Claim calibration in AI-enabled work: match language to evidence

Inventory claims before polishing prose

Extract each material assertion from the draft. Include numbers, comparisons, causal statements, predictions, product capabilities, quotations, and recommendations. For each claim, identify the source, evidence type, population or scope, date, method, and consequence if the claim is wrong.

The Federal Trade Commission's Advertising FAQs explain that advertising claims must be truthful and evidence-based, that both express and implied claims matter, and that the type of support depends on the claim. Those are legal principles for advertising in the United States. This resource uses them only as a concrete example of why wording, implication, and support must be reviewed together; it is not legal advice.

Sources: [1]

Use the SCALE claim record

Record five fields for every material claim:

A claim should be narrowed, qualified, split, or removed when its coverage extends beyond its anchor. Adding a disclaimer after an overstated headline may not repair the impression created by the headline itself.

The FTC's Policy Statement Regarding Advertising Substantiation emphasizes having a reasonable basis before an objective advertising claim is disseminated and considering the type of claim, consequences of error, costs of substantiation, and support experts would expect. Outside advertising, the applicable rules differ, but the pre-publication habit of matching support to wording remains useful.

• Statement: Write the exact sentence and the reasonable implication a reader may take from it.
• Coverage: Define the people, cases, period, product version, geography, and conditions represented by the evidence.
• Anchor: Link the source, calculation, observation, or approved first-party fact supporting the statement.
• Limit: Note missing data, conflicting evidence, uncertainty, method constraints, and alternative explanations.
• Expression: Choose language that matches the support, such as observed, associated, estimated, suggests, intended to test, or unknown.

Sources: [2]

Worked artifact: a synthetic learning-programme update

A fictional programme owner receives attendance records, a confidence survey, AI-tool activity counts, and results from a small controlled work simulation. An AI-assisted draft opens with: "The academy improved workforce AI capability by 35%."

The reviewer creates a SCALE record:

| Field | Review finding | | --- | --- | | Statement | The sentence implies a broad causal improvement in workforce capability | | Coverage | The simulation included 18 volunteers from one function; there was no randomized comparison | | Anchor | Average rubric score increased between two similar simulation tasks; attendance and activity are separate measures | | Limit | Practice effects, task differences, volunteer selection, and reviewer variation remain unresolved | | Expression | "Among 18 volunteers, the average score on the second controlled task was higher; the exercise does not isolate the cause or represent the wider workforce" |

The reviewer also separates four claims that the first draft blended together:

These measures describe different populations and constructs. None can substitute automatically for another. The final update reports participation, self-report, activity, and task-bounded performance in separate rows, with denominators and source links.

This example is synthetic. It does not establish programme impact, causality, learning transfer, or a mature benchmark.

• 42 people completed the course;
• 31 responded to the confidence survey;
• 26 used the approved assistant during the observation window;
• 18 completed both controlled tasks.

Test express and implied meaning

Ask a reviewer who did not draft the document to state what each headline, chart, caption, and recommendation appears to claim. Compare that interpretation with the SCALE record. Check especially for:

Retain the original claim, revised claim, supporting artifact, reviewer rationale, and disposition. If the anchor changes, reopen the review instead of assuming the earlier wording still fits.

• percentages without denominators;
• trends described as causes;
• sample findings generalized to a workforce;
• activity described as capability;
• absence of detected errors described as safety;
• a pilot result described as established performance;
• product features described without checking the current plan and configuration;
• charts whose axes or color choices overstate differences.

Assess claim-calibration behavior

A controlled assessment can provide mixed-quality sources, a persuasive draft, and a deadline. Observable evidence includes whether the participant extracts material claims, checks denominators and scope, distinguishes observation from inference, notices implied meaning, narrows unsupported language, and preserves decision-relevant uncertainty.

Do not score caution as a goal by itself. Excessive qualification can make sound evidence unusable. The skill is to be as definite as the evidence permits and no more. Report only the behavior and artifacts observed in the scenario. Two Prune provides structured assessment evidence for human interpretation and does not make employment decisions.