ChemIllusion Can't Fix Cheating, but It Can Help
No technology can create academic honesty. That takes an instructor's whole approach to teaching, assessment, and classroom culture. ChemIllusion now has tools for in-class and out-of-class assessments that include photographing a model kit, drawing a structure in the air with their hands, or naming the molecule aloud. All with real-time feedback. This post covers what the tools do, what they do not do, and where they fit in a larger strategy of engaging students and keeping them honest.
Software is one piece
ChemIllusion can't fix cheating. No technology can create academic honesty. That takes an instructor's whole approach to teaching, assessment, and classroom culture. As an organic chemistry professor I distrust any product that claims otherwise, and that includes ours. Software can support parts of that strategy, but it cannot stand in for it.
What has changed is the pressure on online homework, and what we can ask students to do in response. Nearly every structure question there is answered by drawing in a browser, and a drawn structure is the easiest kind of answer to get from a friend, a solver or a chatbot. ChemIllusion now has tools for in-class and out-of-class assessments that include photographing a model kit, drawing a structure in the air with their hands, or naming the molecule aloud, all with real-time feedback. These modalities are harder to outsource and closer to how chemistry is taught, without turning ChemIllusion into a proctor.
Build it: answer with a model kit
With Build it, a student answers a structure question by building the molecule with a Molymod kit and photographing it in the homework player. An AI reads the photo and proposes a structure. The student confirms it or fixes it in the same structure editor used for drawn answers, and the confirmed structure is graded by our normal deterministic grader, exactly as a drawn answer would be. The AI never grades.
These photographs are from the Build it demo. The question is HBr addition to propene. The student builds the product with a Molymod kit, photographs it, and the structure the AI reads is graded like any drawn answer.
A few details matter for how it behaves in a course:
- The camera starts only when the student taps Build it, and the only thing sent is the one photo the student has reviewed. In an assessment there is no option to upload from the photo gallery.
- The photo, the AI's proposed structure and whether the student edited it are kept as evidence the instructor can open from question history. Photos are deleted when the course closes.
- The instructor decides, assignment by assignment, whether Build it is allowed. Drawing always stays available, so no question becomes camera-only.
Build it works for homework, quizzes and exit tickets in ChemIllusion courses now. Laptop cameras usually point at the student rather than the desk, so students can also choose to continue a question on their phone: they scan a QR code, take the photo there, and submit back on the computer. Inside an LMS, where the browser frame blocks the camera, the phone is the way in.
You can try it on the Build it demo page, free with a registered account.
The images below come from ChemIllusion's 3D molecule preview, the same renderer used in the Generator. They illustrate the kinds of tasks available on the Build it demo. Pointing at an atom or bond and reading a dihedral conformation from a photo are experimental demo tasks; they are not yet answer types an instructor can enable for graded homework.
Start: eclipsed butane. Task: make it gauche.
Start: methyl axial. Task: ring-flip to equatorial.
Point to the carbon a nucleophile attacks.
Update, November 2026: a model on every question
The Build It / Draw It question set now gives every question the model it is about: 166 questions in 12 modules, from methane to dipeptides. Each chain of questions starts by building that model with the kit, and every question shows a 3D version students can turn, so a ring flip or the back face of a carbonyl is something to look at rather than imagine. A photo of the model is the default answer; students without a kit or a camera switch on Drawing allowed and draw it instead. The models are rendered from real 3D coordinates in the kit's colours; no AI draws them.
Model kits are built for neutral molecules. Amino acids sit at pH 7 mostly as zwitterions, so those questions offer the zwitterion view too and say plainly that building the neutral form is a correct answer.
Model kits are designed to show neutral molecules, so the kit model, and the build each question asks for, is the neutral form. At pH 7 glycine exists mostly as a zwitterion (the amine protonated, the carboxylic acid deprotonated). Building the neutral form is a correct answer; where a question asks you to draw the structure, either form is accepted.
