Student Question of the Week: Is It Safe to Use AI to Simplify Lab Protocols?
Last week I started this new series where I take time to give complete answers to questions received from college students earning their degree to become secondary educators.
Here is the second question submitted, and here is the first question from last week if you missed it.
“Is it safe to incorporate AI into labs to create a more simplified version for students who have never done a lab before?”
It’s the kind of question that sounds simple. When I first heard it, my initial response would have been YES! Bring on VR, AR, and all the things to spark curiosity and interest. But thankfully I’m taking these questions slower with more thought, and I realize this is more complex as this question actually signifies much bigger issues in education. And at the same time makes this topic urgent because it isn’t a hypothetical anymore. Teachers are already doing this. Right now. In real classrooms.
So like I often like to do, let me give you the answer that this question deserves of one with a bit of “Yes, but...” and “Absolutely not, unless...”
The Seductive Promise of AI Simplification
Let’s start with why this idea is so appealing. Picture this: You’re a first-year chemistry teacher. You’ve got a classroom of 30 students some reading below grade level, some who’ve never seen a Bunsen burner, some who are English Language Learners decoding “decant the supernatant liquid” while simultaneously trying not to spill hydrochloric acid. And of course I realize this is just a small slice of differentiated needs that all educators are facing on a daily basis.
The traditional lab manual sits on your desk like a legal document: dense, technical, intimidating. (And yes, please play along as I know this is not really happening in our digital world, so pretend it’s one of many files in your Drive if that makes you happier.) Regardless, here comes ChatGPT, offering to rewrite that acid-base titration lab at a 6th-grade reading level in about 30 seconds.
The hypothesis driving this adoption seems sound: if students spend less mental energy decoding complex sentence structures, they theoretically have more mental capacity to focus on the actual safety protocols. It’s based on solid cognitive science—specifically, Cognitive Load Theory, which tells us that human working memory is finite and easily overwhelmed, and that extraneous cognitive load (like confusing text) should be minimized to free up mental space for learning.
Additional resources on Cognitive Load Theory if you wish to explore further:
But here’s where things get dangerous: AI doesn’t actually understand what it’s simplifying.
The Fundamental Problem: AI Predicts Tokens, Not Chemistry
Let me be blunt about what Large Language Models actually do: they predict the next statistically likely word in a sequence based on patterns in their training data. They do not “know” chemistry; they do not “understand” safety. They predict tokens.
When I facilitate learning sessions about AI, most people think they already understand this. However, when I give them time to explore this interactive resource from the Financial Times that I use all the time, it helps people truly begin to grasp what I’m saying.
This matters enormously in a lab setting. Consider these realities:
The Semantic Trap
In general discourse, “mix” and “combine” are synonyms. In chemistry, they can mean very different things. An AI trained on the entire internet might treat “bleach” and “ammonia” as common household items that frequently appear together in cleaning contexts. Statistically, they’re related tokens that appear in similar contexts.
Physically? They create chloramine gas which is a potentially lethal chemical weapon.
The AI doesn’t know this. It can’t know this. It only knows that these words pattern-match in its training data. To the AI, they’re just words that go together.
The Quantification Nightmare
LLMs struggle with numbers and units. A slight token shift from “0.1M” to “10M” concentration changes a safe reaction into a violent explosion. The AI doesn’t “check” the math; it predicts the number that looks structurally appropriate in the sentence.
This isn’t theoretical. LLMs don’t actually calculate because they memorize patterns. When you see “2+2=4,” the AI isn’t doing arithmetic; it’s recalling that this exact string appeared millions of times in training. That’s a 100-fold concentration difference that could be explosive quite literally. But to the AI, both are just “number + M” patterns.
The Loss Through Lossy Compression
Here’s a real example of what happens when AI “simplifies”:
Original instruction:
“Add 5mL of 6M HCl dropwise with constant stirring to prevent boiling.”
AI-simplified version:
“Slowly add the acid to the mixture.”
Look at what disappeared: the volume (5mL), the concentration (6M), the method (dropwise), and critically, the safety rationale (prevent boiling). A novice might dump 50mL of concentrated acid all at once. The instruction is “simpler,” but the outcome could be catastrophic.
Of course I realize that most classrooms in K-12 space will not have dangerous chemicals for students to use at will let alone if we even have labs happen anymore.
