SWANK AI Guidance Note 19
Correction Rights · AI Governance · Record Integrity

Core Standard

ERROR CORRECTION MUST BE PART OF SYSTEM DESIGN

A system capable of generating consequential information should also provide a meaningful route to:

challenge

correct

annotate

propagate

and

reconsider

material errors.

Preventing mistakes is important.

Governance also requires the ability to correct them when they occur.

Purpose

AI-assisted systems can generate, amplify and distribute factual errors quickly.

An inaccurate statement may begin in:

an AI-generated summary;
a classification;
a retrieved record;
a recommendation;
a generated chronology;
a decision-support output;
or a staff record informed by AI.

That information may then travel into:

reports;
dashboards;
assessments;
correspondence;
later summaries;
management systems;
downstream AI prompts;
and consequential decisions.

A governance framework should therefore consider not only:

How do we prevent error?

but also:

What happens when someone identifies one?

Correction Is an Operational Capability

Correction should not depend upon informal goodwill or exceptional intervention.

A functioning correction pathway should enable a person or reviewer to identify:

what information is said to be inaccurate;
where it appears;
the relevant source;
the proposed correction;
supporting evidence;
whether the original information remains disputed;
and which downstream records or decisions may be affected.

Correction should be designed into the system.

It should not have to be invented after an error becomes consequential.

A Correction Right Must Be Usable

A theoretical ability to request correction is not sufficient if the process is practically inaccessible.

A meaningful route should answer basic questions:

Who do I contact?

What information should I provide?

Who reviews the request?

What evidence will be considered?

When will the record be updated?

Will affected downstream records be checked?

Can the correction affect a decision already made?

If an affected person cannot understand how to raise an apparent factual error, the correction pathway may exist formally without functioning operationally.

Correction Is Not Always Deletion

Different errors require different responses.

An inaccurate or incomplete record may need:

correction;
annotation;
supplementation;
withdrawal;
replacement;
visible qualification;
or preservation alongside a corrected position.

Deleting the original record may sometimes remove important audit history.

For example, an institution may need to preserve:

Original record: X was recorded.

Correction: Later evidence established Y.

Current position: Y should now be relied upon.

That preserves both:

historical accountability;
and current accuracy.

Preserve the Audit Trail

A correction system should be able to show, where proportionate:

what the original record said;
when it was created;
when the apparent error was identified;
who reviewed it;
what evidence supported correction;
when the correction occurred;
and what downstream action followed.

A correction should improve accuracy without making institutional history impossible to reconstruct.

Disputed Information

Not every correction request will result in agreement.

Sometimes the available evidence remains contested.

Where disagreement cannot presently be resolved, the record should preserve that status accurately.

For example:

This information is disputed.

The available evidence does not presently resolve the difference.

That is preferable to presenting one position as settled where it is not.

Correction governance should distinguish between:

demonstrably incorrect

and

materially disputed.

Annotation Can Be Important

Where historical information must remain preserved, annotation may be more appropriate than overwriting.

A useful annotation may identify:

the original statement;
the nature of the challenge;
supporting evidence;
whether the issue was resolved;
the current institutional position;
and the date of review.

This allows later users to understand both the historical and current record.

Downstream Propagation

One of the most important correction questions is:

Where did the error go?

Correcting the first inaccurate record may not be sufficient if the information has already entered:

summaries;
dashboards;
assessments;
reports;
recommendations;
decisions;
AI prompts;
search indexes;
retrieval systems;
or later correspondence.

Material downstream effects may require review.

Correction Should Follow the Information Chain

A useful correction process may proceed:

identify error

↓

verify source

↓

correct or annotate original

↓

identify downstream uses

↓

update affected records

↓

reconsider affected conclusions

↓

document completion

This prevents an institution from correcting the source while leaving the operational consequences of the error unchanged.

Correction and Decision Reconsideration

A factual correction may matter because a decision relied upon the incorrect information.

The institution should therefore ask:

Did this error influence a decision?
Was the information material to the outcome?
Would the decision have been different without it?
Does the decision require reconsideration?
Should affected people be informed?
Are later systems still relying upon the same information?

