🛡️ Kitchen Guardrails

Practical AI safety — not corporate policy. How to use AI responsibly without overthinking it.

8 guardrails

🎯 Accuracy

How to Avoid AI Hallucinations

🍳 Checking the chef's claim before serving it to guests

Read the guardrail

The Problem

AI tools are confident by nature — they generate plausible-sounding text whether or not the underlying claim is accurate. This guardrail helps you work with AI in a way that catches errors before they cause problems.

Practical Techniques

  1. Verify specific facts independently — any statistic, date, name, or quote that matters should be checked in a primary source
  2. Ask for sources — even though AI may fabricate citations, asking “what’s the source for that?” often reveals when it’s uncertain
  3. Try the same question differently — if you get different answers to the same question phrased differently, that inconsistency signals uncertainty
  4. Ask it to identify its own uncertainty — “What are you least confident about in that answer?” often surfaces genuine gaps
  5. Cross-reference important claims — for anything you’ll share or act on, run a quick search

Red Flags to Watch For

  • Specific statistics with no sourcing
  • Historical claims that sound plausible but specific
  • Named experts or studies that seem hard to verify
  • Confident answers to questions that are genuinely contested
  • Details that are too perfect or too neat

Kitchen Tip

Treat AI output like a first draft from a brilliant but fallible assistant — not a final product. The AI is excellent at synthesis and structure; humans are still needed for verification and judgment.

🤝 Trust

Verify Before You Trust

🍳 Tasting the dish before it leaves the kitchen

Read the guardrail

The Problem

AI tools are confident by nature — they generate plausible-sounding text whether or not the underlying claim is accurate. This guardrail helps you work with AI in a way that catches errors before they cause problems.

Practical Techniques

  1. Verify specific facts independently — any statistic, date, name, or quote that matters should be checked in a primary source
  2. Ask for sources — even though AI may fabricate citations, asking “what’s the source for that?” often reveals when it’s uncertain
  3. Try the same question differently — if you get different answers to the same question phrased differently, that inconsistency signals uncertainty
  4. Ask it to identify its own uncertainty — “What are you least confident about in that answer?” often surfaces genuine gaps
  5. Cross-reference important claims — for anything you’ll share or act on, run a quick search

Red Flags to Watch For

  • Specific statistics with no sourcing
  • Historical claims that sound plausible but specific
  • Named experts or studies that seem hard to verify
  • Confident answers to questions that are genuinely contested
  • Details that are too perfect or too neat

Kitchen Tip

Treat AI output like a first draft from a brilliant but fallible assistant — not a final product. The AI is excellent at synthesis and structure; humans are still needed for verification and judgment.

🤝 Trust

Knowing When Not to Trust AI

🍳 Knowing which dishes the chef is actually trained to make

Read the guardrail

The Problem

AI tools are confident by nature — they generate plausible-sounding text whether or not the underlying claim is accurate. This guardrail helps you work with AI in a way that catches errors before they cause problems.

Practical Techniques

  1. Verify specific facts independently — any statistic, date, name, or quote that matters should be checked in a primary source
  2. Ask for sources — even though AI may fabricate citations, asking “what’s the source for that?” often reveals when it’s uncertain
  3. Try the same question differently — if you get different answers to the same question phrased differently, that inconsistency signals uncertainty
  4. Ask it to identify its own uncertainty — “What are you least confident about in that answer?” often surfaces genuine gaps
  5. Cross-reference important claims — for anything you’ll share or act on, run a quick search

Red Flags to Watch For

  • Specific statistics with no sourcing
  • Historical claims that sound plausible but specific
  • Named experts or studies that seem hard to verify
  • Confident answers to questions that are genuinely contested
  • Details that are too perfect or too neat

Kitchen Tip

Treat AI output like a first draft from a brilliant but fallible assistant — not a final product. The AI is excellent at synthesis and structure; humans are still needed for verification and judgment.

🔒 Privacy

Protecting Your Privacy with AI

🍳 Not putting your secret recipe in the window display

Read the guardrail

The Problem

AI tools are confident by nature — they generate plausible-sounding text whether or not the underlying claim is accurate. This guardrail helps you work with AI in a way that catches errors before they cause problems.

Practical Techniques

  1. Verify specific facts independently — any statistic, date, name, or quote that matters should be checked in a primary source
  2. Ask for sources — even though AI may fabricate citations, asking “what’s the source for that?” often reveals when it’s uncertain
  3. Try the same question differently — if you get different answers to the same question phrased differently, that inconsistency signals uncertainty
  4. Ask it to identify its own uncertainty — “What are you least confident about in that answer?” often surfaces genuine gaps
  5. Cross-reference important claims — for anything you’ll share or act on, run a quick search

Red Flags to Watch For

  • Specific statistics with no sourcing
  • Historical claims that sound plausible but specific
  • Named experts or studies that seem hard to verify
  • Confident answers to questions that are genuinely contested
  • Details that are too perfect or too neat

Kitchen Tip

Treat AI output like a first draft from a brilliant but fallible assistant — not a final product. The AI is excellent at synthesis and structure; humans are still needed for verification and judgment.

⚖️ Bias

Spotting AI Bias

🍳 Noticing when the menu only features one type of cuisine

Read the guardrail

The Problem

AI tools are confident by nature — they generate plausible-sounding text whether or not the underlying claim is accurate. This guardrail helps you work with AI in a way that catches errors before they cause problems.