And here’s the important part: The AI thought it was helping. It optimized for readability. It reduced word count. It lowered the reading level. By every metric the AI understands, it succeeded. By every metric that matters in a lab, it failed.
When I Say This Isn’t Hypothetical
This isn’t hypothetical fear-mongering. A landmark study published in ACS Chemical Health & Safety explicitly tested AI models (ChatGPT, Copilot, Gemini) as “Virtual Safety Officers.” The findings were sobering. When asked about PPE, models frequently defaulted to generic advice like “Wear gloves” but in chemistry, glove type matters immensely. Latex is permeable to acetone; nitrile is permeable to dichloromethane. This oversimplification would lead directly to chemical burns.
The Core Safety Study:
Subasinghe, S.M., et al. (2025). “Can Large Language Models (LLMs) Act as Virtual Safety Officers?” ACS Chemical Health & Safety, 32(1), 39-47.
I want to be clear: I think AI LLMs have vastly improved, and things might not be as severe as I’m making this sound. However, what I’m trying to get at in this work is that we need to be experts in our content and fields to be able to verify that indeed whatever is created for us is helpful and accurate.
The Pedagogical Problem: The Super-Cookbook
But even if we could somehow solve all the safety issues (and as these models improvesince that research paper above, the models have already improved a great deal), there’s a deeper educational issue lurking here. By using AI to over-simplify labs, we risk creating what I call the “Super-Cookbook” that are protocols so smoothed over that they eliminate the very friction that creates learning. So robotic in operation, so easy that we feed into the engagement crisis even more.
Let me explain through cognitive science:
Research into “productive struggle” suggests that a certain degree of difficulty is necessary for retention. When students must struggle to interpret a graph or design a data table, they’re engaging in deep cognitive processing. If AI generates everything for them or simplifies inquiry questions into direct commands, it removes this productive struggle.
Studies from Yale and other institutions reveal that using AI to summarize or simplify scientific concepts creates an “illusion of understanding.” Students believe they grasp concepts because the text was easy to read, but testing reveals only superficial knowledge and they cannot apply concepts to new situations. In a lab context, this overconfidence becomes dangerous. And we don’t have time to discuss the real struggle that indeed in our society today not too many like the struggle or the productive struggle to work through things.
Here’s the paradox: While novice learners need scaffolding, over-simplification can eliminate “germane load” which is the productive mental effort dedicated to building schemas and deep understanding. If AI breaks a lab into such granular steps that students never build a mental model of the whole process, they’re not learning science; they’re executing an algorithm.
The “Illusion of Understanding” Research:
Messeri, L., & Crockett, M.J. (2024). “Artificial intelligence and illusions of understanding in scientific research.” Nature, 627, 49-58.
What Labs Are Actually For
Here’s the deeper question I want to pose: What do you think labs are actually for?
If we believe labs exist to teach students to follow directions and get the “right” answer, then AI simplification might seem like a reasonable efficiency gain. Make it easier, make it faster, get to the result. And maybe we should question our goals of a lab in the first place.
But if you believe as I do that labs exist to build scientific thinking, to develop situational awareness, to teach students how to navigate uncertainty and make decisions with incomplete information, to learn from productive struggle.…. to spark curiosity and wonder which I am so passionate about then….
Then AI simplification isn’t just risky. It’s fundamentally opposed to your purpose.
Novice learners don’t just lack content knowledge. They lack schema, the mental models that let experts perceive hazards, predict interactions, and make split-second judgments. That schema is built through experience, through challenge, through mistakes that don’t quite hurt you.
You can’t shortcut that process with better instructions. You can only support it, scaffold it, and protect students while they build it.
But Here’s Where It Gets More Complicated
Because if I stop here, I’ve told you only half the story. And the other half matters for equity.
Not all simplification is dangerous. In fact, some of it is essential.
Let me tell you about Maria (not her real name). She’s a sophomore in chemistry, brilliant with mathematical thinking, came to the US three years ago, and is still building academic English vocabulary. When she reads “decant the supernatant liquid,” she doesn’t see a lab instruction she sees an impossible linguistic puzzle that she must solve while also managing her anxiety about the dangerous chemicals in front of her.
The science isn’t her barrier. The English is.
And this is where AI can be genuinely, powerfully helpful when we use it as an accessibility tool, not as a safety author.