Correction of data and reconsideration of outcome are related but distinct functions.

Both may be necessary.

AI Can Amplify Errors

AI systems can accelerate error propagation because they may reuse earlier information at scale.

An incorrect proposition may appear in:

an original AI summary;
a staff record;
a later AI retrieval;
a generated assessment;
a subsequent decision-support output.

Repeated appearance can create the impression of corroboration.

But the later records may all derive from the same original mistake.

Correction therefore requires source traceability.

Repetition Does Not Make an Error True

An inaccurate statement does not become more reliable because it appears repeatedly.

Organisations should distinguish between:

independent corroboration

and

multiple copies of the same underlying information.

Where AI systems retrieve previous summaries, this distinction becomes particularly important.

A correction pathway should be able to identify whether later apparent support is genuinely independent.

Human-Accessible Correction Routes

People affected by AI-assisted records should not need technical expertise to challenge a factual error.

They should not need to understand:

model architecture;
retrieval pipelines;
databases;
embeddings;
system prompts;
or technical integrations

before being able to say:

This record is inaccurate.

The institution should translate a potentially complex technical process into a clear human-accessible correction route.

Accessibility

Correction mechanisms should account for differing communication needs.

People may require:

written communication;
accessible forms;
plain-language explanations;
representative support;
alternative channels;
additional time;
or assistance organising evidence.

An inaccessible correction route can allow an otherwise correctable error to persist.

Accessibility is therefore part of record integrity.

Correction and Source Evidence

Correction should remain evidence-oriented.

A correction request may rely upon:

original documents;
contemporaneous correspondence;
source records;
official records;
verified chronology;
later evidence;
or other relevant material.

The institution should distinguish between:

a request to change a record

and

evidence demonstrating that the record is inaccurate or incomplete.

Correction rights should preserve evidential discipline rather than replace it.

Corrections to AI-Generated Summaries

Where an AI-generated summary is inaccurate, the review should consider:

what the source actually said;
whether the summary omitted context;
whether disputed information was flattened;
whether chronology was changed;
whether unsupported inference was introduced;
and whether later users relied upon the summary.

The correct response may require more than editing one sentence.

The institution may need to reconsider the summarisation workflow itself.

Corrections to Classifications

AI may classify information as:

high risk;
low priority;
repetitive;
urgent;
non-compliant;
suspicious;
or another operational category.

If the underlying information was wrong, the classification may also require review.

A corrected source should be capable of altering:

labels;
priority;
escalation;
routing;
or downstream recommendations

where the evidence supports change.

Corrections to Chronology

Chronological errors can materially alter institutional understanding.

Examples include:

wrong event dates;
confusion between event and record dates;
events placed in the wrong order;
later information presented as though it was known earlier;
or corrections omitted from the sequence.

Where chronology affects decision-making, correcting temporal information may require reconsideration of the analysis built upon it.

Corrections to Generated Citations

AI systems may produce incorrect:

citations;
source references;
quotations;
page numbers;
or attributions.

Where these materially support an institutional conclusion, correction should include verification against the underlying source.

A polished citation should not be assumed valid because it looks plausible.

Propagating Corrections Through AI Systems

Correction may be particularly difficult where AI systems continue retrieving old information.

Organisations should consider:

whether corrected records are re-indexed;
whether outdated versions remain retrievable;
whether search systems prioritise the corrected record;
whether prior summaries require regeneration;
whether downstream prompts still contain the inaccurate information;
and whether model memory or cached information may preserve the error.

A corrected database entry does not necessarily mean the operational AI system has been corrected.

Current Position vs Historical Record

Good record design should distinguish:

historically recorded

from

currently supported.

An institution may need to retain an earlier inaccurate or disputed record for audit purposes.

But later users should not be left to guess whether it remains current.

The corrected status should be visible.

Correction and Organisational Memory

Organisations need to remember that a correction occurred.

Otherwise, future staff or systems may rediscover the earlier record and repeat the same error.

