On this page
Snapping a photo of a tough problem and asking an AI chatbot about it takes ten seconds. Getting accurate help takes a little more care, because the weakest link is often not the maths but the reading of the figure. If the AI misreads one label, angle or arrow, every step after that can be confidently wrong. This post explains what research shows about AI and diagrams, and gives you a photo checklist and a read-back step that catch many errors before they start.
How good is AI at reading diagrams?
Better than it was, and still uneven. Two studies show where the trouble sits.
MathVista, a benchmark of 6,141 maths examples built around figures, charts and diagrams, tested 12 AI models in 2023 (Lu and colleagues, 2023). The best of them, GPT-4V, scored 49.9%, which the authors report as 10.4% short of human performance. They note that it "often struggles to understand complex figures and perform rigorous reasoning".
A 2024 study called "Vision language models are blind" used seven very simple tasks, such as whether two circles overlap or how many times two lines cross (Rahmanzadehgervi and colleagues, 2024). Four state-of-the-art models averaged 58.07%; the best scored 77.84%, where people would be expected to get everything right. The models did far better when shapes were spread apart. They struggled most when lines and shapes overlapped or sat close together, which describes many physics figures: forces drawn from one point, rays crossing at a lens, wires meeting at a junction.
Models have been updated since these papers, so do not treat the numbers as today's scores. Treat the pattern as the lesson: crowded, small or overlapping details are where misreadings happen.
A photo checklist that helps the AI read your figure
One problem per photo. Crop out neighbouring questions so the AI cannot mix up their data.
Shoot straight down in good light. Avoid shadows and tilted pages that stretch angles.
Make the figure large in the frame. Small labels are the first thing to be misread.
Include the options if it is a multiple-choice question, and the units.
Type the numbers too. Under the photo, write the given values in text: "m = 2 kg, θ = 37°, μ = 0.25". If the photo and your text disagree, you will see it.
Redraw crowded diagrams. If several forces or rays overlap, a quick clean sketch with spaced-out labels is worth it. Drawing it also makes you think, as in why drawing the diagram first helps.
The read-back step: check what the AI saw
Before asking for any solution, ask the AI to describe the figure. A prompt that works:
"Before solving anything, describe this figure: list every object, every labelled quantity with its value and unit, the direction of every arrow, and any angle and what it is measured from. Then stop and wait for me to confirm."
Then compare its list with the page. Example: a block on a 37° incline is attached by a string over a pulley to a hanging mass. A misreading such as "angle of 37° with the vertical" instead of the horizontal swaps sin and cos for every force along the slope. You will catch that in the read-back in five seconds; you might never catch it in a finished solution.
How to ask so you learn, not just get an answer
Once the figure is agreed, say what you have tried: "I got 4.2 m/s², the answer key says 2.6. Here is my free-body diagram. Where does my reasoning go wrong?" This gets you feedback on your thinking rather than a fresh solution to copy. Then check the AI's working the way you would check anyone's: units, a limiting case, and a rough estimate. Our guide on when AI gets physics wrong has a full checking routine, and getting AI feedback on handwritten solutions covers photographing your own working.
At Fermi, a student can bring a problem set or a scanned assignment to Fermi's AI tutor, which teaches through it by writing on a board and drawing the figure. Seeing the figure redrawn is itself a read-back: if it does not match your page, say so before going further.
A 60-second version for exam season
When time is short, do three things: crop to one question, type the given values under the photo, and ask "describe the figure first". Those three habits target the exact weakness the research found, and they cost less than a minute.
