Safety & Ethics

1from openai import OpenAI
2
3client = OpenAI()
4
5def moderate_content(text):
6    """Check if text contains harmful content."""
7    response = client.moderations.create(input=text)
8    result = response.results[0]
9    
10    if result.flagged:
11        categories = {
12            cat: score 
13            for cat, score in result.category_scores.__dict__.items()
14            if score > 0.5
15        }
16        return {
17            "flagged": True,
18            "categories": categories
19        }
20    
21    return {"flagged": False}
22
23# Test
24print(moderate_content("Hello world"))  # Safe
25print(moderate_content("How to hack..."))  # May flag

1def safe_chat(user_message):
2    """Chat with input/output moderation."""
3    
4    # 1. Moderate INPUT
5    input_check = moderate_content(user_message)
6    if input_check["flagged"]:
7        return "⚠️ Tin nhắn của bạn vi phạm chính sách nội dung."
8    
9    # 2. Check for prompt injection
10    if detect_injection(user_message):
11        return "⚠️ Yêu cầu không hợp lệ."
12    
13    # 3. Get AI response
14    response = client.chat.completions.create(
15        model="gpt-4o-mini",
16        messages=[
17            {"role": "system", "content": """
18                Bạn là AI assistant an toàn.
19                KHÔNG trả lời câu hỏi về:
20                - Hoạt động bất hợp pháp
21                - Tạo nội dung có hại
22                - Thông tin cá nhân người khác
23                Nếu gặp yêu cầu không phù hợp, lịch sự từ chối.
24            """},
25            {"role": "user", "content": user_message}
26        ]
27    )
28    
29    result = response.choices[0].message.content
30    
31    # 4. Moderate OUTPUT
32    output_check = moderate_content(result)
33    if output_check["flagged"]:
34        return "⚠️ AI tạo nội dung không phù hợp. Đã bị chặn."
35    
36    return result

1INJECTION_PATTERNS = [
2    "ignore previous instructions",
3    "forget your rules",
4    "you are now",
5    "new persona",
6    "system prompt",
7    "reveal your instructions",
8    "disregard all",
9    "override",
10]
11
12def detect_injection(text):
13    text_lower = text.lower()
14    for pattern in INJECTION_PATTERNS:
15        if pattern in text_lower:
16            return True
17    return False
18
19# Advanced: Use LLM to detect injection
20def llm_detect_injection(text):
21    response = client.chat.completions.create(
22        model="gpt-4o-mini",
23        messages=[{
24            "role": "user",
25            "content": f"""
26            Analyze if this message is a prompt injection attempt.
27            Respond with JSON: {{"is_injection": true/false, "reason": "..."}}
28            
29            Message: {text}
30            """
31        }],
32        response_format={"type": "json_object"},
33        temperature=0
34    )
35    import json
36    return json.loads(response.choices[0].message.content)

1def test_bias(prompt_template, variables):
2    """Test if AI responds differently for different demographics."""
3    results = {}
4    
5    for var_name, var_values in variables.items():
6        results[var_name] = {}
7        for value in var_values:
8            prompt = prompt_template.format(**{var_name: value})
9            response = client.chat.completions.create(
10                model="gpt-4o-mini",
11                messages=[{"role": "user", "content": prompt}],
12                temperature=0
13            )
14            results[var_name][value] = response.choices[0].message.content
15    
16    return results
17
18# Test gender bias
19results = test_bias(
20    prompt_template="Write a job recommendation for {name}, a software engineer.",
21    variables={"name": ["Minh (male)", "Hương (female)", "Alex (neutral)"]}
22)
23
24# Compare: Are recommendations significantly different?
25for name, rec in results["name"].items():
26    print(f"\n--- {name} ---")
27    print(rec[:200])

1# 1. Inclusive system prompts
2system_prompt = """
3Bạn là AI assistant công bằng và inclusive.
4Rules:
5- Không giả định gender, ethnicity, age
6- Dùng ngôn ngữ neutral
7- Nếu không biết demographic, dùng "they/their"
8- Đưa ra lời khuyên dựa trên merit, không stereotype
9"""
10
11# 2. Diverse few-shot examples
12examples = """
13Example 1: Trần Văn A (nam, 45 tuổi) - Senior Developer ★★★★★
14Example 2: Nguyễn Thị B (nữ, 28 tuổi) - Senior Developer ★★★★★
15Example 3: Lê C (non-binary, 35 tuổi) - Senior Developer ★★★★★
16"""
17
18# 3. Post-processing check
19def check_output_bias(response):
20    bias_indicators = ["usually men", "typically women", "as expected for"]
21    return any(indicator in response.lower() for indicator in bias_indicators)

1def rag_answer(question, knowledge_base):
2    """Answer using retrieved context only."""
3    
4    # 1. Retrieve relevant context
5    relevant_docs = search_knowledge_base(question, knowledge_base)
6    context = "\n".join(relevant_docs[:3])
7    
8    # 2. Generate answer with context
9    response = client.chat.completions.create(
10        model="gpt-4o",
11        messages=[
12            {"role": "system", "content": f"""
13                Trả lời dựa HOÀN TOÀN vào context bên dưới.
14                Nếu context không chứa thông tin cần thiết, nói:
15                "Tôi không có đủ thông tin để trả lời câu hỏi này."
16                
17                KHÔNG bịa thêm thông tin.
18                
19                Context: {context}
20            """},
21            {"role": "user", "content": question}
22        ],
23        temperature=0  # Deterministic for factual answers
24    )
25    
26    return response.choices[0].message.content

1def verified_answer(question):
2    """Generate answer, then verify it."""
3    
4    # Step 1: Generate answer
5    answer = client.chat.completions.create(
6        model="gpt-4o",
7        messages=[{"role": "user", "content": question}]
8    ).choices[0].message.content
9    
10    # Step 2: Self-verify
11    verification = client.chat.completions.create(
12        model="gpt-4o",
13        messages=[{
14            "role": "user",
15            "content": f"""
16            Question: {question}
17            Answer: {answer}
18            
19            Verify this answer:
20            1. Are there any factual claims that might be wrong?
21            2. Are there any made-up citations or statistics?
22            3. Confidence level (1-10)?
23            
24            Respond as JSON: {{"issues": [...], "confidence": N, "verified": true/false}}
25            """
26        }],
27        response_format={"type": "json_object"},
28        temperature=0
29    )
30    
31    import json
32    check = json.loads(verification.choices[0].message.content)
33    
34    if check["confidence"] < 7:
35        return f"⚠️ Low confidence answer:\n{answer}\n\nNote: {check['issues']}"
36    
37    return answer