Make a request to the sighted AI so that it can parse the graph correctly. Return JSON {"prompt":"...","read_first":"...","guardrail":"..."}: prompt - what to explain, read_first - how to force the AI to first confirm that it read the graph correctly before drawing conclusions, guardrail - a restriction against fictitious exact values if there are no signatures.
AI understands the graph in two steps: first it recognizes the type (columns, line, circle) and axes labels, then it “reads” the form - where there is growth, where there is failure. The vulnerability is that the model may misunderstand what is plotted along the axes, or invent exact values where there are only bars without labels on the graph. Hence the main trick: make the AI first say the type of graph, what is horizontal and vertical, in what units - and only then draw conclusions. If he mixed up the axes, you will catch it immediately, before making erroneous decisions. The second technique: do not ask “draw a conclusion at all,” but point-by-point – where is the peak, where is the dip, are there any anomalies and how can they be explained. Hidden lever of honesty: directly forbid inventing specific numbers if there are no labels of values on the graph - let him say “approximately” and “you can see from the columns.” A typical beginner mistake: immediately asking for a ready-made output without checking that the AI has read the axes and units correctly.
Unlock access to submit solutions for instant AI review.