Decision Desk · Issue 07 · complete public learning edition

Can People Trust This AI Explanation?

Decide when AI may translate complex authoritative information without changing its meaning, hiding uncertainty, blocking access, or becoming individualized professional advice.

Learning time90 minutes
Teaching12 cumulative lessons
InstrumentFidelity Test
BoundaryExplanation—not advice

Authority → meaning → audience → access → action → correction.

An explanation succeeds only when people can reach, understand, and safely use the meaning the authoritative source actually supports.

01

Freeze source and controlling meaning

Gate
02

Preserve conditions, exceptions, and uncertainty

Gate
03

Define audience and consequence

Gate
04

Provide language, disability, and format access

Gate
05

Keep advice and escalation boundaries visible

Gate
06

Monitor reliance, correct, and retest

Gate

Twelve cumulative lessons.

Use supplied fictional notices only. Never enter personal health, legal, financial, benefits, safety, or eligibility facts.

01
Issue 07

Identify authoritative meaning

Open

Start with the controlling source, owner, version, scope, exceptions, and uncertainty. AI cannot simplify meaning that the organization has not resolved.

02
Issue 07

Define audience and consequence

Open

Reading level alone is not an audience. Identify language, disability and literacy access, context, prior knowledge, likely action, and consequence of misunderstanding.

03
Issue 07

Separate explanation from advice

Open

General explanation describes authoritative information. Individualized advice applies it to personal facts and may require a qualified professional with responsibility for the decision.

04
Issue 07

Preserve conditions and exceptions

Open

Shorter wording can silently remove eligibility conditions, contraindications, deadlines, exclusions, alternatives, or procedural rights that determine the outcome.

05
Issue 07

Represent uncertainty honestly

Open

Do not convert ambiguous, conditional, evolving, probabilistic, or incomplete authority into a confident individualized answer.

06
Issue 07

Test meaning fidelity

Open

Compare each proposition, condition, exception, action, and uncertainty marker to the source. Fluency and similarity scores cannot replace accountable review.

07
Issue 07

Test comprehension and action

Open

Ask whether a representative person understands what the information means, what it does not establish, what action is required, and where to get help.

08
Issue 07

Build language access without meaning drift

Open

Translation requires source-version control, terminology, qualified validation, variant management, accessible correction, and a route when an equivalent expression is uncertain.

09
Issue 07

Design disability and format access

Open

Provide semantic text, keyboard operation, readable structure, captions/transcripts where relevant, nonvisual meaning, alternative formats, and accommodation routes.

10
Issue 07

Escalate individualized questions

Open

Recognize personal eligibility, diagnosis, legal strategy, financial suitability, emergency, safety, and other consequential questions and route them without pretending certainty.

11
Issue 07

Monitor adverse reliance and correct

Open

Track misunderstandings, complaints, omissions, unequal access, escalations, changed sources, and downstream harm. Correct every affected version and notify where necessary.

12
Issue 07

Issue an explanation-specific disposition

Open

Do not use, repair first, preserve human-led explanation, or evaluate a bounded non-advisory pilot. The instrument never approves individualized advice.

Plain language cannot flatten truth.

  • the authoritative source or controlling meaning is unresolved
  • a material condition, exception, deadline, or uncertainty disappears
  • the output sounds certain where the source is conditional
  • a person could reasonably mistake explanation for individualized advice
  • language or disability access changes or withholds material meaning
  • the system cannot recognize and escalate a consequential personal question
  • there is no owner for source change, correction, notice, and retesting

High-Stakes Explanation Fidelity Test

Use fictional source text and audiences only. Do not enter personal facts, eligibility details, records, diagnoses, disputes, credentials, or confidential information.

0/30fidelity gates evaluated
Simpler language is not successful if it changes the decision.
Meaning drift, omitted exceptions, false certainty, inaccessible output, or advice-like personalization fails closed.
Fictional explanationMeaning preservedExceptions retainedUncertainty visibleAction consequence clearAccessible alternativeHuman escalation works
Benefits eligibility notice
Safety instruction
Legal-process overview
Financial-information summary
Translated public-service notice
Bounded disposition

Restoring your local test…

The educational product and reusable Fidelity Test are complete.

Real use still requires current domain authority, qualified review, representative fidelity and comprehension evidence, human language/disability validation, exact configuration testing, escalation operations, correction, and organizational authorization.