A Verification Procedure So You Do Not Use AI Output As Is

AI produces natural sentences. The problem is that naturalness does not guarantee truth. Numbers, dates, quotations and people's names in particular can be wrong while sounding plausible. So reading AI output and verifying AI output have to be separated.

Separate fact, claim and inference

Verification starts by dividing sentences into three kinds.

• Fact: supported by a document, official source or direct confirmation
• Claim: someone's opinion or interpretation
• Inference: something the AI produced by continuing the context

Without that distinction, readers take inferences for facts. In a public post that is especially dangerous. The related principle is set out in separating confirmed fact, claim and inference.

Check four things against the original

1. Source

Where does the basis of this sentence lie? It should be answerable with one of: an official document, a report, a raw note, or direct confirmation. "The AI said so" is not a source.

2. Reference date

Prices, policies, features and statistics change meaning once the reference date passes. Keep the time point — "as of September 2026" — and revise when it changes.

3. Numbers

Amounts, ratios, periods and usage figures are compared with the source once more. Check units and rounding too. Numbers arranged in a table are the hardest to notice.

4. Proper nouns

Check whether names have changed: people's names, titles, company names, product names, names of laws. AI readily invents plausible similar names.

Warning signs

• A sentence asserting something with no source
• Vague phrases like "recent research shows"
• A quotation that is not in the original
• Old prices or policies written as if current
• Links or document numbers that do not exist
• An over-confident tone

When you see one of these, find the basis before polishing the sentence. If you cannot find it, delete it or leave it marked "needs confirmation".

Final checklist before publishing

• Have I compared numbers and dates with the source
• Are names and affiliations correct
• Do quotations match the original
• Do source links still work
• Did I write an inference as a fact
• Is any sensitive information included
• Have I left a reference date and revision history

Verification is not about not trusting AI. It is a procedure for writing a post a person is responsible for. Fix even this procedure briefly, and you keep both the speed of AI and the trust.