The Green Light: Where AI Actually Helps
For English Language Learners:
Using AI to create bilingual glossaries, to rephrase academic vocabulary into simpler language structures while maintaining technical precision, to translate procedural text. this works. This is equity in action.
The key distinction:
Unsafe: “Pour the acid” (stripped of critical details)
Safe: “Carefully pour the acid (liquid) into the beaker” (simplified structure, retained safety information)
For Students with Executive Function Challenges:
Students with conditions affecting executive function often struggle with the multi-step sequencing and working memory demands of labs. AI excels at:
Converting paragraph procedures into numbered checklists
Creating visual schedules and timelines
Generating “social stories” that describe the sensory experience of the lab (helpful for students with sensory processing sensitivities)
This isn’t changing the chemistry. It’s changing the interface. It’s removing barriers to access without removing safety controls.
And this sparks a video I use all the time when supporting project based learning that perhaps what is good for one is good for all
youtube.com/watch?v=a3_Rb86wiH4
For Pre-Lab and Post-Lab Work:
Use AI to generate vocabulary lists, create real-world context hooks, develop hypothesis generation prompts, help students analyze messy data, or critique their own lab reports. Keep AI away from the actual procedure, but use it to bookend the learning experience.
The Experiment I Want You To Try
Don’t take my word for any of this. Test it yourself.
Here’s a real, standard high school chemistry procedure. I want you to feed it to AI and see what comes back:
Original Procedure: Copper Sulfate Crystal Growth Lab
“Carefully measure 50 mL of distilled water in a graduated cylinder and pour into a 150 mL beaker. Using a hot plate set to medium heat (setting 4), warm the water to approximately 60°C (use thermometer to verify; do not exceed 70°C). While wearing heat-resistant gloves, add copper(II) sulfate pentahydrate crystals one spatula-full at a time, stirring continuously with a glass stirring rod until no more crystals will dissolve. You have created a supersaturated solution. Remove from heat using beaker tongs and allow to cool undisturbed overnight on a level surface away from vibrations. The next day, you will observe crystal formation.”
Your mission:
Copy this procedure exactly
Ask your AI of choice: “Please simplify this lab procedure for students who have never done a lab before. Make it easy to understand.”
Compare what you get back to the original
What to watch for:
Did specific temperatures vanish or get rounded?
Did “heat-resistant gloves” become just “gloves”?
Did precise volumes disappear?
Did the rationale for steps (why they matter) evaporate?
Did technical terms get replaced with everyday words that change the meaning?
Did safety equipment get generalized away?
Try it with different AI models. Try different prompts. See what you discover.
Then come back and share in the comments, because I’m genuinely curious what patterns emerge. This is our research now.
Here is what ChatGPT 5.1 Thinking model created which is not too bad and is much more detailed than the original. So perhaps not only do we need to consider the models we are using but also our prompts and how we phrase things.
Measure the water
Use a graduated cylinder to measure 50 mL of distilled water.
Pour the water into a 150 mL beaker.
Heat the water
Put the beaker on the hot plate.
Turn the hot plate to medium heat (about setting 4).
Put a thermometer in the water (do not let it touch the bottom of the beaker).
Heat the water until it is about 60°C.
Do not let it go above 70°C. If it gets close, turn the heat down or off.
Prepare to add the solid (safety)
Turn off the hot plate.
Put on heat-resistant gloves.
Keep safety goggles on at all times.
Add the copper(II) sulfate crystals
Using a spatula, add one small scoop of copper(II) sulfate pentahydrate crystals into the warm water.
Stir with a glass stirring rod until the crystals dissolve.
Keep adding one small scoop at a time, stirring after each scoop.
Stop adding crystals when some crystals stay at the bottom and will not dissolve even after stirring for a while.
At this point, you have made a supersaturated solution (there is more solid than the water can normally hold).
Cool the solution
Using beaker tongs (and your heat-resistant gloves), carefully remove the beaker from the hot plate.
Place the beaker on a level (flat) surface where it will not be bumped or moved.
Leave it overnight
Do not stir, move, or shake the beaker.
Leave the beaker undisturbed overnight.
Observe the crystals the next day
The next day, look into the beaker.
You should see new crystals that have formed in the solution and on the bottom of the beaker.
So What Do We Actually Do?
