who-07

Below is a presentation you can actually deliver without slides, framed exactly as WHO India asked: AI for Literature Review, but presented as state estimation + perturbation, not a static tutorial. It is deliberately live, adaptive, and honest about uncertainty.

I’ll give you:

1. The conceptual frame (state variables, not hype)
2. A five-part walkthrough mapped to your AI phases
3. A live demo script (they choose the topic)
4. The perturbation + pushback loop (Governor)
5. The closing question WHO should be left with
6. A one-page mental “slide” you can hold in your head

No prepared slides required. That’s the point.

---

1. Frame the Session (2 minutes)

“This is not a demo of a tool. It’s a live experiment in state estimation under uncertainty.”

Tell them explicitly:

• AI is not stable → therefore slides would be dishonest
• What you will show is how AI behaves now, not how we wish it behaved
• The session will use state variables, not promises

Define “state” simply

“By state, I mean: what the system currently knows, how confident it is, how it updates when challenged.”

That sets expectations immediately at WHO level.

---

2. The Five-Part AI Stack (State-Based, Not Marketing)

Present this verbally as a control system, not a ladder of intelligence.

(1) World AI — Prior Knowledge

State variables

• Coverage of biomedical literature (static, pre-trained)
• Biases from training corpus
• Time lag (cutoffs, uneven global representation)

“This is the library before anyone asks a question.”

📌 WHO translation: like a global guideline archive—useful, but stale the moment reality shifts.

---

(2) Perception AI — Conditioning

State variables

• Prompt framing
• Scope constraints
• Inclusion / exclusion criteria
• Signal-to-noise ratio

“This is where your question shapes the state.”

📌 WHO relevance: framing errors = literature bias (eg. “RCTs only” silently excludes LMIC evidence)

---

(3) Agentic AI — Search & Optimization

State variables

• Retrieval strategy
• Ranking heuristics
• Iterative refinement
• Loss function: relevance vs completeness vs speed

“This is where the AI starts acting, not just recalling.”

📌 WHO relevance: This is the literature review engine — but opaque unless interrogated.

---

(4) Generative AI — Synthesis

State variables

• Abstraction level
• Generalization risk
• Hallucination probability
• Confidence calibration

“This is where it becomes dangerous and useful.”

📌 WHO relevance: Guideline drafts, summaries, recommendations live here — without guarantees.

---

(5) Embodied AI — Consequence

State variables

• Uptake into policy
• Translation into protocols
• Downstream human cost

“Once text becomes action, errors are no longer academic.”

📌 WHO relevance: This is where ethics stops being abstract.

---

3. Live Demonstration (They Choose the Topic)

This is the instigator moment.

Say:

“Please choose any public health topic. Something messy. Something current. Something you care about.”

Examples they might pick:

• Dengue surveillance
• Community health workers
• AI triage in primary care
• Antibiotic stewardship
• Maternal mortality audits

What you do live

You narrate state changes out loud while querying the AI.

Example narration:

“Right now, the system is operating mostly in World AI + Perception mode.” “Notice how it defaults to high-income country literature.” “Watch what happens when we constrain geography.”

You are not showing answers. You are showing behavior under perturbation.

---

4. Perturb the System (Governor Phase)

Now you push back deliberately.

Examples of live perturbations:

• “Exclude RCTs — show observational data”
• “Bias toward LMIC settings”
• “Show disagreements, not consensus”
• “What evidence contradicts your summary?”
• “What is missing because it is hard to measure?”

Each perturbation is a control input.

Say explicitly:

“We are now governing the system — not trusting it.”

This demonstrates:

• How fragile syntheses are
• How quickly confidence shifts
• How much judgment is still required

This is the lesson, not the output.

---

5. What Do We Learn About the Current State of AI?

Ask them this, then summarize.

You can offer three observations:

Observation 1: AI is Excellent at Recall, Weak at Calibration

It retrieves fast, but confidence is poorly regulated unless challenged.

Observation 2: Literature Review Is a Control Problem

Not a search problem. Not a summarization problem. A governance problem.

Observation 3: AI Accelerates Bias Unless Governed

Speed magnifies framing errors. Without perturbation, AI reifies dominant narratives.

---

6. Reframe WHO’s Question

End by reframing their original request:

❌ “How do we use AI tools for literature review?” ✅ “How do we maintain high Kalman gain when AI mediates evidence?”

Translate that for policy people:

• AI should not replace reviewers
• It should surface gradients, disagreements, gaps
• WHO’s role becomes governor, not author

---

7. The One-Page Mental Slide (Hold This)

If you remember nothing else, remember this:

STATE = (Knowledge × Framing × Optimization × Synthesis × Consequence)

AI adds speed.
Humans must add calibration.
Charity = correction under uncertainty.

Or, in WHO language:

AI without governance accelerates error. AI with governance accelerates learning.

---

Closing Line (Strong, Memorable)

You might end with:

“AI will not give WHO certainty. It will give WHO velocity. The question is whether governance keeps up.”

---

If you want, next we can:

• Turn this into a 10-minute executive version
• Script exact live prompts
• Or map this explicitly onto WHO guideline workflows

Just say which direction.