Using ChatGPT as my "Lab Assistant"
My first attempt at using ChatGPT in a collaborative way. I explore its use as a writing partner and lab assistant by playing with an experiment I have been running in my lab for the past five years.
DiPartures is an opportunity for me to publish the projects which are outside of the scope of my work in photography and style in “east vancouver”. Lately, it has been focused on some of my more “academic” adventures, and I hope to continue to publish some of the other experiments and reflections outside of photography as this section evolves. If you’d like to continue seeing these posts, please consider subscribing!
Pre-Introduction
This “experiment” is a two fold activity for me. Primarily, it is to explore the viability for ChatGPT to assist in writing a formal lab report at the high school level. In doing so, each lab section will be opened with a short explanation of how I used ChatGPT to assist in the lab writing. The second aspect of this lab is to explore the CO2 sensors I have students use in my lab at the beginning of each school year. I have had this research question pertaining to the CO2 sensors on my mind for a couple of years and I have been wanting to write my own lab report, but have never had the time to achieve it. Thanks to ChatGPT, I was able to conduct the lab and write it within a couple of hours.
For the past couple of years, we (myself and the teachers in the science department and our school in general) have debated how we will approach the use of generative AI, how we can use it as educators, and how it will affect our assessments especially for lab assignments. Rather than looking at it as something that aids in “cheating” or “plagiarism”, I aimed to use it as a writing assistant. My main idea was “What if I could just tell ChatGPT the things I was doing, almost using it as a journal, for it to then create my lab sections which require lots of writing for me?”. Rather than giving simple prompts such as “write my introduction to a lab…” I will focus my inputs in ways that are specifically tailored to what I am trying to achieve in the lab. This is something that has reportedly been happening at all levels of science writing over the last couple of years, with specific tools designed for main sections of lab reports, such as protocol generation.
At our school, we have created an “acceptable use scale” which can help the staff and students understand what level AI use is acceptable on any given task. For each section, I will briefly mention which level I believe the section which follows is, and explain why.
There have been many conversations in education today talking about the usefulness, limitations, ethics, and environmental impact of using ChatGPT for example, but like any new technology, I propose it is only as good as the creativity and ingenuity of its user.
The full transcript along with my prompts can be found here.
Writing that is fully written by ChatGPT will be cited as such, other aspects will be harder to cite as I will explain further below.
Abstract
This experiment aimed to determine the minimum amount of yeast required to produce detectable CO₂ levels using Vernier Go Direct CO₂ sensors. A series of trials were conducted with varying concentrations of Fleischmann’s Quick Rise yeast (ranging from 1.000g to 0.007g) in a sugar-water solution to measure CO₂ production during fermentation. Baseline trials confirmed that the probes were stable, showing minimal fluctuation without yeast. The results demonstrated a clear correlation between yeast concentration and CO₂ production, with even the smallest amount of yeast (0.007g) generating a measurable increase in CO₂ levels. These findings indicate that the CO₂ sensors are highly sensitive and can detect fermentation at micro-scale levels, providing valuable insights for both educational and industrial applications where precise control over yeast concentrations is important (OpenAI, 2024).
Using AI for an Introduction to a Lab
The first place I hypothesized that ChatGPT would be useful for writing lab reports was to put together an introduction. For me, an introduction needs to hook the reader into caring about the results of the lab, as well as give any background information the reader could potentially read.
For the prompts here, I mainly focused on telling ChatGPT about what my goals were, some of the variables I would be manipulating and the history of the lab activity in general. It originally gave me a very wordy introduction, so after a couple prompts and a bit of editing, this is the introduction I ended up with. The writing below is about 80% ChatGPT and 20% my additions.
Once I had the introduction set up, I went through the rest of the lab, working with ChatGPT to write sections, and then finally came back to the introduction section and got ChatGPT to do a final clean-up, as you can see in my transcript with the AI.
For the introduction, it started out as Level 3, as I went back and forth with the chatbot to describe the experiment. It is my work, just summarized by the bot. In the end though, I took my final mash-up of ChatGPTs work and my work, and received a summarized and sharpened version, which I believe then elevates this section to Level 4.
Introduction
In AP Chemistry, hands-on experiments are an effective way to connect classroom concepts to everyday life. Over twelve years of teaching, kinetics experiments have consistently been a favorite, particularly those involving the catalyzed reaction of sugar and yeast due to its common uses in baking and brewing. These relatable topics help students develop their ability to ask research questions, run experiments, and analyze results, culminating in full lab reports.