If you’ve read this far, you’re probably feeling some combination of thoughts and feelings. Let me give you something concrete.
If you’re determined to use AI to support novice learners in the lab, here are the non-negotiable protocols:
1. The Human-in-the-Loop Rule
Before any student sees an AI-generated procedure, you must physically perform the lab using only those instructions. Nothing else. No filling in gaps from memory. No assumptions. Like Jeff Robin says about project based learning, it applies to AI learning, Do The Project FIRST!
If you find yourself reaching for equipment not listed, or pausing to verify a safety concern not mentioned, the AI has failed. Full stop.
2. Never Let AI Write the Core Procedure
Use verified, peer-reviewed procedures from reputable sources such Flinn Scientific, NSTA, textbook publishers with liability insurance and expert review processes. These organizations have legal and ethical stakes in getting it right.
Let AI help with accessibility around the edges: vocabulary support, visual schedules, pre-lab context. But the chemical instructions themselves must come from human experts who understand physical reality.
3. Advanced Prompt Engineering
If you’re going to use AI for any lab materials, treat prompting as a technical skill requiring precision:
Use persona prompts:
“Act as a certified Chemical Hygiene Officer with 20 years of experience in high school chemistry labs.”
Use constraint prompts:
“Rewrite this for a 6th-grade reading level. DO NOT remove any safety warnings. BOLD all PPE requirements. INSERT a ‘Stop and Check’ box before any heating step. DO NOT use household analogies.”
Use verification prompts:
After getting output, ask: “Identify any safety information that was present in the original but missing in your simplified version.”
4. Turn AI’s Weaknesses Into Teaching Tools
Give students an AI-generated lab procedure that contains deliberate errors (which you’ve verified and controlled). Ask them to identify the safety risks.
This transforms AI’s hallucination problem into a powerful teaching moment for critical thinking. Students learn to be skeptical, to verify, to think rather than trust.
They learn that “sounds right” and “is right” are very different things possibly the most important lesson in the age of AI.
The Liability Reality Nobody Mentions
Here’s something that should matter to every educator using AI for lab materials:
If a student is injured following AI-generated instructions, legal liability likely rests with you and your school district, not with OpenAI, not with Google, not with Microsoft.
The terms of service for these LLMs explicitly disclaim liability for high-risk advice. You are the licensed professional in the room. You are responsible.
This isn’t about fear-mongering. It’s about informed consent. You should know what you’re accepting when you click “generate.”
My Final Thought
To that college student who asked this question, and to everyone reading this:
Thank you. Thank you for asking before acting. Thank you for caring about both safety and equity. Thank you for recognizing that this is complicated.
AI is a powerful tool.
Use AI to clear the clutter. Use it to translate vocabulary for your ELL students. Use it to create checklists for your students with executive function challenges. Use it to generate hooks and context and post-lab analysis support.
Use it to create more time for the irreplaceable human work of teaching: observing, questioning, adjusting, connecting, caring.
But never ever hand it the keys to the chemical storage cabinet.
Because the ultimate safety device in any laboratory, and the ultimate teaching tool in any classroom, is not an algorithm. It’s the critical, skeptical, and attentive human mind.
Your mind. Your judgment. Your care for the students in front of you.
That’s not something AI can simplify away. And that’s exactly as it should be.
Research Resources & Further Reading
The Core Safety Study:
Subasinghe, S.M., et al. (2025). “Can Large Language Models (LLMs) Act as Virtual Safety Officers?” ACS Chemical Health & Safety, 32(1), 39-47.
The “Illusion of Understanding” Research:
Messeri, L., & Crockett, M.J. (2024). “Artificial intelligence and illusions of understanding in scientific research.” Nature, 627, 49-58.
Related: “Doing more, but learning less: The risks of AI in research” - Yale News
Understanding How AI Works:
Visual Vocabulary of Generative AI - Financial Times Interactive
Cognitive Load Theory Resources:
Professional Guidelines:
ACS Committee on Chemical Safety
Division of Chemical Health and Safety
Have a question you’d like featured? Send it my way. The best questions are the ones that make us stop and think and this one certainly did that.
Try the experiment and share your results in the comments. Let’s build a collective understanding of what AI actually does when we ask it to “simplify” science.
All images created using Gemini 3 and latest version of Nano Banana