TinkerA faxed invoice with a smudge on its total.
Chapter 6 of 7
When pictures mislead
Two images reach the assistant in the same week: a faxed invoice with a smudge on its total, and a returned crate whose label has a sentence written for the AI.
This chapter: How can an image make the assistant wrong, or make it do something it shouldn’t?
You will learn
- Why a misread digit comes back with no sign of doubt
- How a check outside the model catches it
- Why text inside an image is data, never instructions
Words you’ll meet (3)
- Misread
- A reading that comes back confident and wrong, like a smudged 8 read as a 3.
- Cross-check
- A rule the host runs on the reading, such as “the lines must add up to the total”.
- Prompt injection
- Text written to steer a model, hidden in something it reads: an email, a web page, or an image.
The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.
Step 1 of 5
First, a supplier’s invoice that came by fax: grainy, with a smudge across the total. Read it with vision.