Verification practice
Can the Source Carry the Claim?
An AI answer can sound coherent while mixing one real detail with a larger conclusion the evidence cannot support. Follow a fictional claim back to its strongest source, classify each part and repair the wording.
Fictional caseNo account, profile or personal detailsNo result is sent or saved
- Find the original
- Split the bundled claim
- Match each detail
- Repair the wording
- Keep limits visible
Read this boundary first
Practice, Not Proof of Ability
This lab is a learning exercise. It is not a validated assessment of intelligence, judgement, media literacy or workplace skill.
The case is entirely fictional
The organisation, AI answer, sources, dates, people and results were invented for this activity. They are not evidence for or against real meeting-free policies. Your choices stay in this page while it is open: the lab makes no network request and uses no browser or server storage.
The checking method reflects current risk guidance: NIST discusses confidently false generative-AI output and the role of testing, evaluation, verification and validation, while the Office of the Australian Information Commissioner warns that commercial AI can produce inaccurate results and advises care with personal information. Read the NIST Generative AI Profile and OAIC guidance on commercial AI products.
Four evidence relationships
Name What the Source Actually Does
A claim can contain several parts with different evidence status. Classify each part separately instead of accepting or rejecting the whole paragraph at once.
Supported
The Source Directly Matches
The relevant source states the material detail and its context does not reverse or narrow the meaning.
Contradicted
The Source Says Something Different
A date, number, event or other checkable detail conflicts with what the source reports.
Overstated
The Claim Outruns the Evidence
The source offers some related evidence, but the claim makes it larger, more certain, more causal or more general.
Not established
The Source Does Not Answer It
The available evidence cannot settle the claim. That does not automatically make the claim false.
Fictional 2025 case file
One Answer, Three Sources
Read the wording closely. A later page may be easier to find or more confident, but that does not make it the strongest evidence.
AI answer: “A 2025 workplace pilot proved that a meeting-free day increased productivity by 32% for all 240 employees, caused no problems and should be required by every business.”
Original pilot summary · 14 May 2025
Methods and Results
- 240 employees were invited.
- 82 volunteers from two desk-based teams completed a six-week pilot.
- Participants reported an average 12% increase in uninterrupted focus time compared with their own baseline.
- Some participants reported scheduling difficulties.
- There was no control group. The summary calls the pilot exploratory.
Organisation press release · 20 May 2025
Promising Early Signs
The release describes early signs from a “240-person initiative” as promising and says participants felt more productive. It gives no methods and does not explain that 82 volunteers completed the pilot.
Workplace blog · 4 June 2025
Productivity Up 32%
The post says meeting-free days boost productivity by 32%. It links only to the press release, provides no methods and does not link to the original pilot summary.
Ten aligned decisions
Check, Classify and Repair
Make one source decision, classify four claims and choose five evidence-matched repairs. You will receive an outcome band, not a percentage or ability score.
Static learning version
The interactive controls have not loaded, but the full reasoning remains available:
- Use the original pilot summary because it contains the methods, participant count and limits.
- Supported: the pilot happened in 2025.
- Contradicted: 240 people completed it. In fact, 240 were invited and 82 completed it.
- Overstated: it proves productivity rose for every kind of team. The exploratory pilot covered two desk-based teams, used self-reporting and had no control group.
- Not established: every business should require the policy. The sources did not test a universal recommendation.
A source-matched repair says: “A small, exploratory 2025 pilot found an average 12% rise in self-reported uninterrupted focus time among 82 volunteers from two desk-based teams. Some reported scheduling difficulties. The pilot does not establish that every business should require the policy.”
Transfer the method
Use the Same Loop on a Real AI Answer
The topic will change, but the checking work stays recognisable. Increase the depth and independence of verification when the consequences rise.
- Copy the exact claim.
Preserve its numbers, dates, certainty words and recommendations before paraphrasing anything.
- Split bundled statements.
Separate what happened, who was involved, what changed, why it changed and what someone should do.
- Trace derivative pages backwards.
Open the report, dataset, rule or study that can show methods and limits rather than relying on a summary of a summary.
- Classify each relationship.
Use supported, contradicted, overstated or not established. Keep unknowns visible.
- Repair before sharing.
Change the scope, measure and certainty until the source can carry every important word.
Keep learning
Pair Verification with Better Questions
Return to the learning hub for practical guides, or use the Thinking Check to practise judgement across a wider set of AI-use scenarios.