---
title: "AI Mistakes to Watch For Before You Trust the Answer"
date: "2026-06-15"
author: "Graham"
description: "A practical checklist of AI mistakes: hallucinations, stale facts, missing context, fake citations, overconfidence, privacy risk, and wrong-tool problems."
tags: ["AI Safety", "Power of AI"]
url: "https://powerofai.ca/ai-mistakes-to-watch-for"
readTime: "8 min"
---

# AI Mistakes to Watch For Before You Trust the Answer

The right attitude toward AI is not blind trust and not cynical dismissal. It is useful skepticism.

AI can be incredibly helpful and still be wrong. The skill is learning where it tends to fail, then building a simple check before you use the output.

## Quick picks

- **Most common mistake: Confident wrong answers.** The answer can sound polished even when the facts are off.
- **Most avoidable mistake: Not providing context.** If the model does not have the document, quote, policy, or file, it may fill gaps from general patterns.
- **Most dangerous mistake: Using AI as final authority.** Legal, medical, financial, safety, and employment decisions need human and professional review.
- **Best habit: Ask what it is assuming.** Assumptions reveal where the answer is weak.

## Hallucinations

A hallucination is a confident wrong answer. It can be a fake fact, fake citation, wrong date, invented feature, made-up quote, or explanation that sounds right but is not.

The fix is to ground the answer. Paste the source, upload the file, ask for citations from provided material, or use a search-connected workflow when the question depends on current facts.

## Stale facts

AI systems may answer from training data, memory, or old assumptions unless they can search or you provide current material. This matters for prices, laws, policies, model releases, sports, companies, products, and anything changing quickly.

When freshness matters, ask: "What date is this based on? Do you need to search? What could have changed?"

## Missing context

A model trained on the internet still does not know your customer, your policy, your assignment, your codebase, your invoice, your business rules, or your voice unless you give it that context.

This is why many people get generic answers. They ask generic questions.

## Privacy and sensitive data

Before you paste private information, ask whether the AI really needs it. Remove names, account numbers, addresses, health details, financial identifiers, student information, secrets, and passwords unless you understand the tool and the risk.

Use source-grounded tools and enterprise settings where appropriate, but still keep the habit: share the minimum context needed.

## Copyable prompts

### AI answer audit prompt

```text
Audit your last answer before I use it. List possible hallucinations, stale facts, missing context, assumptions, privacy risks, and parts that need human or professional verification.
```

### Source-grounding prompt

```text
Answer using only the source material I provide. If the answer is not in the source, say "not in the source." Source material: [SOURCE]. Question: [QUESTION].
```

### Uncertainty prompt

```text
Separate your answer into: what seems clear, what is uncertain, what you are assuming, what could change the answer, and what I should verify before acting.
```

## Related Power of AI pages

- [How to Question AI](/guide/questioning-ai): The deeper verification guide.
- [Use AI For This](/use-ai-for-this): Apply the checklist to practical playbooks.
- [AI Glossary](/ai-glossary): Definitions for hallucination, grounding, RAG, context, and guardrails.
- [What Is an AI Agent?](/what-is-an-ai-agent): Agents make verification even more important.

## Sources and official references

- [Google people-first content guidance](https://developers.google.com/search/docs/fundamentals/creating-helpful-content)
- [OpenAI ChatGPT capabilities overview](https://help.openai.com/en/articles/9260256-chatgpt-capabilities-overview)
- [Anthropic Claude Constitution](https://www.anthropic.com/constitution)