On this page
If you cannot see a diagram, a physics question that begins "In the figure shown..." can stop you before you start. The WHO World report on vision, published in 2019, estimates that at least 2.2 billion people worldwide have a vision impairment. Three kinds of tools help: maths-aware screen readers, image-description apps and AI chatbots that explain steps. None is reliable enough to trust without checking.
What tools help visually impaired students read equations?
A screen reader turns on-screen text into speech or braille. NVDA, a screen reader for Windows, is free to use. Equations are harder: fractions, powers and roots have structure that plain speech can flatten.
MathCAT generates speech and braille, including Nemeth and UEB Technical, from MathML, the code web pages use to store equations. It lets you move through an expression piece by piece or hear an overview first. One gap matters in India: its README lists speech languages such as English, French and Vietnamese, but no Hindi or other Indian language is supported yet; Urdu is listed as in progress.
The practical point: an equation is only readable if it is stored as maths, not as a picture of maths. Ask for digital notes and papers in an accessible format. For the history, see how Louis Braille's six-dot code let blind students read.
Can AI describe diagrams and graphs accurately?
Sometimes. Be My AI, announced by Be My Eyes on 7 August 2023, uses GPT-4 to describe photos you take, answers follow-up questions, and lets you reach a human volunteer to check.
For STEM content, results vary a lot. A 2024 study by Marina Buzzi and colleagues tested five AI tools on four STEM images, rated by the authors. Seeing AI did not identify the STEM content, Be My Eyes described simple images well but complex ones less accurately, Microsoft Copilot gave the best results though with some misinterpretations, and Gemini failed to recognise a finite state machine diagram. In this preliminary study, the authors conclude that for STEM educational content "more effort is required".
Charts are harder still. In the VizAbility study presented at UIST 2024, GPT-4V answered questions from chart images alone, and only 27.96% of its answers were rated "Very Good". The researchers' own system, which also used the chart's underlying data, had 77.36% rated "Very Good".
Where do AI explanations still fall short?
A 2026 preprint survey of 20 peer-reviewed studies on AI descriptions of STEM images lists challenges that persist: factual inaccuracies and hallucinations, too few datasets co-designed with blind and low-vision users, and automatic scores that poorly capture usefulness and trust.
In practice, watch for three failures:
Invented numbers: a value guessed from an unlabelled axis or scale.
Missing relationships: every part named, but not which arrow points where.
Confident slips in the maths that follows. Our guide on why AI chatbots get maths wrong has a five-step check.
How do you get a useful description of a diagram?
Guidelines for describing STEM images from NCAM, the National Center for Accessible Media at GBH, say descriptions should follow a "drill-down organization": a brief summary first, then extended description or specific data. Ask an AI to do the same, one request at a time:
"In one sentence, what kind of diagram is this and what is it for?"
"List every labelled part and every number with its unit, exactly as written. Say 'not labelled' instead of guessing."
"Describe how the parts connect, one relationship per line."
"If this is a graph or chart, give the data as a table."
"Which parts are you unsure about?"
Worked example, a block resting on a ramp. A good answer sounds like this:
Summary: a free-body diagram of a block at rest on a ramp inclined at 30 degrees.
Parts: the block, the ramp surface, three force arrows.
Relationships: weight points straight down; the normal force points away from the ramp, perpendicular to its surface; friction points up the slope.
Now check it. If the AI says friction points down the slope while the block sits at rest, or gives an angle that differs from the question text, stop and ask a teacher or classmate. If a sighted helper photographs problems for you, our guide to getting the figure right when you photograph a problem helps them take a clearer picture.
Where does an AI tutor fit in?
Once the diagram is clear, the hard part is the reasoning. Ask for an explanation you can interrupt: slower, a step restated, each symbol said in words. Fermi's AI tutor teaches rather than just answering: it notices where you are stuck and decides whether to explain, ask or show. The Fermi Problem Library has worked problems with a step-by-step solver that explains the reasoning. As with any tool, try it with your own screen reader before relying on it.
Parents: ask the school for accessible digital material, and check exam accommodations on the current official notice from your board.
Start with one diagram from this week's homework and the five requests above.
