AI Prompting · Level 1: Prompting Basics · Lesson 5
Check, question and improve AI answers
An AI response is a strong first draft, not a final deliverable. In this lesson, you will build a critical evaluation habit — a systematic 5-point check that catches errors before they reach your audience, and a follow-up prompting toolkit to turn a good first answer into an excellent final one.
By the end of this lesson, you can:
- Apply a 5-point verification checklist to any AI-generated response.
- Identify the three most common AI failure modes: hallucination, recency gaps, and embedded assumptions.
- Write effective follow-up prompts to iteratively refine, shorten, or fact-check an AI answer.
Why AI responses always need a human check
AI language models are pattern-matching engines, not fact databases. They are trained to produce statistically plausible text — which means a confident-sounding response and a factually accurate one can look identical on the surface. Understanding where errors come from helps you know where to look:
Hallucinations
The model generates specific facts, statistics, or citations that sound credible but cannot be traced to a verified source. They are especially common in niche topics or when the prompt implies that certain data exists.
Recency Gaps
AI models have a training data cut-off date. Any events, price changes, policy updates, or product releases after that date are invisible to the model — unless you supply them directly in your prompt.
Embedded Assumptions
When context is missing, AI fills the gaps with statistically common assumptions. A question about "our software" may produce an answer based on the most common software type in training data, not yours.
Tone Drift
Without explicit tone instructions (covered in Lesson 3), AI defaults to a neutral, slightly formal voice that may not match your brand, audience, or relationship with the reader.
The 5-point verification framework
Rather than reading an AI response with a vague sense of unease, use this structured checklist. Apply it systematically and it becomes fast — most outputs can be checked in under two minutes:
Check 1
Factual accuracy
Isolate every specific claim: statistics, dates, prices, names, and causal statements. Ask yourself: can I verify this with a reliable external source? For business-critical outputs, cross-reference at least one primary source before publishing.
Check 2
Source grounding
If you pasted a document or data set into your prompt, confirm the AI used only that material. Ask the AI directly: "Which part of the text I provided supports this claim?" If it cannot answer, the claim is likely generated, not extracted.
Check 3
Recency
Check whether time-sensitive information (regulations, market data, software versions, prices) could have changed after the model's training cut-off. Add a disclaimer or verify against a live source if recency matters for your use case.
Check 4
Completeness
Re-read your original prompt and confirm every question was answered. AI assistants sometimes address the most prominent part of a multi-part prompt and quietly drop the secondary questions. If something is missing, ask for it explicitly in a follow-up.
Check 5
Bias & framing
Consider whether the response reflects your intended perspective and tone. AI models can default to over-optimistic, overly cautious, or culturally biased framings depending on training data. Read the response as your audience would — not as someone who wrote the prompt.
Follow-up prompts: the iterative improvement toolkit
A first AI response is rarely the last. These follow-up prompt patterns are the most effective tools for converting a good first draft into a polished deliverable:
- Verify a specific claim: "You stated [X]. Please identify the exact sentence in my original text that supports this. If you cannot find it, retract the claim and flag it as [UNVERIFIED]."
- Shorten without losing meaning: "Reduce this to under 100 words. Keep the three most important points. Do not add any new information."
- Adjust tone: "Rewrite the second paragraph in a warmer, more direct tone. Avoid corporate jargon. Keep the same facts."
- Request a gap check: "Review your response against my original question. List any parts of my question that you did not fully address."
- Challenge an assumption: "Your response assumes [X]. Rewrite assuming [Y] instead, and note how the conclusion changes."
Developing calibrated trust in AI
The goal of this lesson is not to make you distrust AI — it is to help you develop calibrated trust: knowing precisely when to rely on an output and when to verify it independently.
Low-stakes, creative, or internally-facing tasks (brainstorming, drafting internal notes, rephrasing copy) warrant minimal verification. High-stakes outputs — anything involving financial figures, legal language, medical information, or public-facing claims — always warrant at least a targeted spot-check using the 5-point framework.
The most effective AI users are not those who trust the AI blindly, nor those who distrust it reflexively. They are the ones who have developed a fast, systematic habit for knowing the difference.
Illustrative example
Prompt Makeover
A marketing manager receives an AI-generated summary of a competitor's pricing model. Before sharing it with their director, they apply the 5-point verification check — and find a critical error.
Original prompt
Summarise Competitor X's pricing tiers from the information below.What is missing
- Fact verification: are the stated prices current or from outdated training data?
- Source grounding: did the AI use only the pasted text, or blend in its own assumptions?
- Completeness: were any pricing tiers mentioned in the source document left out?
- Recency check: does the AI acknowledge any knowledge cut-off limitations?
Improved prompt
Summarise Competitor X's pricing tiers using ONLY the text I have pasted below.
After the summary, add a section titled 'Verification Flags' and list:
1. Any pricing figures you could not confirm from the pasted text alone (mark as [UNVERIFIED]).
2. Any assumptions you made that are not explicitly stated in the source.
3. A reminder if this information may have changed since your training data cut-off.Why it is better
- It restricts the AI to the pasted source, reducing hallucination risk.
- It forces the AI to self-report any gaps or unverified figures.
- The 'Verification Flags' section makes the human review step fast and systematic.
- It builds a transparent audit trail — crucial for anything shared with stakeholders.
What still needs checking
- Cross-reference any key pricing figures against Competitor X's live website.
- Check if the competitor has announced pricing changes since the document was written.
Practice
Identify the stronger follow-up prompt
A researcher has received an AI-generated response about renewable energy adoption rates. They want to improve it. Read both follow-up prompts and select the one that will produce a more accurate, verifiable revision.
Illustrative example
The AI returned a 4-paragraph answer on global solar panel adoption. One statistic (35% year-on-year growth in 2023) seems unusually high and the researcher cannot recall seeing that figure in their source documents.
Check the model answer
Option B directly challenges the specific claim with a verification request.
Option A is vague — 'make it more accurate' gives the AI no direction and it may simply rewrite with equal or greater confidence. Option B identifies the exact suspect statistic, asks the AI to trace it back to the source, and provides a clear protocol ([UNVERIFIED]) to flag anything it cannot support. This is targeted, reproducible, and leaves a clear audit trail.
Knowledge check
Check the answer
Option B is the correct answer.
AI hallucinations occur when a model generates confident-sounding statements that are not supported by its training data or the context you provided. They are not intentional deception — but they can be equally damaging if not caught. The 5-point verification habit is your primary defence.
Keep this
The 5-Point AI Output Verification Checklist
Run every important AI response through this checklist before sharing or publishing:
✅ 1. FACTUAL ACCURACY — Can I verify key claims against a reliable source?
✅ 2. SOURCE GROUNDING — Did the AI stay within the context I provided, or invent extras?
✅ 3. RECENCY — Could this information be outdated due to a training cut-off?
✅ 4. COMPLETENESS — Did the AI answer every part of my question?
✅ 5. BIAS & FRAMING — Does the response reflect my intended tone, not the AI's default?
If any box is uncertain → use a targeted follow-up prompt to investigate before publishing.Build this habit once and it becomes second nature. The 5-point check takes under 2 minutes for most work outputs and can prevent costly corrections later.
Optional browser progress
Your progress
0 of 5 available lessons completed
Progress stores only the course identifier and completed lesson identifiers in this browser. It does not save your answers, prompt text or identity, and it does not sync to another device. Clearing browser data removes it.