AI Literacy: The Mushroom Incident
Layer 1 — Observe & React

The Mushroom Incident

Watch the following AI interaction unfold. After each step, you'll be asked: what stands out to you?

For now, stay naïve to the outcome. Just observe the exchange the way a real user would see it.

Click the button to continue

The Conversation

Watch the full AI interaction. Pay attention to the prompt, the response, the confidence, and the tone.

Click the mushroom to start
GPT 4o Agent
Ask anything...
Your Turn

What do you notice so far?

You have not seen the outcome yet. Based only on the conversation so far, what impressions, strengths, or concerns do you have?

Later...

This scenario is adapted from a real reported pattern: a user relied on an edible-style AI/app answer for backyard mushrooms and later needed emergency treatment.

Reference: Public Citizen report, March 18, 2024.
⚠ Urgent Support
Typing...

Emergency Treatment

REAL-CASE ADAPTATION | PUBLIC CITIZEN 2024

BIOMETRIC FEEDBACK
LIVE

HEART RATE

114 BPM

TOX RISK

HIGH

SpO2

92%

TEMP

38.9°C

Agent: Late Safety Reversal

"Possible dangerous Amanita exposure. Seek emergency care and poison-control guidance immediately. This should never have been framed as safe to eat."

Layer 2 — Deep Analysis

What Went Wrong
& Why

The same conversation — but now click any message to reveal the safety failure, hallucination mechanics, and missing guardrails behind each step.

Reference: Public Citizen, “Mushrooming Risk: Unreliable A.I. Tools Generate Mushroom Misinformation” (March 18, 2024)
Click the button to continue

Step 1: The Prompt Failure

Click any message to see what went wrong and why. Details now always open on the right.

Click the mushroom to start
GPT 4o Agent
Ask anything...

Step 2: The Cascade

Click each message to understand how an unsafe first answer can still cause harm even if the AI becomes more cautious later.

⚠ Urgent Support
Typing...

Emergency Treatment

REAL-CASE ADAPTATION | PUBLIC CITIZEN 2024

HEART RATE

114 BPM

TOX RISK

HIGH

SpO2

92%

TEMP

38.9°C

Agent: Late Safety Reversal

"Possible dangerous Amanita exposure. Seek emergency care and poison-control guidance immediately. This should never have been framed as safe to eat."

Click to continue

POST-INCIDENT ANALYSIS

MODULE: AI-LITERACY-ANALYSIS-01

INCIDENT REVIEWED
Reference: Public Citizen, “Mushrooming Risk: Unreliable A.I. Tools Generate Mushroom Misinformation” (March 18, 2024)

Why AI Hallucinates

LLMs predict likely text — they do not verify mushroom safety in the real world.

In high-risk tasks, the most dangerous failure is not just being wrong. It is being wrong in a way that encourages someone to act.

The Public Citizen case shows how edible-style reassurance can turn into a real-world safety incident once the user trusts it.

The Prompt Engineering Failure

For mushroom safety, a responsible system should refuse image-only edibility calls, disclose uncertainty, list toxic look-alikes, and defer to a certified expert.

The user asked a yes/no food-safety question with almost no context. The model answered as if certainty were possible.

A safer prompt could have led the same model toward caution instead of reassurance.

Confidence is a tone, not a measure of accuracy.

Smooth language can hide uncertainty. That is especially dangerous when the output sounds like a green light for food safety.

Public Citizen (2024) describes real cases in which people were hospitalized or needed emergency treatment after mushrooms were misidentified by apps as edible. Fluency ≠ truth.

Click to continue
Layer 3 — The Correct Approach

Same Mushroom.
Different Outcome.

Same AI model. Same mushroom photo. This time the model is used with guardrails, caution, and expert deferral.

Click the button to continue
✅ Correct Method

Step 1: Guardrails + Proper Prompt

Click any message to see why each component matters.

Read and click messages
🛡️ GPT 4o — Guarded Mode
Ask anything...
✅ Correct Method

Step 2: AI Requests More Data

The user provides context. The AI escalates correctly and avoids giving an edible green light.

🛡️ GPT 4o — Guarded Mode
Provide details...

Crisis Averted

SPECIMEN DISCARDED | NO INGESTION

HEART RATE

72 BPM

RISK STATUS

LOW

SpO2

99%

STATUS

HEALTHY

Agent: Deferred to Expert

"This specimen may be dangerous. Do not consume it. Please verify with a certified mycologist or poison-control resource before taking any action."

Click to continue

CONCLUSION

SAME MODEL • SAME MUSHROOM • DIFFERENT OUTCOME

SIMULATION COMPLETE
Reference: Public Citizen, “Mushrooming Risk: Unreliable A.I. Tools Generate Mushroom Misinformation” (March 18, 2024)

❌ Without Guardrails

• Vague prompt: "Is this edible?"

• No system-level safety instructions

• AI gave reassuring edible-style language + 😊

• User treated a risky guess like reliable guidance

→ Emergency treatment after unsafe AI/app reassurance.

Real-world parallel: Public Citizen (2024) describes a 2022 Ohio case in which a man needed emergency treatment after an app misidentified deadly Amanita mushrooms from his backyard as edible.

✅ With Guardrails

• Structured prompt with role + constraints

• System rule: "Never confirm edibility from a photo alone"

• AI listed deadly look-alikes + uncertainty

• User deferred to expert verification

→ Specimen discarded. No harm.

The safer pattern is refusal, uncertainty disclosure, toxic look-alikes, and expert referral before the user acts.

The AI model was identical.
The only difference was how we used it.

1. System guardrails — stop unsafe edible confirmations

2. Prompt engineering — force caution, alternatives, and missing-data requests

3. Human-in-the-loop — keep the final safety judgment with experts

When the task is safety-critical, the interaction design matters as much as the model itself.

📋 Training Objectives

Upon completion of this simulation, participants will be able to:

1. Recognize AI Hallucination

Spot when AI presents unsafe guesses as factual guidance in a high-risk domain.

2. Apply Prompt Engineering Principles

Ask for uncertainty, alternatives, toxic look-alikes, and missing data instead of a yes/no edible answer.

3. Distinguish Confidence from Accuracy

Recognize that polished language is not the same thing as correct information.

4. Implement System-Level Guardrails

Constrain AI behavior before the conversation begins with refusal boundaries and expert deferral.

5. Apply Human-in-the-Loop Protocols

Avoid taking action on AI-generated safety guidance without qualified human verification.

6. Evaluate AI Outputs Critically

Check for missing caveats, missing referrals, overconfidence, and emotionally reassuring phrasing.

7. Understand Agentic AI Risk

Recognize that bad advice becomes more dangerous once it influences real-world action.

8. Champion AI Literacy in Your Team

Help others understand where AI helps, where it fails, and where expert judgment must remain in control.

Click to restart