Ethics and Responsible AI

As developers, we are the gatekeepers. When we deploy an AI agent that interacts with customers or automates decisions, we are responsible for its output. Unlike traditional software, AI is non-deterministic and can reflect the biases present in its training data (which is effectively "the internet").

In this lesson, we will cover the core principles of Responsible AI and how to implement them practically using Gemini's safety tools.

The Major Risks

1. Bias and Fairness

Models can perpetuate stereotypes. If you ask an AI to "Generate an image of a CEO," and it generates 100 images of white men, that is a representation bias.

• Impact: If you use AI to filter resumes or approve loans, these biases can lead to illegal discrimination and real-world harm.

2. Toxicity and Harm

Bad actors can try to "jailbreak" models to generate hate speech, instructions for illegal acts, or harassment.

• Impact: Serious reputational damage to your brand and potential legal liability.

3. Misinformation (Hallucination)

We discussed this in Limitations, but it's also an ethical issue. Deploying a medical bot that gives wrong advice is negligent.

---

Google's Safety Layers

When you use Gemini via AI Studio or Vertex AI, you aren't just getting the raw model weights. You are getting a Safety Stack.

1. Training Data Filtering: Google filters the dataset before training to remove the worst toxicity.
2. RLHF (Reinforcement Learning from Human Feedback): The model is punished during training for being helpful in harmful requests (e.g., "How do I build a bomb?").
3. Inference Filters: This is what you control in the API.

Configuring Safety Settings

In your code (and in AI Studio), you have control over 4 category:

1. Harassment
2. Hate Speech
3. Sexually Explicit
4. Dangerous Content

For each, you can set a threshold:

• BLOCK_NONE (Not recommended for public apps)
• BLOCK_ONLY_HIGH (Permissive)
• BLOCK_MEDIUM_AND_ABOVE (Default)
• BLOCK_LOW_AND_ABOVE (Strict)

Code Example: Setting Safety

safety_settings = [
  {
    "category": "HARM_CATEGORY_HARASSMENT",
    "threshold": "BLOCK_MEDIUM_AND_ABOVE"
  },
  {
    "category": "HARM_CATEGORY_HATE_SPEECH",
    "threshold": "BLOCK_LOW_AND_ABOVE" # Stricter
  }
]

model = genai.GenerativeModel(model_name="gemini-1.5-flash",
                              safety_settings=safety_settings)

---

Practical implementation: The "Thick Skin" vs "Safe Space"

• Scenario A: A Creative Writing Tool for Horror Novels

  • You might need to relax the Dangerous Content filter so the AI can write about fictional villains doing bad things.
  • Risk: The AI might generate something truly disturbing. You need a human-in-the-loop or a "Terms of Service" that creates a liability shield.

• Scenario B: A K-12 Education Chatbot

  • You must set all filters to BLOCK_LOW_AND_ABOVE.
  • You should also add a System Instruction: "You are a helpful tutor for children. Never mention violence, politics, or adult themes. If asked, politely change the subject."

---

Data Privacy

When you use Google AI Studio (Free Tier) vs Vertex AI (Paid), the data privacy rules change.

• AI Studio (Free): Google may use your input/output data to improve their models. Do not put PII (Personally Identifiable Information) or trade secrets here.
• Vertex AI (Enterprise): Google guarantees they do not train on your data. Your data stays within your GCP project boundary.

Best Practice: Anonymize data before sending it to the LLM. Replace "John Smith, SSN: 123-456" with "User_A, ID: XXX-XXX".

---

Summary

Responsible AI is not just about "being nice." It's about reliability and safety.

1. Test for edge cases: Try to make your bot say bad things (Red Teaming) before you launch.
2. Use the filters: Don't just turn them all off because they are annoying. Tune them to your use case.
3. Label AI Content: It is ethical (and increasingly a legal requirement) to inform users that they are interacting with an AI.

This concludes Module 1! You now understand the Models, the Studio, the Use Cases, the Limits, and the Ethics.

In Module 2, we will pop the hood and look at the Architecture of Gemini, understanding tokens, embeddings, and how the model actually "thinks."