Give examples of the data formats (modalities) that multimodal AI works with.
Explain that generative AI output has randomness, and how this brings varied expression but makes results hard to reproduce.
Distinguish XAI, a technology that asks the AI itself for an explanation, from human-in-the-loop, a mechanism that puts people into the processing flow.
Multimodal AI (マルチモーダルAI) is AI that can combine and handle data of different formats, such as text, images and audio (each format is a modality, モダリティ).
Input format
Output format
Example exchange
Image and text
Text
Show a photo, ask what is in it, and get a written answer
Text
Image
Describe a scene in words and have the AI draw it
Video (picture and sound)
Text
Have the AI turn a recorded meeting into a written summary
AI that takes in and puts out only one format is called single-modal (シングルモーダル), the counterpart of multimodal AI. Multimodal AI can take a mix of formats as input, or answer in a different format from the input, so exchanges can cross formats.
AnalogyLike a person who hears a question about something they are looking at and answers out loud, it links several senses together.
How to remember"Modal" refers to a mode or format; add "multi" (many) and it reads as "many formats."
Exam points
Mix-upThe "multi" in multimodal AI (マルチモーダルAI) counts data formats, not the number of AI systems or users.
Randomness (ランダム性) is the property that generative AI builds its output from probabilities, so the same prompt can give a different output each time.
Enter the same prompt
Assign a probability to each candidate next word
Pick a word according to probability
A different text comes out each time
Because the chosen words vary, the finished text can differ from run to run.
4 / 4
For example, a text-generating AI assigns a probability to each candidate next word and picks words according to those probabilities, so when the chosen words change, the finished text changes too. This variation makes expression more varied, but it also makes the same result hard to reproduce. How much the output varies can be turned down or up, for instance through settings at generation time.
AnalogyIt is like ordering the same chef's choice at the same restaurant: the plating and side dishes differ a little from day to day.
How to remember"Random" usually means haphazard or arbitrary, but here it means variation that follows probabilities.
Exam points
Mix-upRandomness (ランダム性) means the output varies according to probabilities. It does not mean the AI is broken or malfunctioning.
XAI (Explainable AI, 説明可能なAI) is the technology and approach of showing people, in a form they can follow, the grounds on which an AI reached its conclusion.
Enter data
The AI makes a decision
Show the result and its grounds
A person checks the grounds and decides whether to accept
Looking at those grounds, a person decides whether to accept the result.
4 / 4
Advanced AI runs complex internal calculations, so it easily becomes a black box (ブラックボックス) where no one outside can see why it gave a result. XAI makes the grounds visible, raising the transparency (透明性) of decisions so people can check results and be accountable for them.
AnalogyLike an answer sheet that shows the working, not just the answer. With the working visible, you can find exactly where the reasoning went wrong.
How to rememberThe X at the front comes from "eXplainable."
Exam points
Mix-upXAI (説明可能なAI) is technology for making AI decisions explainable. The term does not refer to technology that raises the accuracy of those decisions.
Human-in-the-loop (ヒューマンインザループ) is a mechanism that, instead of leaving processing entirely to the AI, builds human checks and decisions into the processing flow.
← Scroll the figure sideways →
For example, a person reviews a proposal from the AI, approves it for execution if it is sound, and sends it back with corrections if not. With people stepping in at key points, AI errors and bias are stopped before anything is executed, and it is clear who made each decision and who is responsible.
AnalogyIt resembles approval inside a company, where a manager reads and signs off on a document a staff member drafted before it goes out.
How to rememberThe Japanese term simply writes the English "Human in the loop" in katakana. Picture a human standing inside the loop.
Exam points
Mix-upHuman-in-the-loop (ヒューマンインザループ) adds people to the processing flow, the opposite direction from full automation, which hands all human work to AI.
XAI vs. Human-in-the-loop
Aspect
XAI
Human-in-the-loop
Key differenceWhat is added
A function on the AI's side that shows the grounds for its decisions
Human checks and decisions inside the processing flow
Mainly improves
Transparency of decisions
Ability to stop errors and bias, and clarity about who is responsible