The yeast and sugar experiment is especially relevant, as students often explore questions such as "Which type of yeast produces more CO₂?" or "Does yeast have a preferred sugar?" These investigations, using Vernier Go Direct CO₂ sensors, allow students to track CO₂ production as an indicator of fermentation and connect their findings to the kinetics of reaction rates. Over the years, results have shown that different yeast strains and sugar types can indeed affect the rate of CO₂ production.
However, students often start by following package instructions, using large amounts of yeast and sugar, which leads to an overproduction of CO₂ and overwhelming the sensor. This prompted my investigation to determine the minimum amount of yeast required for the sensor to detect CO₂ production when plenty of sugar is present.
The Vernier CO₂ sensor, with a resolution of 1 ppm and a range of 0-10,000 ppm, is sensitive enough to detect even small amounts of CO₂. Therefore, this experiment will test whether a fractional amount of yeast can still generate detectable CO₂. The Fleischmann’s yeast package instructions recommend using about 11g of yeast with a teaspoon of sugar for baking, but this is far too much for small-scale experiments. The goal here is to identify the minimum yeast amount that produces detectable CO₂ levels.
This experiment will measure CO₂ production in a simple sugar solution with varying amounts of yeast. A baseline trial will account for any background noise in the sensor readings, and subsequent trials will determine the smallest amount of yeast detectable using consistent controls. By doing so, the experiment will assess both the sensitivity of the CO₂ sensor and the minimal yeast quantity required to initiate fermentation.
Using AI for the Methods Section
I believe this section is where ChatGPT became most useful. The way I approached using it for the methods section, was to describe the steps I was taking to run the experiment, being clear about my variables and actions. I instructed ChatGPT to “store” that information, and eventually it knew I wanted it to write me a methods section. This mimics very closely what a lab workbook would look like, storing my rough notes and thoughts, to eventually go back and write formally.
In this case though, ChatGPT was able to summarize my entire method section in a fraction of the time. Keeping me organized. It was impressive in the way it was able to organize, chronologically, my procedure.
Originally, it produced a “step-by-step” methods section, which I don’t love (personal preference), so I asked for a narrative style and it came out almost perfect.
For this section it is much harder for me to decided which level I believe this section is. At the end of the day, it is all my work, just summarized by ChatGPT. It did not offer any thoughts or insights, it just summarized what I did. That being said, if I go by the chart above, it would fall into the Level 3, as it did produce the sentences that I would eventually edit and paste below.
One note: all graphs and tables produced below are my own.
Methods
To investigate the minimum detectable amount of yeast required for CO₂ production, a series of trials were conducted using Vernier Go Direct CO₂ probes. The experiments began by preparing three probes (numbers 13, 15, and 16), which were washed with tap water and calibrated following the manufacturer’s guidelines. Tap water was used as it would be the most common water used in any baking application, as since this experiment is in the realm of baking, it was decided that tap water would be sufficient. Calibration was performed indoors, and the probes were placed in the Vernier standardized bottles which accompany the probes. Each trial contained 75 mL of 36°C tap water with 4.5 grams of pre-dissolved sugar, optimal living temperature for the yeast, and an abundant amount of sugar for the reaction. These initial trials aimed to assess any background CO₂ levels or "noise" in the system before introducing yeast. The probes were observed for a few minutes to notice any fluctuations in data and then allowed to adjust for 5 minutes before data collection began. The yeast data was intended to be then collected over a 20-minute period. The baseline trials revealed minimal fluctuations in CO₂ concentration, confirming a stable setup (See Graph 1 in Data Section). The probes stayed within an acceptable range over a 10 minute period.
For the first set of yeast trials, varying amounts of Fleischmann’s Quick Rise yeast were introduced to the sugar solution. The yeast was pre-dissolved in 10 mL of 36°C water before being added to the containers. Probe 13 (red on graph) received 1.000g of yeast, probe 15 (yellow on graph) received 0.500g, and probe 16 (blue on graph) received 0.050g. After adding the yeast, the mixtures were gently stirred and left to stabilize for one minute before starting data collection. Stirring was repeated every 5 minutes over the course of an intended 25-minute trial, during which CO₂ levels were closely monitored. Notably, one probe (13) initially registered higher readings after adding the water, but this discrepancy resolved after a brief adjustment period. The trial was ended after 900s (15 minutes) as the results were clear that the probes were all registering a change in CO2 concentration much above that of the baseline levels when no yeast was involved.