Organisational memory should preserve:

the original issue;
the correction;
why it occurred;
what changed;
and what should now be relied upon.

Correction is therefore part of institutional memory, not merely record administration.

Learning From Correction

Repeated correction requests may reveal more than isolated factual mistakes.

They may indicate:

poor source quality;
weak summarisation;
recurring model errors;
ambiguous workflow design;
inadequate staff verification;
poor chronology handling;
weak data integration;
or insufficient human review.

Correction data can therefore become a governance signal.

Correction Metrics

Organisations may find it useful to examine:

number of correction requests;
proportion accepted;
time to resolution;
recurring types of error;
affected AI systems;
number of downstream records requiring update;
number of decisions reconsidered;
repeated errors after correction;
and common causes.

The objective is not simply performance measurement.

It is to identify systemic weaknesses.

When Errors Repeat

Repeated correction of the same type of error should trigger broader review.

Questions may include:

Is the source unreliable?
Is the AI summarisation process defective?
Is staff verification inadequate?
Is the same model limitation recurring?
Are corrections failing to propagate?
Is the workflow inherently producing the problem?

A governance system should learn from patterns rather than treat every correction as unrelated.

Correction and Incident Management

Some corrections may also constitute AI incidents.

This is more likely where an error:

materially affects a person;
influences a consequential decision;
spreads through multiple systems;
exposes confidential information;
repeats systematically;
or indicates a significant governance weakness.

Correction and incident response should therefore connect where appropriate.

High-Stakes Environments

Correction becomes especially important where AI-assisted records influence:

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

In these settings, delayed or ineffective correction can cause continuing harm even after the original error has been identified.

The greater the consequence, the stronger the need for:

accessible challenge;
rapid review;
source traceability;
downstream propagation;
and decision reconsideration.

Correction Before and After Consequence

Where an apparent error is identified before a consequential decision, governance should consider whether the decision should be paused pending review.

Where the error is identified afterwards, the organisation should consider whether the outcome requires reconsideration.

Correction governance therefore has two important questions:

Can we prevent an identified error from becoming consequential?

and

Can we repair the consequence if it already has?

Questions for Organisations

Where AI-assisted systems generate or rely upon institutional records, organisations may ask:

Can someone easily identify and challenge a factual error?
Is the correction route understandable without technical expertise?
Who reviews correction requests?
Is source evidence examined?
Can disputed information remain visibly disputed?
Are corrections annotated appropriately?
Can downstream records be identified?
Can corrected information propagate through AI systems?
Are affected decisions reconsidered where necessary?
Can later users distinguish historical information from the current position?
Are recurring corrections analysed as governance data?
Can the institution demonstrate that a material correction actually took effect?

SIAAF Relevance

This Guidance Note principally relates to:

Domain 03 — Evidence & Traceability

Can the organisation identify the source of an error and trace where it travelled?

Domain 04 — Communication & Feedback Integrity

Can corrected information move accurately through the institution?

Domain 05 — Escalation, Challenge & Contestability

Can an affected person or reviewer challenge an inaccurate record and obtain meaningful reconsideration?

It may also engage:

Domain 02 — Decision Integrity & Human Oversight

Where corrected information requires reconsideration of an AI-supported decision.

Domain 06 — Risk, Harm & Operational Resilience

Where inaccurate information creates continuing or consequential harm.

SWANK AI Standard

ERROR CORRECTION MUST BE PART OF SYSTEM DESIGN

A system capable of creating consequential information should also be capable of:

receiving challenge

checking source evidence

correcting or annotating the record

preserving dispute where unresolved

propagating material corrections

reconsidering affected decisions

and

learning from recurring error.

The purpose is not to erase institutional history.

It is to ensure that later users can distinguish:

what was originally recorded

from

what the evidence now supports.

Related SWANK AI Guidance

Guidance Note 02 — AI Summarisation and Record Integrity

Guidance Note 05 — Preserving Disagreement in AI-Assisted Records

Guidance Note 08 — Reassessment Alongside Escalation

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 20 — Governing AI in High-Stakes Environments

SWANK AI

Independent AI & Institutional Assurance

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