regions of the brain and even the spinal cord, these fibers convey different types of information, including EOCD, proprioception, and electrosensory signals (Bell 1981; Szabo 1979, 1990; Bell 1992; Kennedy 2014; Sawtell 2010). A major question under investigation was whether certain mossy fibers relay movement-related information, specifically, corollary discharges. To address this, we used a combination of hypothetical preparations and extracellular recordings from putative mossy fiber axons in the EGp, a method previously mentioned by Bell et al. Our observations revealed that a subset of tonically active mossy fibers exhibit changes in firing rate during spontaneous movement (Figure S1) or in response to microstimulation-evoked motor activity. Notably, 23 of 48 fibers exhibited this behavior (Fig. 2C). Of these, 19 of the 23 fibers showing rhythmic motor activity showed rhythmic firing rate changes synchronized with motor neuron activity at specific correlation values. [INSERT VALUE HERE]We found a direct correlation between the frequency of mossy fiber firing rate changes and the frequency of smoothed motor neuron bursts (Fig. 2D). Some fibers showed this pattern even during spontaneous bursts. When exposed to various microstimulation intensities, certain mossy fibers showed that the frequency of their firing rate changes increased in conjunction with the microstimulation intensity. [INSERT VALUE HERE] In conclusion, our data strongly suggest that mossy fibres transmit locomotor-related information that is tightly coupled to the rhythm of swimming movements.
ChatGPT Link: https://chat.openai.com/share/e88bd31e-ca63-4d7d-b356-4bbbc6c01443
The generated paragraphs may not be in a state suitable for publication (the style is uneven, sometimes terribly flat or overly decorative, and some words can be removed), but the basic structure is in place and it's a good state to start revising. You might argue that you can achieve the same results in less time without using ChatGPT, and that's probably true. However, in my experience, I've found that using AI in this way helps writers avoid two common problems with first drafts: Some writers (including me) get bogged down in a blank page. For this group, speaking their thoughts out loud and using ChatGPT's templates to create a workable draft is a great psychological relief and a great way to overcome procrastination. Other writers tend to jump too far into a draft without a proper structure plan and then struggle to fix the resulting mess of words. Having ChatGPT's templates ready forces these writers to think about the structure first, making subsequent revisions easier.
Prime prompt: By providing the AI with details about the context, format, tone, and template required for a writing task, it will output a more usable draft.
Indeed, the resulting paragraph is, in some ways, the simplest scientific writing we have seen. But using this strategy in its simple form helps users develop good practice. This is especially important when using AI tools to assist with more complex writing. Think of scientific writing as a cognitive act that requires ongoing decisions at multiple levels: content (central ideas and their connections), structure (how to structure words, sentences, and paragraphs for clarity and impact), and style (word choices that balance personal preferences and expert expectations). These three levels are not completely separable, but it is important to think in advance at which levels you want a tool like ChatGPT to intervene.
In the results paragraph, the content is set pretty strictly by the findings, the structure is mostly predetermined by scientific conventions, and it remains primarily stylistic. Tools like ChatGPT may not produce a stylistically perfect text, but they can be further refined to your specifications. And because we have prepared clear templates and organized thoughts, we can ensure that the AI is not limited to the tedious work of drafting and does not overly affect the creative process. Certainly, you may want to leverage AI for structural or conceptual assistance, but that may require a different approach, which we will discuss in future columns. The key is to recognize and respect these differences.
