SWANK AI Guidance Note 02
AI Governance · Evidence · Record Integrity

Core Standard

SOURCE OVER SUMMARY

AI-generated summaries may improve efficiency.

They should not obscure the evidence, uncertainty, disagreement or chronology contained in the underlying record.

Purpose

Artificial intelligence can help organisations process large volumes of information by summarising:

correspondence;
reports;
case records;
meeting notes;
policies;
complaints;
assessments;
research;
and other documentary material.

This can substantially improve administrative efficiency.

It can also alter operational meaning.

When complex records are condensed, important distinctions may disappear between:

established fact;
allegation;
interpretation;
disputed information;
uncertainty;
and recommendation.

In consequential environments, those distinctions matter.

A Summary Is Not the Source

An AI-generated summary is an interpretation of source material.

It should not silently become a replacement for that material.

Where a summary contributes to a consequential institutional process, the underlying records should remain available so that a later reviewer can determine:

what the original source said;
who created it;
when it was created;
what context surrounded it;
whether it was disputed;
what qualifications were present;
and whether later evidence changed the position.

The summary should improve access to evidence.

It should not make the evidence harder to inspect.

Source Traceability

Material statements in an AI-generated summary should remain traceable to their original source.

Depending upon the context, traceability may include:

document title or identifier;
author or origin;
date;
page, paragraph or section;
underlying record;
evidential status;
relevant qualification;
and identified uncertainty.

A reviewer should ideally be able to move from:

summary → proposition → source

without reconstructing the entire evidential history manually.

Summary-to-Source Drift

One significant risk is summary-to-source drift.

This can occur gradually.

For example:

Original record:
A qualified or disputed statement is recorded.

First summary:
The statement is shortened.

Later record:
The qualification disappears.

Subsequent summary:
The shortened statement is repeated as though established.

Over time, information that began as uncertain, disputed or contextual may appear definitive simply because it has been repeatedly condensed.

The apparent consistency of later records may therefore arise from repetition rather than independent corroboration.

Repetition Does Not Create Evidence

An AI-generated error should not acquire greater evidential authority merely because it has been copied into multiple records.

If one inaccurate summary is subsequently reproduced in:

case notes;
dashboards;
reports;
assessments;
meeting records;
later AI prompts;
or decision-support material,

the existence of multiple copies does not transform the original error into corroboration.

The relevant question remains:

What does the underlying source evidence establish?

Preserve Evidential Status

AI-assisted summaries should preserve distinctions such as:

The source records…

The individual states…

The organisation records…

The evidence indicates…

This remains disputed…

This has not been independently verified…

The available material does not resolve the difference…

These distinctions prevent a summarisation system from creating false certainty.

A coherent narrative is not necessarily an accurate one.

Preserve Disagreement

Complex institutional records often contain conflicting accounts.

For example:

Source A records X.

Source B records Y.

If the available evidence does not establish which account is correct, an AI system should not produce a blended narrative that conceals the disagreement.

The stronger summary may be:

The sources conflict and the available information does not presently resolve the difference.

Preserving contradiction is sometimes more accurate than resolving it.

Preserve Chronology

Chronology is part of evidential meaning.

Consider:

Decision → new evidence → review

and:

New evidence → decision → no review

The same events appear in both sequences.

Their institutional meaning is very different.

AI-assisted summarisation should therefore preserve material temporal distinctions, including where relevant:

event date;
record creation date;
date information became known;
decision date;
review date;
correction date;
and date circumstances materially changed.

A summary that preserves facts but loses sequence may still distort the record.

Preserve Uncertainty

AI systems often produce fluent, confident language.

That fluency can make uncertainty less visible than it was in the source material.

Where source information is:

incomplete;
ambiguous;
contested;
provisional;
unverified;
or contradictory,

the summary should preserve that status.

Appropriate language may include:

The available information is incomplete.

The sources conflict.

This remains unverified.

The record does not establish a reliable conclusion.

Refusing to create false certainty can be a feature of good record design.

Correction

Where AI introduces a material error into a summary:

the underlying source should be checked;
the inaccurate statement should be corrected or clearly annotated;
affected downstream records should be identified where proportionate;
material corrections should be propagated;
the correction history should remain visible where auditability requires it.

Correction should not necessarily mean deleting the original record.

In some environments, preserving the original alongside a visible correction is more important for accountability.

Downstream Error Propagation

A correction may need to extend beyond the first inaccurate summary.

Organisations should consider whether the information has already entered:

later summaries;
reports;
assessments;
dashboards;
recommendations;
decision records;
AI prompts;
retrieval systems;
or other downstream workflows.

Correcting only the original output may leave the operational error active elsewhere.

Human Review

Meaningful human review requires access to source material.

A reviewer cannot adequately challenge a summary if the underlying information is unavailable.

Where summaries contribute to significant decisions, reviewers should be capable of examining:

original records;
material competing evidence;
relevant qualifications;
identified uncertainty;
and subsequent corrections.

Human oversight is weakened when a reviewer can see only the AI-generated interpretation.

High-Stakes Environments

The consequences of summarisation error increase where records contribute to decisions concerning:

healthcare;
safeguarding;
education;
employment;
public services;
disciplinary processes;
financial access;
regulatory action;
legal rights;
or other significant individual interests.

In these environments, organisations may require stronger standards for:

source preservation;
traceability;
uncertainty;
chronology;
correction;
human review;
and documentation of AI involvement.

The greater the potential consequence, the weaker the case for allowing summaries to replace source evidence.

Questions for Organisations

Where AI is used to summarise institutional records, organisations may ask:

Can every material proposition be traced back to its source?
Does the summary distinguish fact from allegation or interpretation?
Is disagreement preserved?
Is material uncertainty visible?
Has chronology been retained accurately?
Are original records still accessible?
Can a human reviewer inspect the underlying material?
What happens when an AI-generated summary is wrong?
Are corrections propagated to downstream records?
Could an independent reviewer later reconstruct how the summary was produced and relied upon?

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 information retain its meaning as it moves through the institution?

Domain 05 — Escalation, Challenge & Contestability

Can inaccurate or disputed summaries be challenged and corrected?

It may also engage:

Domain 02 — Decision Integrity & Human Oversight

Where human decision-makers rely upon AI-generated summaries.

Domain 06 — Risk, Harm & Operational Resilience

Where summarisation errors could materially affect consequential outcomes.

SWANK AI Standard

SOURCE OVER SUMMARY

Summaries may improve efficiency. Source material should remain available wherever context, dispute, accountability or consequential decision-making requires it.

AI-generated summaries should preserve:

source

status

context

chronology

uncertainty

disagreement

correction

A summary should make evidence easier to use.

It should not make evidence disappear.

Related SWANK AI Guidance

Guidance Note 01 — Handling AI-Assisted Complaints and Correspondence

Guidance Note 05 — Preserving Disagreement in AI-Assisted Records

Guidance Note 14 — AI Incident Reporting and Error Correction

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.