Starter path
Lesson 2 of 3

Inspect before you trust

Separate facts, assumptions, and recommendations so uncertainty stays visible.

Confidence is not evidence

Fluent writing can hide a weak claim. Read the answer once for meaning, then again for statements that could be wrong, outdated, or missing context.

Visual explainer

Separate the answer before you judge it

An answer becomes easier to verify when you sort its sentences by what they are doing.

Diagram, step by step
  1. AI answerConfident language is only the outer layer
  2. FactsClaims that can meet primary evidence
  3. AssumptionsConditions that have not been proved
  4. RecommendationsJudgments shaped by priorities and risk
Relationships
  • AI answercheck: Facts
  • Factsexpose: Assumptions
  • Assumptionsdecide: Recommendations
Different sentence types need different checks; they should not share one confidence score.
  • Facts: claims that can be checked against a primary source.
  • Assumptions: conditions the answer relies on but did not prove.
  • Recommendations: judgments that depend on your priorities and risk.

A source link is a lead, not proof that the source supports the sentence beside it.

Triage the answer before checking it

You do not need to verify every sentence with the same effort. First separate claims that could be checked from opinions, instructions, and assumptions. Then prioritize claims that are central to the decision, likely to change over time, or costly if wrong. This turns verification from an endless search into a focused review.

Different sentence types need different review actions
Sentence typeExampleBest next action
FactThe plan includes 10 GB of storageCheck current official documentation
AssumptionYour users prefer a mobile appAsk what evidence supports the assumption
RecommendationChoose the cheaper planCompare it with your priorities and risk
InstructionRun this command as administratorUnderstand effects and test in a safe context

Trace an important claim to primary evidence

Visual explainer

The claim-to-evidence chain

A citation becomes useful only when the exact claim survives every link in the chain.

Diagram, step by step
  1. Exact claimCopy the statement you need to verify
  2. Primary sourcePrefer the owner, original data, or direct observation
  3. Supporting passageLocate the words or data that address the claim
  4. Context matchCheck date, version, region, and conditions
  5. Review judgmentSupported, contradicted, or unresolved
Relationships
  • Exact claimsearch: Primary source
  • Primary sourcelocate: Supporting passage
  • Supporting passagecompare: Context match
  • Context matchdecide: Review judgment
Opening a link is not the end of verification; it is the start of a claim-level comparison.
  1. Freeze the wording

    Write the exact claim before searching so the target cannot quietly change.

  2. Find the closest primary source

    Use official documentation, original research, raw data, or observed product behavior.

  3. Read around the evidence

    Check headings, definitions, exceptions, and footnotes—not one isolated sentence.

  4. Record the result

    Mark the claim supported, contradicted, partly supported, or still unresolved.

Check freshness, scope, and missing conditions

A statement may once have been true and still be wrong for your case. Product limits change, laws differ by jurisdiction, features vary by plan, and research results depend on the tested population. Always ask “true for whom, where, when, and under which conditions?” before transferring a claim into your work.

Visual explainer

Four context layers around a claim

The words can match while the surrounding conditions make the evidence unusable.

Diagram, step by step
  1. Date and versionIs the evidence current for this product or rule?
  2. Region and jurisdictionDoes it apply where the decision will be used?
  3. Plan and permissionsIs the feature available to this account and role?
  4. Conditions and exceptionsWhat must be true for the claim to hold?
Relationships
  • Date and versionnarrow: Region and jurisdiction
  • Region and jurisdictionqualify: Plan and permissions
  • Plan and permissionsconfirm: Conditions and exceptions
Evidence supports a claim only inside the conditions it actually covers.

Handle uncertainty without hiding it

False certainty

This is definitely supported because three articles repeat it.

Useful uncertainty

Two secondary sources repeat the claim, but the official documentation does not confirm it. Treat it as unresolved until the provider clarifies.

Practice an answer audit

A compact worksheet for reviewing an AI answer
ClaimRisk if wrongEvidence neededStatus
Copy one exact statementLow, medium, or highName the best primary source or direct testSupported, contradicted, partial, or unresolved
Expose one assumptionWhat decision could it distort?User data, requirement, or observationConfirmed or still assumed
Separate one recommendationWhose priorities does it serve?Decision criteria and trade-offsAccepted, changed, or rejected
Knowledge check

Check your understanding

Three practice questions. They are not saved to an account or awarded points.

01Which sentence is a checkable factual claim?
02An answer includes a source link. What does the link prove by itself?
03What should you do with an assumption hidden inside an AI answer?

This open path saves completion on this device only. It does not add account data or points.