SWANK AI Guidance Note 05
AI Governance · Evidential Traceability · Record Integrity

Core Standard

EVIDENCE OVER INFERENCE

A coherent narrative is not necessarily an accurate one.

Where disagreement exists, it should remain visible until the evidence resolves it.

Purpose

Complex institutional environments frequently contain competing accounts.

Different people may describe:

the same event differently;
different parts of the same event;
information available at different times;
conflicting interpretations;
or facts that have not yet been independently verified.

AI-assisted summarisation can make this information easier to process.

It can also remove important distinctions by producing a single fluent narrative from material that was never actually agreed.

Operational coherence should not be created by eliminating legitimate disagreement.

Disagreement Is Evidential Information

The existence of disagreement is itself relevant information.

A record may appropriately distinguish between statements such as:

The individual states…

The organisation records…

The available evidence indicates…

This remains disputed…

This has not been independently verified…

These formulations preserve the evidential status of the information.

They allow a later reviewer to understand not only what was recorded, but also how certain or contested it was.

AI Should Not Resolve Disputes by Writing Style

Generative AI is designed to produce coherent language.

That strength can become a weakness when the underlying material is contradictory.

Where two sources disagree, a system may be tempted to:

merge them;
smooth over the difference;
choose the more detailed account;
choose the more recent account;
prefer the more confidently written account;
or create a blended narrative.

None of those processes necessarily establishes which source is correct.

A linguistically coherent account may still be evidentially inaccurate.

Preserve Source Attribution

Where disagreement exists, source attribution should remain visible.

For example:

Source A records X.

Source B records Y.

The available evidence does not presently resolve the difference.

This is often analytically stronger than producing a single statement that implies agreement.

Source attribution helps a later reviewer determine:

who said what;
when it was recorded;
what evidence supported it;
whether it was disputed;
and whether later information altered the position.

Preserve Evidential Status

AI-assisted systems should distinguish between categories such as:

Established Information

Material directly supported by reliable evidence.

Reported Information

A statement recorded from an identifiable person or source.

Interpretation

A conclusion or explanation drawn from information.

Disputed Information

A material proposition actively contested by another relevant source.

Unverified Information

A proposition that has not yet been independently established.

Later Correction

Information that changes, qualifies or corrects an earlier record.

These categories should not silently collapse into one another during summarisation.

Repetition Does Not Resolve Disagreement

A disputed statement may appear repeatedly in later records.

That repetition does not necessarily make it independently established.

For example:

an initial record contains a disputed proposition;
a later summary repeats it;
another report relies on that summary;
the same proposition appears in a subsequent decision record.

The existence of several later references may create the appearance of corroboration.

But if all of them originate from the same initial source, they may represent repetition rather than independent evidence.

Source traceability is therefore essential.

Summary-to-Source Drift

Disagreement can become less visible each time material is summarised.

An original record may say:

The organisation records X. The individual disputes this.

A later summary may say:

The record states X.

A subsequent document may say:

X occurred.

The evidential status has changed even though no new evidence was introduced.

This is a form of summary-to-source drift.

AI-assisted record systems should be designed to prevent it.

Chronology Matters

Disagreement may also change over time.

A record should distinguish between:

what was believed initially;
what was later challenged;
what new evidence emerged;
whether the original position changed;
and what remains unresolved.

For example:

Initial record → challenge → new evidence → revised conclusion

is different from:

Initial record → challenge → no reassessment

Both contain disagreement.

Their governance implications are different.

New Evidence Should Be Able to Change the Record

A stable institutional system should not preserve an earlier interpretation simply because it appeared first.

Where materially relevant new evidence emerges, systems should permit:

reassessment;
correction;
annotation;
reconsideration;
competing-source review;
and updating of summaries.

The objective is not to erase historical records.

It is to ensure that later users can distinguish between:

what was recorded at the time

and

what the current evidence supports.

Disagreement and AI-Assisted Summaries

Where AI is used to summarise contested material, organisations may consider requiring the system to preserve:

source attribution;
uncertainty;
disagreement;
evidential status;
chronology;
later corrections;
contradictory evidence;
and unresolved questions.

The system should not infer factual resolution merely because producing a unified narrative is easier.

Human Review

Where disagreement materially affects a consequential decision, human reviewers should be able to access the original records.

A reviewer should be able to determine:

what each source actually said;
whether the accounts genuinely conflict;
whether one source has stronger evidential support;
whether later evidence changes the position;
and whether the disagreement remains unresolved.

Human oversight is weakened when reviewers receive only a blended AI-generated summary.

High-Stakes Environments

Preserving disagreement becomes especially important where records may influence:

safeguarding;
healthcare;
education;
employment;
disciplinary action;
eligibility decisions;
public services;
regulatory intervention;
legal processes;
or other consequential outcomes.

In these environments, an unresolved dispute should not silently become a settled institutional fact.

The greater the consequence, the greater the need for visible source attribution and evidential discipline.

Correction and Annotation

Not every disagreement can be resolved immediately.

Where a factual correction can be established, the record may need to be corrected.

Where the evidence remains contested, an annotation may be more appropriate.

For example:

This information remains disputed.

That preserves both the historical record and the current evidential status.

The purpose is not to force artificial agreement.

It is to keep the record accurate about what is and is not established.

Avoiding Institutional Consensus by Repetition

AI-assisted systems can contribute to feedback loops.

An initial interpretation may be:

summarised;
stored;
retrieved;
reused as context;
and presented to later decision-makers.

If later systems treat the earlier summary as established fact, the institution may appear to have reached consensus.

But the apparent consensus may simply reflect repeated reuse of the same unresolved proposition.

Organisations should therefore ask:

Is this conclusion independently supported, or merely repeatedly recorded?

Questions for Organisations

Where AI processes disputed or contradictory records, organisations may ask:

Are competing accounts preserved separately?
Can each material proposition be traced to its source?
Is disputed information visibly labelled?
Are interpretations distinguishable from facts?
Does summarisation preserve uncertainty?
Can later evidence change the operational record?
Are corrections propagated?
Does repetition create the appearance of false corroboration?
Can a human reviewer access the original source material?
Can an independent reviewer determine what remains unresolved?

SIAAF Relevance

This Guidance Note principally relates to:

Domain 03 — Evidence & Traceability

Can the organisation prove how an important conclusion was reached?

Domain 04 — Communication & Feedback Integrity

Does important information retain its meaning as it moves through the institution?

Domain 05 — Escalation, Challenge & Contestability

Can an affected person or reviewer challenge an inaccurate or incomplete institutional account?

It may also engage:

Domain 02 — Decision Integrity & Human Oversight

Where human decision-makers rely upon disputed or AI-summarised information.

Domain 06 — Risk, Harm & Operational Resilience

Where unresolved disagreement could materially affect consequential outcomes.

SWANK AI Standard

EVIDENCE OVER INFERENCE

A coherent narrative is not necessarily an accurate one.

Where disagreement exists:

preserve the sources

preserve the distinction

preserve the uncertainty

preserve the chronology

preserve the correction pathway

until the available evidence justifies resolution.

AI should help institutions understand complex records.

It should not create certainty that the evidence does not support.

Related SWANK AI Guidance

Guidance Note 02 — AI Summarisation and Record Integrity

Guidance Note 08 — Reassessment Alongside Escalation

Guidance Note 17 — Source Traceability by Design

Guidance Note 18 — AI, Chronology and Organisational Memory

Guidance Note 19 — Correction Rights in AI-Assisted Systems

SWANK AI

Independent AI & Institutional Assurance

We do not just review AI. We review the institutional systems responsible for governing it.