Practical Techniques

  1. Verify specific facts independently — any statistic, date, name, or quote that matters should be checked in a primary source
  2. Ask for sources — even though AI may fabricate citations, asking “what’s the source for that?” often reveals when it’s uncertain
  3. Try the same question differently — if you get different answers to the same question phrased differently, that inconsistency signals uncertainty
  4. Ask it to identify its own uncertainty — “What are you least confident about in that answer?” often surfaces genuine gaps
  5. Cross-reference important claims — for anything you’ll share or act on, run a quick search

Red Flags to Watch For

  • Specific statistics with no sourcing
  • Historical claims that sound plausible but specific
  • Named experts or studies that seem hard to verify
  • Confident answers to questions that are genuinely contested
  • Details that are too perfect or too neat

Kitchen Tip

Treat AI output like a first draft from a brilliant but fallible assistant — not a final product. The AI is excellent at synthesis and structure; humans are still needed for verification and judgment.

🛡️ Safety

Handling Sensitive Topics

🍳 Knowing which dishes require extra care in the kitchen

Read the guardrail

The Problem

AI tools are confident by nature — they generate plausible-sounding text whether or not the underlying claim is accurate. This guardrail helps you work with AI in a way that catches errors before they cause problems.

Practical Techniques

  1. Verify specific facts independently — any statistic, date, name, or quote that matters should be checked in a primary source
  2. Ask for sources — even though AI may fabricate citations, asking “what’s the source for that?” often reveals when it’s uncertain
  3. Try the same question differently — if you get different answers to the same question phrased differently, that inconsistency signals uncertainty
  4. Ask it to identify its own uncertainty — “What are you least confident about in that answer?” often surfaces genuine gaps
  5. Cross-reference important claims — for anything you’ll share or act on, run a quick search

Red Flags to Watch For

  • Specific statistics with no sourcing
  • Historical claims that sound plausible but specific
  • Named experts or studies that seem hard to verify
  • Confident answers to questions that are genuinely contested
  • Details that are too perfect or too neat

Kitchen Tip

Treat AI output like a first draft from a brilliant but fallible assistant — not a final product. The AI is excellent at synthesis and structure; humans are still needed for verification and judgment.

🤝 Trust

When (and When Not) to Trust AI

🍳 Knowing your sous-chef's actual training vs. what they claim to know

Read the guardrail

The Problem

AI tools are confident by nature — they generate plausible-sounding text whether or not the underlying claim is accurate. This guardrail helps you work with AI in a way that catches errors before they cause problems.

Practical Techniques

  1. Verify specific facts independently — any statistic, date, name, or quote that matters should be checked in a primary source
  2. Ask for sources — even though AI may fabricate citations, asking “what’s the source for that?” often reveals when it’s uncertain
  3. Try the same question differently — if you get different answers to the same question phrased differently, that inconsistency signals uncertainty
  4. Ask it to identify its own uncertainty — “What are you least confident about in that answer?” often surfaces genuine gaps
  5. Cross-reference important claims — for anything you’ll share or act on, run a quick search

Red Flags to Watch For

  • Specific statistics with no sourcing
  • Historical claims that sound plausible but specific
  • Named experts or studies that seem hard to verify
  • Confident answers to questions that are genuinely contested
  • Details that are too perfect or too neat

Kitchen Tip

Treat AI output like a first draft from a brilliant but fallible assistant — not a final product. The AI is excellent at synthesis and structure; humans are still needed for verification and judgment.

🎯 Accuracy

Fact-Checking Creative Work

🍳 Making sure the garnish matches what's on the menu

Read the guardrail

The Problem

AI tools are confident by nature — they generate plausible-sounding text whether or not the underlying claim is accurate. This guardrail helps you work with AI in a way that catches errors before they cause problems.

Practical Techniques

  1. Verify specific facts independently — any statistic, date, name, or quote that matters should be checked in a primary source
  2. Ask for sources — even though AI may fabricate citations, asking “what’s the source for that?” often reveals when it’s uncertain
  3. Try the same question differently — if you get different answers to the same question phrased differently, that inconsistency signals uncertainty
  4. Ask it to identify its own uncertainty — “What are you least confident about in that answer?” often surfaces genuine gaps
  5. Cross-reference important claims — for anything you’ll share or act on, run a quick search

Red Flags to Watch For

  • Specific statistics with no sourcing
  • Historical claims that sound plausible but specific
  • Named experts or studies that seem hard to verify
  • Confident answers to questions that are genuinely contested
  • Details that are too perfect or too neat

Kitchen Tip

Treat AI output like a first draft from a brilliant but fallible assistant — not a final product. The AI is excellent at synthesis and structure; humans are still needed for verification and judgment.

Further Reading

AI Safety
Google's AI Overviews Can Scam You. Here's How to Stay Safe Wired

AI-generated search summaries are being exploited to surface scam sites and misinformation at scale. How the attack works and practical steps to protect yourself.
Recipe
Make AI Check Its Work The Prompt Kitchen

A practical recipe for forcing AI to verify its own output before you trust it — layered fact-checking built into the prompt itself.
Security
OWASP LLM Top 10: Prompt Injection & Agent Safety OWASP

The definitive security reference for LLM applications — prompt injection, insecure output handling, training data poisoning, and the 7 other attack vectors every AI builder needs to understand.