Session 1 slides

Session 1 slides

Responsible AI foundations

Humanitarian AI Training Kit | Session 1 | Level 1

ParalleX Labs Inc. | Learning materials only

Ground rules

  • Use synthetic and public examples only
  • Do not share real beneficiary or case data in chat
  • Camera and microphone optional
  • Participation by voice or typed chat
  • You remain accountable for decisions

Diagnostic (4 questions)

  1. Name one AI tool you have used or heard about at work.
  2. What is one concern about AI in humanitarian action?
  3. Who approves new digital tools in your organization?
  4. What is one task you might support with AI?

What is generative AI?

  • Creates text, images, code, or audio from prompts
  • Does not guarantee truth
  • Can sound confident when wrong
  • Examples: drafting, summarizing, translating

Not the same as every “AI” label on software.

Prediction and automation

Type Does Humanitarian example
Prediction Scores risk from patterns Flood exposure mapping
Automation Runs fixed rules Auto-routing forms
Generative Produces new content Draft report from notes

All require governance. Generative tools need extra verification.

Accountability

Question: If AI supports a decision, who is accountable?

  • The vendor? No
  • The model? No
  • A named person in your organization? Yes

Human approval, override, and remedy are required for consequential uses.

Protection lens

Before the tool, ask:

  • Who benefits? Who may be excluded?
  • Did affected people and local partners shape the task?
  • How can someone challenge or correct the result?

Fundamental Principles: humanity, impartiality, neutrality, independence.

Worked example: the decision

Task: Review a draft evacuation SMS. You may not send it.

Source W1 (10:00): Reception Centre A open. Only Entrance E is step-free. No road route confirmed.

Draft: “Road B open. All entrances step-free. Go now.”

Worked example: source check

Statement W1 Mark
Road B open No route info Not established
All entrances step-free Only E Contradicted
Go now No send authority Not established

Seven-step screen (overview)

  1. Purpose and alternative
  2. People and partners
  3. Data
  4. Tool and vendor
  5. Verification
  6. Human control
  7. Monitoring and stop

Decision: Green / Amber / Red + owner + review date

Group exercise

Case: Wildfire evacuation SMS (synthetic)

12 minutes in groups

Apply the seven-step screen

Prepare: one protection issue + G/A/R decision

Debrief prompts

  • What was contradicted in the draft?
  • Who could be harmed?
  • Who approves a corrected message?
  • What manual route remains if the tool fails?

One rule to remember

Stop if the source is unverified or a human approval point is missing.

Your action plan

Write one safe action before Session 2:

  • Apply the screen to a fictional task
  • Read your organization’s AI or data policy
  • Ask your manager about approved tools

Optional: 10-minute practice between sessions