Pre-launch · Training and the free Foundations exam are open now. Levels 1–5 open soon. Get notified on any level page.
20 min read
Many AI tools now accept and produce images, audio, video, and long documents. The same discipline applies as with text: say what evidence counts, separate what's observed from what's inferred, and make outputs checkable. What changes is how each medium fails. This module covers inputs you analyze, outputs you generate, and how to keep people in charge of decisions built on either.
[inaudible 01:42-01:45] instead of filling in a plausible word.When multimodal analysis feeds a decision about money, safety, or someone's job, a person reviews the evidence and makes the call.
Ruth Abara manages claims at a regional moving company. When a customer reports damage, the file includes photos, a crew bodycam clip, the customer's voicemail, and the service contract. A team lead wants AI to "review the claim and decide payment."
Weak approach.
Before:
Look at the photos, video, voicemail, and contract and decide whether
we should pay this claim. Explain what happened.
The tool may blend observation with guesswork ("the crew dropped the cabinet"), fill an inaudible phrase in the voicemail, miss a liability cap in a contract appendix, and issue a payment decision nobody reviewed.
Better approach.
After:
Prepare an evidence summary for a human claims reviewer. Do not
recommend or decide payment.
Photos: For each, describe visible damage and location on the item.
Label each "damage visible", "no damage visible", or "unclear".
Video: List timestamped observations of the item. Put any
interpretation in a separate section labeled "Inference".
Voicemail: Transcribe. Mark unclear words as [inaudible mm:ss].
Contract: Before summarizing, list every section and appendix. Then
quote, with section numbers, the clauses on liability limits,
claim deadlines, and exclusions.
If any file is missing or unreadable, say which and stop that part.
Why this is stronger. Each medium has its own evidence rule. Observations and inferences are separated, so the reviewer can see what's actually shown. Uncertain audio is marked, not invented. The contract coverage list catches the appendix with the liability cap. And the payment decision stays with Ruth's reviewer, who checks the summary against the original files.
Setup. Gather a photo of a room or object, a chart with labeled axes, and a short audio clip or voice memo you recorded.
Task. Write one prompt per item. For the photo, ask for a description of visible content only, with a separate "inference" list. For the chart, ask for title, axes, and units before any reading of values. For the audio, ask for a transcript with unclear spans marked and timestamped.
What a good result looks like. Each output can be checked directly against its source. Nothing in the photo description is outside the frame. The chart reading names units. If any span is unclear, the transcript marks it rather than smoothing it over.