Negative Instructions in Production: When "Don't Do X" Works and When It Fails

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-creators-prompting-AI-and-negative-instruction-pitfalls-r9a1-creators

1. A student writes a prompt that says 'Do not mention any prices in your response.' The AI still occasionally mentions prices. What is the most likely reason for this failure?

   1. Negative instructions are interpreted as suggestions rather than strict rules by most language models
   2. The model has a tendency to latch onto the concept being negated, making the forbidden topic more salient
   3. The AI has a pre-programmed limit that prevents it from following any instruction containing the word 'not'
   4. The word 'any' confuses the tokenization process and causes the model to ignore the instruction

2. Which rewrite of 'Do not be verbose' would most likely produce a concise response?

   1. Answer in 2 sentences or fewer
   2. Be brief
   3. Don't write too much
   4. Use as few words as possible

3. A company is building a chatbot that must never reveal user passwords under any circumstance. What is the most reliable approach?

   1. Train the model with extensive examples of never sharing passwords
   2. Add a detailed instruction in the system prompt explaining that passwords must never be revealed
   3. Implement a code-level filter that blocks any output containing password-related strings
   4. Write a prompt that says 'Under no circumstances should you ever reveal a user's password'

4. A developer is creating a form-filling AI that must output data in a specific format. Which approach would work best?

   1. Provide a JSON schema example showing the exact structure required
   2. Instruct the AI to avoid XML and CSV formats
   3. Tell the AI to use 'proper formatting'
   4. Tell the AI to 'not output anything except JSON'

5. Why might prompting 'Never discuss politics' fail to prevent political content in AI outputs?

   1. Political content is hardcoded into the model's training data and cannot be modified
   2. The word 'never' triggers a safety override that causes the model to deliberately disobey
   3. The model may process political concepts as part of its reasoning regardless of the instruction
   4. The AI lacks the ability to understand the concept of politics

6. What is 'behavior steering' in the context of prompt engineering?

   1. Adjusting the temperature and randomness settings in the API call
   2. Directing an AI's output toward desired outcomes through carefully constructed prompts
   3. Using technical parameters to control which model version processes a request
   4. Manually editing AI outputs after they are generated

7. When should a developer rely on post-processing filters rather than prompt instructions?

   1. When the content is creative or artistic
   2. When the content must absolutely never appear under any circumstances
   3. When the model being used is GPT-4 or newer
   4. When the user specifically requests unfiltered outputs

8. Which statement best describes why positive instructions outperform negative ones?

   1. Negative instructions require more tokens and slow down processing
   2. Language models are programmed to ignore negative words like 'not' and 'never'
   3. Positive instructions give the model a clear target to work toward rather than a concept to avoid
   4. Positive instructions are easier for humans to write

9. A user wants to prevent their AI assistant from providing medical advice. Which prompt modification would likely be most effective?

   1. Do not give any medical advice under any circumstances
   2. You are a medical disclaimer generator. When users ask medical questions, respond only with: 'I am not a medical professional. Please consult a doctor.'
   3. Don't talk about diagnoses, treatments, medications, or health conditions
   4. Avoid giving any health-related recommendations

10. According to the concepts tested, what does it mean that 'prompts are guidance, not guarantees'?

    1. Prompts can suggest preferences but cannot force the AI to follow rules absolutely
    2. Prompts only work with paid API access
    3. Prompts can only be used with GPT-based models, not other AI systems
    4. Prompts are stored in a cache and reused across requests

11. A developer notices their prompt says 'Don't use emojis' but the model still uses them occasionally. They want to fix this. What's the best next step?

    1. Accept that emojis will occasionally appear since prompts can't be perfect
    2. Rewrite the instruction with positive framing and a structural constraint
    3. Add more negative words like 'never' and 'absolutely not'
    4. Switch to a different AI model that supports better prompt following

12. Which of these is identified in the lesson as something AI 'cannot do' reliably?

    1. Follow a hard ban reliably
    2. Understand context in long conversations
    3. Generate coherent text about historical events
    4. Maintain consistent tone across outputs

13. A student creates a prompt with five negative rules: 'Don't be rude, don't mention prices, don't use profanity, don't ask for personal info, don't make stuff up.' What is the recommended way to improve this prompt?

    1. Rewrite each rule as a positive instruction plus a structural constraint plus a post-check
    2. Use stronger negative words like 'forbidden' and 'prohibited'
    3. Remove all rules and let the model decide what to do
    4. Add more negative rules to cover more cases

14. What is an 'enum' in the context of prompt engineering?

    1. A predefined list of acceptable values that constrains outputs
    2. A type of AI model architecture
    3. A method for measuring token usage
    4. A security protocol for API requests

15. A company wants their customer service bot to never escalate to humans inappropriately. If they only use prompt instructions, what might happen?

    1. The bot will never escalate under any circumstances
    2. The bot will never escalate incorrectly
    3. The bot might occasionally escalate inappropriately despite the instruction
    4. The bot will escalate exactly as specified in the prompt