Confident with AI course contents
Part Three: Understand Images and Visual Information
10. Judge and Verify an Image Description
Classify claims as visible, contextual, inferred or unsupported, then verify important details proportionately.
The lesson
Verification without sight means using narrower questions, independent evidence and human help where needed.
Use four categories consistently: visible, contextual, inferred and unsupported. Uncertainty is a property a claim can acknowledge, not a separate evidence category.
A model may present an inference or unsupported guess with the texture of visible fact.
Classify the claim before relying on it
- Visible: the image itself supplies the evidence, such as printed words, colours or objects.
- Contextual: the evidence comes from a caption, filename, surrounding page or other source rather than the image alone.
- Inferred: the claim interprets visible evidence, such as mood, genre, relationship or intention.
- Unsupported: the available image and context do not justify the claim, even if it is stated confidently.
Ask how likely the detail is to be wrong
- Could the image establish the claimed name, relationship or location?
- Is exact text small, stylised, curved or obscured?
- Does the answer depend on counting or left and right?
- Is an inference being presented as visible fact?
- Does the wording acknowledge appropriate uncertainty?
Decide what is proportionate
Check a claim when being wrong could cost money, affect safety, require an apology or cause a public correction. A decorative detail with no effect on meaning may not need checking. Use the green, amber and red risk bands from Module 6 when the decision is unclear.
Ways to verify without sight
- Ask a narrower question about the disputed detail.
- Ask whether the evidence came from the image, context or inference.
- Request an independent description.
- For exact text, compare the answer with optical character recognition, or OCR.
- Check captions, filenames and surrounding content.
- Use human assistance or an authoritative source when the consequences justify it.
Agreement, disagreement and no verdict
Agreement increases confidence but does not prove correctness. If two descriptions disagree, a third independent reading may help, but it does not automatically settle a consequential dispute. Escalate important unresolved details to an authoritative source or person.
If verification is not possible, either omit the claim or carry it forward with its provenance and uncertainty attached. For example: “One image-description model reported a red logo, but I could not verify that detail independently.” Do not repeat it as established fact.
Worked example
Return to the book cover from Module 9. “The title reads The Stranger Times” is a visible-text claim, but stylised lettering makes the transcription fragile. “C. K. McDonnell” is visible printed text; confirming that it names the correct author and edition requires contextual or authoritative evidence. “The cover suggests comic fantasy” is inferred. A confidently reported quotation that cannot be read in the image or supported by context is unsupported.
Whether to verify each claim depends on purpose. A casual impression of the cover may need no further check. Reproducing the title, author or quotation in published work does.
Try it
Take the description you produced in Module 9, or another non-sensitive description. Label its important claims as visible, contextual, inferred or unsupported. Choose one claim, explain what would happen if it were wrong and decide whether to accept it, verify it or carry it forward with an explicit qualification.
Checkpoint
You can name the claim and category, explain its fragility and consequences, and justify the proportionate response you chose.
About this course
Confident with AI was created by Charli-Jo Tyrer in association with the Technology Association of Visually Impaired People (TAVIP).
Licensed under CC BY-SA 4.0.