Plain English
Treat fluent text as a starting point
An AI-generated research paragraph can organise ideas while mixing correct facts with mistakes. A citation may exist but refer to a different period, and an accurate number may be used in an unsupported explanation. Verification therefore works claim by claim. Asking whether the whole paragraph sounds convincing is a much weaker test than checking what each sentence actually asserts.
The joint investor alert on artificial intelligence and investment fraud warns that generated information can be inaccurate, incomplete or misleading. This does not mean every AI-assisted statement is false, nor that every error is fraud. It means polished language and apparent authority should not substitute for inspecting evidence. The same principle applies to a human-written summary.
Worked example
Audit a deliberately faulty paragraph
The following is an invented AI-style paragraph written for this exercise, not output attributed to a real service: “Birch Systems doubled quarterly revenue from £100m to £110m. Its £440m annual revenue confirms the same quarterly trend. The increase proves its AI product caused stronger customer demand.” Each sentence contains a problem that requires a different check.
The fictional source packet contains quarterly revenue of £100m for the three months ended 30 June 2025 and £110m for the equivalent period in 2026. It also contains £440m revenue for the full year ended 31 December 2025. Currency and accounting scope are unchanged. No source in the packet breaks down demand by product or establishes a causal effect from an AI product.
First, recompute growth: (110 − 100) / 100 = 10%, not 100%. Second, label the £440m figure as annual; it cannot directly confirm a quarterly trend. Do not divide it by four and invent a reported quarter, because revenue may be seasonal. Third, mark the product explanation as unsupported. A revenue increase alone does not identify its cause.
| “Doubled” |
Calculation error: the comparable quarter grew 10% |
| Annual figure used as quarterly evidence |
Period mismatch: retain the full-year label |
| AI product caused demand growth |
Unsupported causal claim: no product-level evidence provided |
Checking the Claim
Write the corrected claim and its limits
A defensible replacement is: “In this hypothetical source packet, revenue for the comparable quarter increased by 10%. The packet does not establish which products or other factors explain that increase.” The annual amount can be reported separately with its own dates, but it should not be recruited as evidence for a claim it does not address.
Use a repeatable verification trail: copy the claim, open the original source, identify the relevant passage and period, check the units and arithmetic, and state what remains unresolved. Retain the source version and the date it was accessed. A page title or search excerpt is not enough when the underlying table supplies a different definition.
A second AI response is not an independent source merely because it agrees. It could draw on the same material or reproduce the same error. Asking for citations may help locate documents, but each link still needs inspection. If the source is inaccessible, record that limitation rather than labelling the claim verified. Lack of access is not proof that the statement is false.
Limits and Assumptions
Verification has a boundary
US SEC filing guides describe how annual, quarterly and current reports can be located through EDGAR. These document types belong to the US system; other jurisdictions have different filing arrangements. Even an authentic report may contain estimates or management assertions. Confirming that management made a claim is different from confirming that the claim will prove correct.
Arithmetic checks cannot resolve every question. Establishing a causal explanation may require evidence absent from public documents. An honest review can stop at “not established by these sources”. Avoid uploading confidential research or personal information to a tool without understanding its applicable data handling. This exercise describes a method of reading evidence, not an endorsement or certification of any AI system.
Common mistake
Do not mistake citations for completed checks
A common mistake is placing a link beside a sentence and considering the work finished. Check whether the document supports that exact wording, timeframe and level of certainty. A source that mentions a product does not prove the product caused growth. Keep observations and interpretations separate when rewriting the paragraph, so the correction does not introduce a new unsupported claim.
Self-check
Check your understanding
Why is £440m not evidence of a particular quarter’s revenue?
It covers a full year. Dividing by four would create an average, not a disclosed quarterly result, and would ignore seasonality.
Does agreement from another AI verify the claim?
No. Agreement is not independent source evidence; both responses may rely on the same material or repeat an error.
What should replace an unsupported causal explanation?
Report the supported observation and explicitly state that the available evidence does not establish its cause.
Optional workflow context
A connection to Strata Research
Strata Research organises company research into reports. The evidence checklist above can also be used independently with source documents.
Strata Research
Disclaimer
Educational Use Only
This article is for information and financial education only. Examples are hypothetical and are not personalised investment advice or recommendations to buy, sell or hold a security.