Drag either model to turn it, or focus it and use the arrow keys. Switch glycine to its zwitterion to see what the kit cannot show.
Browse the full Build It / Draw It curriculum: its modules, the textbook chapters each question maps to, and the Chapter 3 questions as students see them, open to anyone. Teachers can open every question and add the set to a course from the instructor preview.
Sign it and Say it
Sign it uses hand gestures. On the gesture drawing page, the camera follows your hand while you draw the structure in the air: pinch to place an atom, drag to draw a bond, and hold a gesture to change a bond order or an element. Recognition runs on your device, camera frames never leave it, and nothing is sent to an AI service.
In a direct ChemIllusion course an instructor can allow Sign it as an answer source. When a student submits that way, a snapshot of the structure they drew is saved as evidence. The snapshot is a picture of the molecule, not of the student. Gesture drawing does not handle wedge and dash bonds or charges, so it suits connectivity questions best.
A one-page student handout explains the machine learning behind Sign it: a local hand tracker proposes a gesture, then chemistry rules decide whether the requested edit is allowed. Download the student handout (PDF) to share before students try the camera.
Say it is the voice version. On the Say it demo page, you look at a structure and name it out loud. Your browser's speech recognition turns the words into text, we match the name to a structure, and that structure is checked like any other answer. Naming a molecule aloud is a different skill from drawing it, and it is harder to hand off quietly. The demo is open to everyone; Say it is not yet an answer source an instructor can turn on for graded homework. For drawing by voice, Quick voice edit is in beta on the Generator and in the Research ELN drawing editor.
What these tools do not do
Overclaiming here would hurt instructors and students, so I want to be specific about the limits.
These tools raise the effort needed to outsource an answer. They do not prove authorship. A student can photograph a model someone else built, or have a friend draw in the air for them. Blocking gallery uploads removes the easiest shortcut and nothing more.
There is no cheating detection. Nothing flags a student, scores how suspicious an answer looks, compares one student with another, or proctors anyone. The instructor sees the evidence, when it was captured and whether the student changed the AI's reading, and deciding what that means is left to a person.
The cameras read a model kit or a hand. They never analyze faces, emotions or attention, and if a face appears in the background of a model photo, nothing looks at it. Europe's AI Act treats software that monitors students during tests as high-risk and bans inferring emotions in schools.
The AI can also misread a photo, with bond orders the hardest part to read from a picture. That is why the student always checks the structure first, and why grading runs only on the structure the student submits.
A comprehensive strategy
No single measure works on its own. What follows is how I think about honest work as an instructor, and almost none of it needs software.
Design the assessment around reasoning. A question that asks only for a product structure is easy to outsource. Asking for the mechanism, a sentence on why one product forms over another, or a variant students have not seen in the practice set makes their own reasoning the thing being graded.
Keep practice low-stakes. When homework is where students are allowed to be wrong, with feedback and another try, the reason to borrow an answer gets smaller. The high-stakes moments can then be ones I see directly.
Make the policy clear. Students should know which tools, AI included, are allowed on which assignments and why. A vague policy leaves room for misunderstanding on both sides.
Use more than one modality. Physical model kits work well in the classroom and for homework. A student who has built chair conformations of cyclohexane in class has a reason to reach for the kit at home, and an answer built by hand sits closer to the understanding I am trying to grade.
Build relationships. I believe students are less likely to cut corners in a course where they know the instructor and trust that the grading is fair. That trust is built in class and in office hours, long before any tool sees an answer.
Discuss the need for mental effort in learning and the risk of offloading that effort to AI too early.
Where the tools fit
Build it, Sign it, and Voice mode are one piece of that strategy. Students get a way to show their work with their hands or their voice, the easiest shortcuts take more effort, and instructors get something concrete to discuss with a student when a question comes up. Everything else on the list above stays with the instructor.
If you teach chemistry and want to pilot Build it or Sign it in a course, start on the teachers page or try the Build it demo first. I would like to hear what works in your classroom and what does not, so please get in touch.