Introduction

​

The problem

AI tools now generate many of the graphics students see: charts, maps, diagrams, geometry figures, and worked solutions. Most checks on generated content read only the text, so a graphic passes through unexamined. A graphic can fail a student in two ways:

• It’s wrong. A miscounted feature, a mislabeled axis, or a flawed step can look right at a glance and still teach the wrong thing.
• It’s too hard to read. A passage can sit on grade level on every language measure, yet put its key idea in a graphic that students at that grade haven’t learned to read, like a dual-axis graph.

At the volume AI produces, no team can review every graphic by hand. Research shows that graphics help students learn when they carry information relevant to the text, not when they decorate it (Levie & Lentz, 1982 ↗), and only when readers connect what they see with what they read (Hannus & Hyönä, 1999 ↗; Mason et al., 2013 ↗). Reading a graphic is also a skill in its own right: the same display can be clear to one reader and opaque to another, depending on what each knows about how that kind of graphic works (Shah & Hoeffner, 2002 ↗). That barrier falls hardest on students with less exposure to graph and diagram conventions.

​

What we’re building

Each Graphics evaluator takes images as input, assesses a different aspect of the graphics in instructional content, and returns its reasoning, so you can see what to fix and not just whether something’s wrong. The Graphics evaluators accept PNG, JPEG, and WebP images. Images go directly from your environment to the model provider, using your own API key, and are never sent to Learning Commons.

​

Use cases

• Check graphics before they reach students: Run generated graphics through an evaluator. When it flags a graphic as incorrect, or as more complex than you intended for the grade, send that graphic to be regenerated or reviewed by a person.
• Find problems at the source: Look across the graphics an evaluator flagged as incorrect or too complex for recurring issues, like misread axes or key content that appears only in a graphic, and fix the prompt that produced them.
• Match graphics to instructional goals: Use an evaluator’s complexity rating to check whether a graphic fits your instructional goal: simple enough to support reading the prose, or demanding enough to teach students to read and integrate visual information.
• Monitor changes over time: Compare results across model updates or prompt changes to catch shifts in the graphics your product generates.

​

Limitations

​

Related topics