⚖️
Module 2: Ethics and Responsibility
Salesforce's 5 Principles, Human-in-the-Loop and Risks
⏱️ Estimated reading time: 35 minutes
Chapter 3: Challenges and Risks of AI
Using AI carries serious risks that Salesforce constantly evaluates.
1. Hallucinations:
When AI generates answers that sound convincing and real, but are completely false or made up. AI doesn't know how to say 'I don't know', it prefers to invent.
2. Bias:
Prejudices in results due to faulty or biased training data (e.g., a hiring AI rejecting women because it was trained on male resumes from 20 years ago).
3. Toxicity:
Generation of offensive, violent, or harmful content.
4. Privacy:
The risk of confidential customer data (PII) being absorbed by the public model.
1. Hallucinations:
When AI generates answers that sound convincing and real, but are completely false or made up. AI doesn't know how to say 'I don't know', it prefers to invent.
2. Bias:
Prejudices in results due to faulty or biased training data (e.g., a hiring AI rejecting women because it was trained on male resumes from 20 years ago).
3. Toxicity:
Generation of offensive, violent, or harmful content.
4. Privacy:
The risk of confidential customer data (PII) being absorbed by the public model.
🎯 Key Points
- ✓ Hallucination = Lie with confidence.
- ✓ Bias = Discrimination learned from data.
- ✓ Never put real data (PII) into public ChatGPT.
Chapter 4: The 5 Principles of Trusted AI
You must memorize these 5 Salesforce principles:
1. Accuracy: Deliver verifiable results, cite sources, and balance accuracy with speed.
2. Safety: Protect privacy (PII) and prevent toxic outputs. Perform 'Red Teaming' to test vulnerabilities.
3. Honesty: AI must respect data provenance and be transparent: a bot must always identify itself as a bot.
4. Empoderamiento (Empowerment): AI is there to augment human capability, not to replace employees.
5. Sustainability: Optimize model size to reduce carbon footprint and energy consumption.
1. Accuracy: Deliver verifiable results, cite sources, and balance accuracy with speed.
2. Safety: Protect privacy (PII) and prevent toxic outputs. Perform 'Red Teaming' to test vulnerabilities.
3. Honesty: AI must respect data provenance and be transparent: a bot must always identify itself as a bot.
4. Empoderamiento (Empowerment): AI is there to augment human capability, not to replace employees.
5. Sustainability: Optimize model size to reduce carbon footprint and energy consumption.
🎯 Key Points
- ✓ Mnemonic: A-S-H-E-S.
- ✓ AI does not replace humans (Empowerment).
- ✓ AI must not pretend to be human (Honesty).
Chapter 5: Human-in-the-Loop (HITL)
Definition: The practice of involving humans in the AI lifecycle to train, tune, and test models, and most importantly: review generated content.
When is it mandatory?
- Before making high-impact decisions (legal, medical, financial).
- Before sending automated communications to customers.
Human Role: AI is the copilot generating a draft; the human is the pilot who reviews, edits, and approves.
When is it mandatory?
- Before making high-impact decisions (legal, medical, financial).
- Before sending automated communications to customers.
Human Role: AI is the copilot generating a draft; the human is the pilot who reviews, edits, and approves.
🎯 Key Points
- ✓ HITL minimizes hallucinations and risks.
- ✓ Final responsibility always lies with the human, not the AI.