0.7.15 Modeling AI and Human Judgment With a UML Activity Diagram

AI Use Is a Process

AI interaction can be modeled as a sequence of actions and decisions.

A useful process includes both:

• the AI system;
• the human user.

The AI generates an output.

The human evaluates and decides what happens next.

A UML activity diagram can make those responsibilities visible.

Use Activity Partitions for Responsibility

You have already used activity partitions, or swimlanes, for browser/server interactions.

The same idea can separate:

Human
AI System

Actions go in the lane of the participant responsible for them.

For example:

Human lane

Define task
Prepare permitted input
Evaluate output
Verify claim
Choose final action

AI System lane

Receive prompt
Generate response
Return response

Begin With the Human Goal

The flow starts before the prompt.

A person should first identify the task.

Conceptually:

HUMAN
Start
  ↓
Identify task
  ↓
Decide what information is appropriate to share
  ↓
Create prompt

This prevents the diagram from implying that responsible AI use begins only after the output arrives.

Hand the Prompt to the AI System

The flow crosses into the AI lane:

HUMAN                          AI SYSTEM

Create prompt ----------------> Receive input
                                ↓
                                Generate output
                                ↓
Receive output <--------------- Return output

The handoff shows the responsibility change.

The human supplies input.

The AI system generates the response.

Evaluation Returns to the Human

After receiving the response, the human evaluates:

• relevance;
• accuracy concerns;
• credibility;
• missing context;
• bias;
• privacy or rule issues.

The AI does not make the final decision about whether its own response should be used.

Add a Verification Decision

A useful decision is:

Does an important claim need verification?

Conceptually:

Evaluate response
      ↓
Important claim needs evidence?
      ├─ [yes] → Verify with authoritative source
      └─ [no]  → Continue decision

The verification action belongs in the human process.

The person selects and interprets the evidence.

Model Accept, Revise, Verify, or Reject

A simplified decision section could be:

Evaluate output
      ↓
Need verification?
      ├─ [yes] → Verify claim → Re-evaluate
      └─ [no]
            ↓
Usable as-is?
      ├─ [yes] → Accept
      └─ [no]
            ↓
Can focused revision make it usable?
      ├─ [yes] → Revise
      └─ [no]  → Reject

The exact diagram can vary.

The important requirement is that the human judgment remains visible.

Revision Can Create Another AI Interaction

If the human chooses to revise the prompt, the flow may return to the AI system.

For example:

Revise prompt
      ↓
AI receives new prompt
      ↓
AI generates new output
      ↓
Human evaluates again

This forms an iteration loop.

The loop should not imply:

Keep asking until the AI says something you like.

Each new output still needs evaluation.

Verification May Use a Source Outside the AI System

If the AI makes a factual claim, the human may consult an authoritative source.

The diagram can include an action such as:

Verify claim with authoritative source

You do not need to create a separate swimlane for every possible website or document unless the purpose of the model requires it.

Keep the current diagram focused on AI and human responsibility.

End With a Human Decision

A strong model should end with an action owned by the human.

For example:

Use revised output

or:

Reject output

Then:

End

This preserves the central principle:

AI generates; the human decides.

Read the Diagram as a Responsibility Story

A complete explanation might be:

The human defines the task, minimizes the input, and sends a prompt. The AI system generates a response. The human evaluates the result, verifies important claims when needed, and decides whether to accept, revise, or reject the output.

If the diagram communicates that story clearly, it is representing responsible AI use rather than only the technical prompt/response exchange.