When it fails: yearbook photo mix-ups

Common failures are often pattern mistakes, not cartoon meltdowns.

In the yearbook draft:

• Inputs: faces, old tagged photos, maybe club rosters
• Job: propose who appears in each image
• Outputs: name labels on thumbnails
• Failure modes: lookalikes, twins/siblings, hats and shadows, rare faces with few examples, names copied from the wrong album

A confident tag is still a proposal. If the club prints without checking, the mistake becomes permanent paper.

1. Lookalike swap — similar haircuts, uniforms, or angles confuse the model.
2. Missing representation — students who rarely appear in past tagged sets get worse accuracy.
3. Context blindness — the tool does not know “this is a visitor,” only “this face matches a pattern.”
4. Overconfidence — a high “match” score can still be wrong.
5. Silent miss — a person is in the photo but never tagged, so the error is absence, not a wrong name.

Each flavor can still fit your mental model: designed system, bounded task, inputs, outputs, consequences. Failure diagnosis starts by naming which box broke.

For any suspicious tag, ask:

• What evidence did the system likely use?
• What similar-looking example could fool it?
• Who should confirm before publishing?
• Is the cost a funny typo — or a public embarrassment?

Then pick a fix that matches the failure: more human review for cover photos, a “needs check” badge on low-data faces, or blocking auto-publish entirely.

Do not “fix” a yearbook error by yelling that AI is fake. Fix the process around a bounded tool.

A strong poster line might read: “Some AI tools guess from patterns. Guesses need a human check before they become the yearbook.”

That is more useful than “AI is dangerous” or “AI is magic.” It teaches classmates to expect plausible mistakes in photo tools, feeds, and chatbots alike.