Encouraged by the results from the initial trials, a second set of experiments was conducted to explore smaller increments of yeast. The same sugar-water mixture (4.5 grams of sugar in 75 mL of water) was used, with yeast pre-dissolved in 10 mL of 36°C water. Probe 16 received 0.007g of yeast, probe 15 was assigned 0.012g, and probe 13 received 0.044g. These masses were chosen as they were visibly different when putting some yeast grains on the weigh boats. Any less either seemed too miniscule, or did not register sufficiently on the scales.
After allowing the probes to stabilize again for 5 minutes then the probes were recalibrated to 400 ppm after a 15-minute period with just water and sugar, ensuring accuracy before the trials with yeast began. The second yeast trials ran for 25 minutes, with readings taken every 5 minutes to ensure consistency across the experiments.
Data Section
Graph 1: Baseline Data for the three probes
Graph 2: 1.00 g Yeast (Blue), 0.500g Yeast (Yellow), 0.05g Yeast (Red)
Graph 3: 0.04 g Yeast (Red), 0.012g Yeast (Blue), 0.007g yeast (Yellow)
Analysis
I did not use ChatGPT for this section as I figured it would be more work to enter in all my analysis thoughts for it to then consolidate it. If I had a more complex research question that required more data analysis, it could have been useful to run it through some of the specific Chatbots on ChatGPT.
A1. Baseline Trials
The purpose of the baseline trials was to analyze the “noise” of these probes. Once I ran the baseline for a while, I could see that the probes do not fluctuate outside of 10 ppm of their starting point. Even after removing the probes and changing the conditions between adding new ingredients, the probes do not fluctuate much. This was an important aspect as in previous years some of the probes would “go rouge” and continue to steadily increase despite no reaction actually happening.
A2: 0.05g - 1.00g
It is clear by the graph above that the probes work quite well in differentiated amounts of yeast. It would be assumed that the more yeast present, the more CO2 that could be produced with the same amount of sugar in a set amount of time. The graph shows that the 1.00 g trial produced much more, and much faster, than the 0.500 g and 0.05 g trials. Although the scale for the 1.00 and 0.500 g trials throw off the 0.05 g trial, a change was still seen in concentration, so further trials needed to be done to see “how little” could really be detected.
A3: < 0.04g Trials
Again, it can clearly be seen that all three trials produced some amount of CO2, even the 0.007g of yeast produced a 60 ppm increase (~10%) over a 25 minute period. Again, the validity of the experiment can be seen by the larger amounts of yeast producing more CO2 than the minimal amounts of yeast.
Using AI for the experiment Discussion
For this section, I copied and pasted everything I had written above, including the data and analysis, to ask ChatGPT what useful subheadings for a “discussion section” would be. It ended up producing a very good summary and some highlights of my lab which I will include here. This would fall under Level 4, as it is ChatGPT which summarized the lab and explained some of the strengths of the lab, with me slightly editing some of the sections. I wrote the introduction to the section, but the headings and the content under are mostly ChatGPT.
Discussion
The results for the lab were conclusive in understanding the minimum amount of yeast that could be detected by the CO2 sensors. I believe the senors could even detect smaller amounts. The 0.007g trial had a “countable” amount of yeast on it, probably between 10-30 grains of dried yeast. If the scale was even more sensitive, I believe it could have detected CO2 generated by a few grains of yeast.
Although there was not a fully fleshed out “hypothesis”, there are three main highlights I believe this lab archives:
1. Effectiveness of Probes and Experimental Setup
The CO₂ probes demonstrated consistent stability, with fluctuations within 10 ppm during baseline trials, reinforcing their reliability for detecting small-scale CO₂ production. Recalibrating the probes to 400 ppm after the initial baseline was crucial in ensuring accurate readings, particularly for the trials involving minimal yeast concentrations. This stability is essential when working with small amounts of yeast, as even slight probe instability could obscure meaningful data, which makes these sensors well-suited for this type of experiment (OpenAI, 2024).
2. Yeast and CO₂ Production Relationship
The experiment showed a clear relationship between yeast concentration and CO₂ production, with detectable fermentation from as little as 0.007g of yeast. This highlights both the sensitivity of the CO₂ sensors and the efficiency of yeast in metabolizing sugar, even in very small quantities. These findings could have broader implications for optimizing fermentation processes in various industries, where controlling yeast efficiency at different concentrations can lead to more consistent results in baking, brewing, or other fermentation-driven applications (OpenAI, 2024).
