AI ETHICS

Coherence Is Not Truth

Fluent language is a proposed answer. Reliability comes from checking individual claims against suitable evidence and documenting their limits.

Flowing blue text lines pass through an inspection opening toward separate evidence cards.
An image motif for separating linguistic form from external evidence.

The review question

Which suitable source supports an individual claim for this question—and how will we know when it must be corrected?

The fictional course case

An assistant claims that Python is a prerequisite for an introductory course. The approved programme does not require it; an optional supplementary sheet recommends Python only for an advanced exercise. The case is entirely invented.

Keep four things separate

Coherence

Sentences fit together linguistically. That does not show that they match the facts.

Claim

One verifiable statement. Facts, value judgements, forecasts, and recommendations need different review methods.

Evidence

A suitable, current source that supports, contradicts, or leaves open the specific claim within the right scope.

What research and guidance contribute

Studies of summarisation, truth benchmarks, and RAG identify errors in the tasks and model cohorts they examined. Source-attribution research evaluates whether a statement is supported by an identified source. Calibration, uncertainty, and self-correction research offer limited signals—not a general guarantee of truth.

Apply F to the case: A → B → D → C

Path from a verifiable claim through alternative explanations, source fit, and independent review to a bounded conclusion.
A review path with feedback: counterevidence can reopen an earlier explanation.
A · Narrow the questionSpecify the course, version, and meaning of “requirement.”
B · Form alternativesCompare correct source use, version mixing, and unsupported gap-filling.
D · Check and correctSplit claims, compare source and passage, and record errors and open limits.
C · Bound the answerWrite a reviewed, version-specific response and define a revision trigger.

How the answer changes

“The approved introductory programme lists no Python prerequisite. The supplementary sheet uses Python for an optional advanced exercise. The current version must be checked for any other course run.”

Match review depth to impact

Fully reviewing every reversible internal wording choice can be disproportionate. Claims about access, safety, or hard-to-reverse decisions need stronger evidence and effective approval. Without a suitable source, leave the answer open or refer it onward.

K and R have different roles

K · Governance

The four-step core cycle considers purpose, affected people, and knowledge gaps; reviews values and evidence; sets a provisional organisational rule; and defines feedback and revision.

R · Expert handoff

The separate advisory model R structures a specific expert-review assignment, for example when a real legal, contractual, or regulatory question arises. It does not replace K.

Course exercise · 60–75 minutes

0–10Read the fictional answer; mark facts and value judgements.
10–20Write individual claims and assign a provisional status.
20–35Use F-B to develop alternatives and distinguishing evidence.
35–50Use F-D to record source checks, counterevidence, and correction.
50–65Use F-C to define the answer, role, stop condition, and revision.
65–75Peer review: check source, version, scope, and remaining uncertainty.

Before use or publication

  • Each decision-relevant claim can be checked individually.
  • Source, version, passage, and scope fit the claim.
  • A link or repeated model answer does not count as review.
  • Facts and value judgements are separate; uncertainty stays visible.
  • Review, approval, and stop roles are named.
  • Correction path and revision trigger are defined.

Limits of this overview

Research findings apply to the tasks and models studied; they are not blanket metrics for current systems. F, K, and R are applied didactically here and are not presented as empirically validated or legally binding procedures. The case work and checklist are editorial teaching tools.