AI GUIDEDifferent media, different operations
Multimodal means working with more than one kind of input or output, such as text, images and audio. It does not mean that every product can perform every operation. Reading an image, editing it and generating a new one are different capabilities. Transcribing speech and speaking a generated response are different stages too.
A model may process several media directly, or an application may combine specialised steps. An audio recording might become a transcript, the transcript might become notes, and a speech system might narrate those notes. At each transformation, information can be lost or changed.
At Advanced, ask what representation crosses each boundary, what evidence survives and how an error is detected. This lesson follows AD02. It does not require image-generation credits or a recorded lecture: all source material for the exercise is supplied below.
A worked source pack
Fictional transcript T1: “We held three workshops. The first had 8 attendees, the second 12 and the third 10. We have not investigated why attendance changed.” Fictional slide S1: “Attendance: 8, 12, 10. Capacity: 15 per workshop.” CSV C1 contains rows W1,8,15; W2,12,15; W3,10,15, with columns workshop, attendees, capacity.
Task: prepare a short learning note and a chart of attendance. The transcript supplies the claim, the slide confirms the values, and the table provides an explicit calculation input. Expected total attendance is 30 visits, not necessarily 30 unique people. Capacity utilisation by workshop is approximately 53.3%, 80% and 66.7%.
Do not derive a cause from the chart. Nothing in T1 says that a new speaker, advertising or AI improved attendance. Do not infer the identities of attendees from an imagined photograph. A multimodal workflow gains capabilities, not permission to invent missing evidence.
Keep transformations visible
For a real recording, keep the original and references to relevant time positions. Speech recognition may confuse names, accents or numbers. For a slide image, retain the image and identify which value was read from it. Optical character recognition extracts text from pixels; visual interpretation may also infer layout or relationships, so ask which observations are explicit and which are interpretations.
For numerical work, use an explicit table and inspect the calculation rather than asking an image generator to draw exact data bars. A code-execution tool can calculate and generate a chart; an image generator can produce an illustrative scene. Swapping these operations can create visually convincing but inaccurate data graphics.
Finally, create the prose or presentation from the checked facts. Keep source references in speaker notes or an accompanying source list. If you generate a spoken summary, listen for changed numbers and missing qualifications. Reformatting a claim does not increase its truth.
Inject and resolve a conflict
Now change S1's second attendance value to 21 while T1 and C1 still say 12. Do not vote two sources against one automatically: they might share the same original error. The correct first result is a visible conflict, with references and a request to identify the authoritative record.
If an authorised organiser confirms C1 as the approved record, use it for calculations and keep a note about the corrected slide. Without that confirmation, report the disputed workshop separately and do not publish a definitive total as established fact.
Exercise: draw a pipeline with input type, transformation, intermediate output, check and failure action for each stage. Mark which stages need a model, which need ordinary arithmetic and which need a person. Compare it with a simpler workflow using only the table. More input types are useful only when they add necessary evidence or a useful output format.
Finish and go deeper
You complete this lesson when another person can trace each chart value and substantive claim to an identified source, reproduce the arithmetic and see unresolved conflicts. Your outputs are a checked table, a short note and a pipeline diagram, not an automatically verified presentation.
For deeper study, compare the linked image-understanding documentation with image-generation guidance and code-execution documentation. Notice the different inputs, outputs and limitations. For speech, inspect your chosen product's current transcription and voice documentation rather than assuming a feature from another product is included.
Continue with A09 on tools and agents. Then P05 and P06 connect technical capability to evidence and practical value, before AD04's final project.
Sources
Read the original sources to explore the topic and check specific claims.