3. Comparison of Different Yeast Concentrations
CO₂ production increased proportionally with yeast concentration, as expected, with higher yeast amounts producing faster and more significant CO₂ levels. The results confirm that yeast concentration directly influences fermentation rates, which can be crucial for adjusting yeast amounts in different practical settings. However, while even very small yeast quantities were detectable, there may be practical limitations in real-world applications, where controlling such small amounts could be challenging or impractical outside of laboratory conditions (OpenAI, 2024).
For the Sources of Error, I believe that I had standardized my methodology well enough that I would have avoided any critical impingements that would have nullified my results. If I had to then use the data to analyze a certain RATE or expected value per unit of yeast, I would have included a stir bar and ensured the yeast and sugar was being interacted with sufficiently rather than my occasional stirring. As the purpose of this lab was to just see if there was a “detectable” amount, the rate of production, as well as the values were less important, however, the magnitudes collected allowed me to still think that the results were valid. To increase the validity of the trials, more trials, and repeatable trials could have been used, but from experiment to experiment, the expected results (more yeast = more CO2) were conclusive, and SOME CO2 was detected in all trials that contained yeast and sugar.
For future experiments, I would consider; a) exploring the minimum amount of sugar required, b) comparing the different yeast strains, and if more time allowed c) running multiple trials to see how repeatable the results using similar amounts really is.
The experiment results will conclusively serve as an aid for future students to decide how much yeast is required for their experiments.
Using AI for the Conclusion
Again, the strength of the ChatGPT (or any generative AI) is getting a summary of work, either your own or by way of a literature review.
For this section I took the entire lab and asked for a conclusion section in less than three paragraphs. This would be a level 4 use of AI.
Lab Conclusions
The purpose of this experiment was to determine the smallest amount of yeast necessary for detectable CO₂ production using Vernier Go Direct CO₂ sensors. Through a series of trials involving different concentrations of yeast in a sugar-water solution, baseline CO₂ levels were first established to ensure that any subsequent changes in CO₂ concentration were due to fermentation. The experiment successfully demonstrated that the CO₂ probes were stable and capable of detecting even minimal amounts of yeast, with as little as 0.007g producing measurable CO₂ levels. This sensitivity highlights the effectiveness of the probes for small-scale fermentation experiments.
The results showed a clear relationship between yeast concentration and CO₂ production, with higher yeast amounts leading to faster and more substantial increases in CO₂ levels. The experiment confirms that even very small amounts of yeast are sufficient to initiate fermentation, making it possible to fine-tune yeast usage in practical applications. The evaluation of the experiment also demonstrated that the probes were reliable, with minimal fluctuations in baseline readings. While practical applications of such small yeast amounts may be limited, the findings provide valuable insights into both the sensitivity of the sensors and the efficiency of yeast, offering potential for further research in fermentation processes and educational settings (OpenAI, 2024).
ChatGPT/Generative AI as a lab assistant Conclusions
I believe this was a very efficient and effective use of generative AI and honestly, it was kind of fun. As someone who has recently tried to dive deep into the writing process, I’ve been avoiding using ChatGPT as I’d rather be honing my own skills. In this application, using this tool allowed me to streamline the lab writing process in a way that still enabled me to be the generator of the novel ideas. Clearly, my knowledge of the lab writing process, comfort with the tools and technologies involved in the lab, and the understanding of the reaction allowed me to work efficiently in collaboration with ChatGPT.
I approached it as a “writing partner” rather than an “idea generator”, which, in the future, I would encourage my students to try. My hesitation though, is to which level students will not allow themselves to be the generator of novel ideas, and instead rely too much on the generative AI to produce the work.
Creating a lab report like this has taken students in the past over a month, while they struggle with creating procedures, tracking their variables and procedures, using the tools and technologies, and deciding what needs to be entered into each section of the report. They often cite staring at a blank page as a hurdle to overcome in the writing process, even though they know what they did, and how they did it, they struggle in putting it together in a cohesive way.
In running this experiment, I was able to run all the trials (some which took around 20 minutes), write the lab report, and write this entire reflective piece in three class periods (around four hours).
In doing some further research, it is clear that there are many ways generative AI is being used in the sciences, especially in the lab writing realm, already. Looking at Using ChatGPT Wisely with Your Lab Work by Zareh Zurabyan (2024), there are many aspects in which I did in this lab, and I believe all levels of lab work will continue to use to streamline the lab writing process.
As always, thanks for reading.
DiPartures continues to be a way for me to publish the projects I am working on independently, if you enjoyed reading aspects of this, please consider subscribing to keep up to date with the random assortment of interests I write about.