Free AI foundations · Lesson 8 of 12
Check claims before you use them
Find unsupported additions and verify consequential facts.
Understand
A hallucination is an unsupported or incorrect generated claim presented as if it were grounded. It can be a plausible citation, a date, a calculation or a statement about an action the system never took. Confidence of tone is not a reliable measure of truth. Asking the same model whether it is correct is not an independent check.
Build a claim list from the answer. For each important claim, record the source, publication or retrieval date and what you actually checked. Prefer primary evidence for official rules, prices or product features. Open a cited link and confirm it supports the claim; a real link can still be irrelevant. When a claim cannot be verified, remove it or mark it uncertain.
Choose the depth of review based on consequences. A harmless brainstorming idea needs different scrutiny from a customer promise or a consequential recommendation. For specialist decisions, use qualified review. Stop when the source is missing rather than allowing the AI to fill the gap. Check that a claimed action has a real receipt or observable result before calling it complete.
Try it in your life or work
Personal practice
Take a fictional trip suggestion and list what would need current verification: hours, availability, prices and travel time. Do not book anything. Mark each as checked or unknown; do not invent sources.
Business practice
Audit the parcel draft above. Underline each unsupported statement and replace it with a review question. Try the customer-reply demo and inspect its placeholders rather than sending the result.
Keep: A claim-check list with sources and unresolved questions.
Use fictional or non-sensitive information. Written practice does not require an AI account. The live demos are optional; other tools have their own terms and costs.
Show the answer and explanation
Open the relevant primary source and check the exact claim.. A source must actually support the statement. Repeated model agreement is not independent evidence.
This is a self-check, not a scored certification. Answers and exercise notes are not submitted to AstraCruz.